
The Bacillus Calmette-Guerin (BCG) vaccine has been in use for 99 years, and is regarded as one of the oldest human vaccines known today. It is recommended primarily due to its effect in preventing the most severe forms of tuberculosis, including disseminated tuberculosis and meningeal tuberculosis in children; however, its efficacy in preventing pulmonary tuberculosis and TB reactivation in adults has been questioned. Several studies however have found that asides from its role in tuberculosis prevention, the BCG vaccine also has protective effects against a host of other viral infections in humans, an effect which has been termed: heterologous, non-specific or off-target.
As we approach 100 years since the discovery of the BCG vaccine, we review the evidence of the non-specific protection offered by the vaccine against viral infections, discuss the possible mechanisms of action of these effects, highlight the implications these effects could have on vaccinology and summarize the recent epidemiological correlation between the vaccine and the on-going COVID-19 pandemic.
Several epidemiological studies have established that BCG does reduce all-cause mortality in infants, and also the time of vaccination influences this effect significantly. This effect has been attributed to the protective effect of the vaccine in preventing unrelated viral infections during the neonatal period. Some of such viral infections that have been investigated include: herpes simplex virus (HSV), human Papilloma virus (HPV), yellow fever virus (YFV), respiratory syncytial virus (RSV) and influenza virus type A (H1N1). These effects are thought to be mediated via induction of innate immune memory as well as heterologous lymphocytic activation. While epidemiological studies have suggested a correlation, the potential protection of the BCG vaccine against COVID-19 transmission and mortality rates is currently unclear. Ongoing clinical trials and further research may shed more light on the subject in the future.
BCG is a multifaceted vaccine, with many numerous potential applications to vaccination strategies being employed for current and future viral infections. There however is a need for further studies into the immunologic mechanisms behind these non-specific effects, for these potentials to become reality, as we usher in the beginning of the second century since the vaccine's discovery.
Citation: Oluwafolajimi A. Adesanya, Christabel I. Uche-Orji, Yeshua A. Adedeji, John I. Joshua, Adeniyi A. Adesola, Chibuike J. Chukwudike. Bacillus Calmette-Guerin (BCG): the adroit vaccine[J]. AIMS Microbiology, 2021, 7(1): 96-113. doi: 10.3934/microbiol.2021007
[1] | Qian Zhang, Haigang Li, Ming Li, Lei Ding . Feature extraction of face image based on LBP and 2-D Gabor wavelet transform. Mathematical Biosciences and Engineering, 2020, 17(2): 1578-1592. doi: 10.3934/mbe.2020082 |
[2] | Fang Zhu, Wei Liu . A novel medical image fusion method based on multi-scale shearing rolling weighted guided image filter. Mathematical Biosciences and Engineering, 2023, 20(8): 15374-15406. doi: 10.3934/mbe.2023687 |
[3] | Haohao Xu, Yuchen Gong, Xinyi Xia, Dong Li, Zhuangzhi Yan, Jun Shi, Qi Zhang . Gabor-based anisotropic diffusion with lattice Boltzmann method for medical ultrasound despeckling. Mathematical Biosciences and Engineering, 2019, 16(6): 7546-7561. doi: 10.3934/mbe.2019379 |
[4] | Michael James Horry, Subrata Chakraborty, Biswajeet Pradhan, Maryam Fallahpoor, Hossein Chegeni, Manoranjan Paul . Factors determining generalization in deep learning models for scoring COVID-CT images. Mathematical Biosciences and Engineering, 2021, 18(6): 9264-9293. doi: 10.3934/mbe.2021456 |
[5] | Auwalu Saleh Mubarak, Zubaida Said Ameen, Fadi Al-Turjman . Effect of Gaussian filtered images on Mask RCNN in detection and segmentation of potholes in smart cities. Mathematical Biosciences and Engineering, 2023, 20(1): 283-295. doi: 10.3934/mbe.2023013 |
[6] | Jimin Yu, Jiajun Yin, Shangbo Zhou, Saiao Huang, Xianzhong Xie . An image super-resolution reconstruction model based on fractional-order anisotropic diffusion equation. Mathematical Biosciences and Engineering, 2021, 18(5): 6581-6607. doi: 10.3934/mbe.2021326 |
[7] | Chen Yue, Mingquan Ye, Peipei Wang, Daobin Huang, Xiaojie Lu . SRV-GAN: A generative adversarial network for segmenting retinal vessels. Mathematical Biosciences and Engineering, 2022, 19(10): 9948-9965. doi: 10.3934/mbe.2022464 |
[8] | Hao Wang, Guangmin Sun, Kun Zheng, Hui Li, Jie Liu, Yu Bai . Privacy protection generalization with adversarial fusion. Mathematical Biosciences and Engineering, 2022, 19(7): 7314-7336. doi: 10.3934/mbe.2022345 |
[9] | Hui Yao, Yuhan Wu, Shuo Liu, Yanhao Liu, Hua Xie . A pavement crack synthesis method based on conditional generative adversarial networks. Mathematical Biosciences and Engineering, 2024, 21(1): 903-923. doi: 10.3934/mbe.2024038 |
[10] | Wei-wei Jiang, Guang-quan Zhou, Ka-Lee Lai, Song-yu Hu, Qing-yu Gao, Xiao-yan Wang, Yong-ping Zheng . A fast 3-D ultrasound projection imaging method for scoliosis assessment. Mathematical Biosciences and Engineering, 2019, 16(3): 1067-1081. doi: 10.3934/mbe.2019051 |
The Bacillus Calmette-Guerin (BCG) vaccine has been in use for 99 years, and is regarded as one of the oldest human vaccines known today. It is recommended primarily due to its effect in preventing the most severe forms of tuberculosis, including disseminated tuberculosis and meningeal tuberculosis in children; however, its efficacy in preventing pulmonary tuberculosis and TB reactivation in adults has been questioned. Several studies however have found that asides from its role in tuberculosis prevention, the BCG vaccine also has protective effects against a host of other viral infections in humans, an effect which has been termed: heterologous, non-specific or off-target.
As we approach 100 years since the discovery of the BCG vaccine, we review the evidence of the non-specific protection offered by the vaccine against viral infections, discuss the possible mechanisms of action of these effects, highlight the implications these effects could have on vaccinology and summarize the recent epidemiological correlation between the vaccine and the on-going COVID-19 pandemic.
Several epidemiological studies have established that BCG does reduce all-cause mortality in infants, and also the time of vaccination influences this effect significantly. This effect has been attributed to the protective effect of the vaccine in preventing unrelated viral infections during the neonatal period. Some of such viral infections that have been investigated include: herpes simplex virus (HSV), human Papilloma virus (HPV), yellow fever virus (YFV), respiratory syncytial virus (RSV) and influenza virus type A (H1N1). These effects are thought to be mediated via induction of innate immune memory as well as heterologous lymphocytic activation. While epidemiological studies have suggested a correlation, the potential protection of the BCG vaccine against COVID-19 transmission and mortality rates is currently unclear. Ongoing clinical trials and further research may shed more light on the subject in the future.
BCG is a multifaceted vaccine, with many numerous potential applications to vaccination strategies being employed for current and future viral infections. There however is a need for further studies into the immunologic mechanisms behind these non-specific effects, for these potentials to become reality, as we usher in the beginning of the second century since the vaccine's discovery.
At present, image color rendering as a major branch of image processing has attracted much attention. With the development of deep learning, image color rendering based on neural network has gradually become a research hotspot [1,2,3,4,5]. Because traditional color rendering methods require manual intervention and have high requirements of reference images. Moreover, when the structure and color of the image are complex, color rendering effect is not ideal [6,7,8,9,10]. Color rendering methods based on deep learning can be easily deployed in the actual production environment, and the limitation of the traditional methods can be solved [11,12,13]. By using the neural network model and the corresponding dataset training model [14,15], the image can be automatically rendered according to the model, without being affected by human or other factors [16,17,18,19].
Larsson [20] used the convolutional neural network to consider the brightness of the image as input, decomposed the color and saturation of the image by the super-column model, to realize color rendering. Iizuka [21] combined the low-dimensional feature and global feature of the image by using the fusion layer in the convolutional neural network, for generating the color of the image and processing images of any resolution. Zhang [22] designed an appropriate loss function to handle the multi-mode uncertainty in color rendering and maintain the color diversity. However, when the grayscale image features are extracted using the above mentioned method, up-sampling is adopted to make the image size consistent, resulting in the loss of image information. Moreover, the network structure cannot well extract and understand the complex features of the image, and the rendering effect is limited [23,24,25].
Isola [26] improved conditional generative adversarial networks (CGAN) to achieve the transformation between images. The proposed pix2pix model can realize conversion between different images, for example, color rendering can be realized by learning the mapping relationship between grayscale image and color image [27,28]. But the pix2pix model based generative adversarial networks (GAN) has the disadvantage of training instability. Moreover, the current image rendering methods based on deep learning are not good at rendering robust images. Gabor filter can easily extract texture information in all scales and directions of the image, and reduce the influence of light change and noise in the image to a certain extent.
Therefore, we propose a color rendering method using Gabor filter based improved pix2pix for robust image. The contributions of this paper are mainly there-folds:
(1) The improved pix2pix model can not only automatically complete image rendering and achieve good visual effect, but also achieve more stable training and better image quality.
(2) Gabor filter was added to enhance the robustness of model rendered images.
(3) The metric data of a series of experiments show that the proposed method has better performance for robust image.
The rest of the paper is organized as follows. Section 2 introduces the previous work, including Gabor filter and pix2pix model. Section 3 describes the method and its design details. Section 4 introduces the experiment and comparison experiment, and evaluates the image quality. Section 5 conclusions the paper and outlooks the future work.
Fourier transform is a powerful tool in signal processing, which can help us transform images from spatial domain to frequency domain, and extract features that are not easy to extract in spatial domain. However, after Fourier transform, frequency features of images at different locations are often mixed together, but Gabor filter can extract spatial local frequency features, which is an effective texture detection tool [29,30]. The Gabor filter is derived by multiplying a Gaussian by a cosine function [31,32,33], it is defined as
g(x,y,λ,θ,φ,σ,γ)=exp(−x′2+γ2y22σ2)exp(i(2πx′λ+φ)) | (2.1) |
greal(x,y,λ,θ,φ,σ,γ)=exp(−x′2+γ2y22σ2)cos(i(2πx′λ+φ)) | (2.2) |
gimag(x,y,λ,θ,φ,σ,γ)=exp(−x′2+γ2y22σ2)sin(i(2πx′λ+φ)) | (2.3) |
where, x′=xcosθ+ysinθ,y′=−xsinθ+ycosθ. Where, x, y represent the coordinate position of the pixel, λ represents the wavelength of the filter, θ represents the tilt degree of the Gabor kernel image, φ represents the phase offset, σ represents the standard deviation of the Gaussian function, and γ represents the aspect ratio.
In order to make full use of the characteristics of Gabor filters, r filter extracts the texture features of the image in 6 scales and 4 directions. Namely, the Gabor it is necessary to design Gabor filters with different directions and scales to extract features. In this study, the Gaboscales are 7, 9, 11, 13, 15 and 17. The Gabor directions are 0°, 45°, 90° and 135°, as shown in Figure 1(a). Extract effective texture feature sets from the output results of the filter. The extracted texture feature sets are shown in Figure 1(b), with 24 texture feature maps in total.
At present, image rendering based on generative adversarial networks [34] attracts much attention because it can directly generate color images by using mapping relations. Therefore, it is widely used in image processing, text processing, natural language processing and other fields. pix2pix model [26] is a model for image-to-image conversion based generative adversarial networks. It can better synthesize image or generate color image. The following are the main features of the pix2pix model.
(1) Both the generator and discriminator structure use the convolution unit of Conv-Batchnorm-ReLU, namely, convolutional layer, batch normalization and ReLU Loss are used.
(2) The input of the pix2pix model is the specified image, such as the label image to the real image, the input is the label image, the input is the grayscale image to the color image, and the input is the grayscale image. The grayscale image as the input of the generator, and the input and output of the generator as the input of the discriminator, so as to establish the corresponding relationship between the input image and the output image, realize user control, and complete image color rendering.
(3) PatchGAN was used as discriminator for pix2pix model. Specifically, the image is divided into several fixed-size blocks, and the authenticity of each block is determined. Finally, the average value is taken as the final output. A network structure similar to U-net is adopted as a generator, and skip connections are added between i and n−i at each layer to simulate U-net, where n is the total number of layers of the network. Not only can the path be shrunk for context information, but the symmetric extension path can be positioned precisely.
(4) The loss function of the pix2pix model is as follows, which is composed of L1 loss and Vanilla GAN loss. Where, let x be the input image, y be the expected output, G be the generator, and D be the discriminator:
G∗=argminGmaxDLcGAN(G,D)+λLL1(G) | (2.4) |
LcGAN(G,D)=Ex,y(logD(x,y))+Ex(log(1−D(x,G(x)))) | (2.5) |
LL1(G)=Ex,y(∥y−G(x)1∥) | (2.6) |
In view of the detail problems existing in the generative adversarial networks based image color rendering method in complex scenes, this paper proposes an image color rendering method using Gabor filter based improved pix2pix for robust image. The network framework is shown in Figure 2. The rendering process is shown in Figure 3. After selecting the data set for training, the trained generator is used for color rendering.
Firstly, we preprocessed the image with Gabor filter, and extracted the texture feature set of the image as input for training and verification. By comparing 24 Gabor texture feature maps with 6 scales and 4 directions, the texture map with 7 scales and 0° direction has the best color rendering effect. Secondly, this paper utilizes the existing pix2pix model architecture for image transformation to perform color rendering by learning the mapping relationship between grayscale image and color image. Finally, although the pix2pix model solves some problems existing in the generative adversarial networks, it still has the instability problem of training on large-scale image dataset. Therefore, the least square loss in LSGAN [35] is used in the objective function of pix2pix model, and the penalty term similar to WGAN_GP [36] is added. We improve the overall model framework, it is shown that the proposed method has a better performance on the rendering of robust images by a series of comparison experiments.
The generator in generative adversarial networks hopes that the output data distribution can be more close to the distribution of the real data. Meanwhile, the discriminator of generative adversarial networks needs to make a judgment between the real data and the output data by the generator to find the real data and the fake data. The loss function can generate more real data through the Lipschitz constraint generative adversarial networks. The traditional generative adversarial networks uses the cross entropy loss or Vanilla GAN loss as the loss function. The classification is correct, but gradient dispersion occurs when the generator is updated [36,37]. LSGAN uses the square loss as the objective function, and the least square loss function penalizes the samples (fake samples) that are in the discriminant true but far away from the decision boundary, and drags the false samples far away from the decision boundary into the decision boundary, to improve the quality of the generated image.
Therefore, compared with the traditional generative adversarial networks, the image generated by LSGAN has higher quality and a more stable training process. So the least square loss function is adopted in the framework of this paper.
{minDVLSGAN(D)=12ExPdata(x)[(D(x)−b)2]+12EzPz(z)[(D(G(z))−a)2]minDVLSGAN(D)=12EzPz(z)[(D(G(z))−c)2] | (3.1) |
where, the input image is x, expected output is y, generator is G, discriminator is D, noise is z, labels of generated sample and real sample are a and b, respectively. c is the value set by the generator to let the discriminator think the generated image is real data.
Generative adversarial networks can generate better data distribution, but it has the problem of training instability. Improving the training stability of generative adversarial networks is a hot topic in deep learning. Wasserstein generative adversarial networks (WGAN) [38] uses Wasserstein distance to generate a value function with better theoretical properties than JS divergence in order to constrain the Lipschitz constant of the discriminator function, which basically solves the problems of generative adversarial networks training instability and model collapse and ensures the diversity of generated samples [39]. WGAN_GP continues to improve on WGAN, and its penalty term is derived from the Wasserstein distance, where the penalty coefficient is 10.
The objective function of WGAN_GP is as follows, adding the original critic loss and the gradient penalty term of WGAN_GP.
L=E˜x∼Pg[D(˜x)]−E˜x∼Pr[D(x)]+λE˜x∼Pˆx[(∥▽ˆxD(ˆx∥2−1)2] | (3.2) |
where, E˜x∼Pg[D(˜x)]−E˜x∼Pr[D(x)] is the original critic loss, λE˜x∼Pˆx[(∥▽ˆxD(ˆx∥2−1)2] denotes the gradient penalty term of WGAN_GP, ˆx=tˆx+(1−t)x, 0≤t≤1, and λ is the penalty coefficient.
To verify the effectiveness and accuracy of the proposed method, we conducted extensive experiments on summer dataset [40], with 1231 pieces of train set, and 309 pieces of test set. Experiment 1 is conducted to test the effect of application of the Gabor filter and different objective functions in the pix2pix model environment. Experiment 2 is performed to test the rendering effect when different Gabor texture feature maps are given as input. Experiment 3 is conducted to test whether the penalty term should be added in the discriminator. Experiment 4 is the rendering effect of low-quality or robust images was tested by adding noise and dimming the brightness of the image for assessing the robustness of this model.
Training parameters: The experiment was performed on a PC with Intel(R) Core(TM) i7-9750H CPU @ 2.6 GHz 2.59 GHz, a graphics card NVIDIA GeForce GTX 1650, and CUDA+Cudnn for acceleration training. The proposed method is implemented based on Python 3.7 and Pytorch framework. The number of experimental training iterations is 200, optimizer is Adam, batch_size is 1, learning rate is 0.0002, and number of processes is 4.
Network structures and implementation details: All the models we train are designed to 256 × 256 images. The input image of the model is 512 × 256; the left one is the original color image, and the right one is the texture feature map processed by the Gabor filter, as shown in Figure 4. By default, the pix2pix model uses a generator similar to U-net, PatchGAN, and Vanilla GAN loss.
Evaluation Metrics: To reflect image color rendering quality of different models more objectively, peak signal to noise ratio (PSNR) and structural similarity (SSIM) indexes are adopted to evaluate the rendered images [41,42]. These two indexes are often used in the evaluation metrics of image processing. PSNR is an objective standard to evaluate the quality of the color image produced. The calculation formula is as follows:
PSNR=10log10(2n−1)2MSE | (4.1) |
MSE=1H×WH∑i=1W∑j=1[X(i,j)−Y(i,j)]2 | (4.2) |
where, H and W represent the width and height of the image respectively, (i,j) represents each pixel point, and n represents the number of bits of the pixel, X and Y represent two images respectively.
Because PSNR index also has its limitations, it cannot completely reflect the consistency of image quality and human visual effect, so SSIM index is used for further comparison. SSIM is a metric to measure the similarity of two images. By comparing the image rendered by the model with the original color image, the effectiveness and accuracy of this algorithm are demonstrated. The calculation formula is as follows:
SSIM=(2μxμy+c1)(2σxy+c2)(μ2xμ2y+c1)(σ2xσ2y+c2) | (4.3) |
where, μx and μy respectively represent the average value of the real image and the generated image, σ2x and σ2y respectively represent the variance of the real image and the generated image, σxy represents the covariance of the real image and the generated image, c1=(k1,L)2 and c2=(k2,L)2 are constants that maintain stability, and L is the dynamic range of pixel value, k1=0.01, k2=0.03.
In this study, the Gabor filter extracts the texture features of the image in 6 scales and 4 directions. For convenience, according to the texture feature set shown in Figure 1(b), the images are numbered from left to right and from top to bottom. The direction is assumed to be d and the size is s, as shown in Figure 5. For example, G1 means "s = 7, d = 0°". So the direction is 0° and the scale is 7. G6 means "s = 17, d = 0°". So the direction is 0° and the scale is 17. By default, the pix2pix model uses Vanilla GAN loss. Based on pix2pix model, the model using least squares loss function is called LSpix (least squares pix2pix). Based on Gabor filter, the model using Gabor texture maps is called pixGn (pix2pix Gabor n), n = 1, 6, 7, 13.
To test the effect of application of the Gabor filter and different objective functions in the pix2pix model environment, we divided the experiment into adding Gabor filters (Figures 6(c), (e)), not adding Gabor filters (Figures 6(b), (d)), using least squares loss (Figures 6(d), (e)) or Vanilla GAN loss (Figures 6(b), (c)). By comparing the images in Figure 6, it can be confirmed that the rendering effect preprocessed by least square loss or Gabor filter is better, which is the LSpixG1 model. This is because Gabor can preprocess images and obtain multi-scale and multi-direction features of images, so as to achieve good and fast feature extraction and learning during network model learning. Moreover, compared with other loss functions, the least square loss function only reaches saturation at one point, which is not easy to cause the problem of gradient disappearance.
Tables 1 and 2 compare the distortion and structural similarity between the rendered image and the ground truth, show the maximum, minimum, and average indexes. This is an additional interpretation of Figure 6. The LSpix model has the highest score in the maximum and average PSNR, which is 3.591dB and 1.083dB higher than that of the pix2pix model. Meanwhile, the LSpix model has the highest score in SSIM, which is 1.618%, 15.649% and 3.848% higher than that of the pix2pix model, respectively. This proves that our model is closer to ground truth in structure, and the colors are more reductive.
Network | MAX PSNR | MIN PSNR | AVE PSNR |
pix2pix | 29.204 | 11.126 | 23.024 |
pixG1 | 28.225 | 10.477 | 19.981 |
LSpix | 32.795 | 11.003 | 24.107 |
pixG6 | 27.874 | 9.883 | 20.012 |
LSpixG6 | 32.616 | 10.632 | 21.409 |
LSpixG1 | 32.524 | 11.238 | 21.354 |
Note: Bold font is the best value for each column. |
Network | MAX SSIM | MIN SSIM | AVE SSIM |
pix2pix | 92.888 | 52.474 | 82.163 |
pixG1 | 86.101 | 36.592 | 69.145 |
LSpix | 94.506 | 68.123 | 86.011 |
pixG6 | 85.625 | 33.117 | 68.845 |
LSpixG6 | 91.312 | 56.897 | 78.387 |
LSpixG1 | 91.757 | 54.785 | 78.485 |
Note: Bold font is the best value for each column. |
In order to test the rendering effect when different Gabor texture feature maps are input, we use different feature images as input. Figure 7 shows how different Gabor texture images are rendered when Vanilla GAN loss is the target function of the pix2pix model. Figures 5(c), (d), that is, the direction is the scale is 7 and 45° or 90°, contain incomplete details of the original image, resulting in incomplete input texture features. Therefore, the generated image is blurred, as shown in Figures 7(a), (b). Although the 7th and 13th texture images were considered as training sets (pixG7+G13 model) with a total of 1231 × 2 images taken together, the rendering effect was not significantly improved, as shown in Figure 7(b). Evidently, by comparing the images in Figure 7, it can be found out that the visual effect of Figures 7(c)–(e) is good and not blurred. And Table 3 and 4 show the evaluation indexes after the input of different feature maps. The data show that incomplete input of texture feature map is not desirable.
Network | MAX PSNR | MIN PSNR | AVE PSNR |
pixG1 | 28.225 | 10.477 | 19.981 |
pixG6 | 27.874 | 9.883 | 20.012 |
pixG7 | 27.565 | 9.232 | 17.600 |
pixG1+G13 | 28.960 | 9.947 | 20.682 |
Note: Bold font is the best value for each column. |
Network | MAX SSIM | MIN SSIM | AVE SSIM |
pixG1 | 86.101 | 36.592 | 69.145 |
pixG6 | 85.625 | 33.117 | 68.845 |
pixG7 | 91.188 | 4.964 | 39.057 |
pixG1+G13 | 87.615 | 42.562 | 71.630 |
Note: Bold font is the best value for each column. |
And to compare the operation efficiency of input different texture maps, the training time is shown in Table 5, in hours. Regardless of whether the Gabor filter was used, which texture map was entered, the operation time was around 9 hours. However, if two texture maps are used for training, such as G1 and G13 are used in pixG1+G13 model, the training set doubles and the pre-training time doubles. Even though the results shown in Figure 7(d) are good, the method is not desirable. This is because when we use filtering, we need to extract multi-scale and multi-direction features and remove redundant information. Once the important information is removed, it will certainly have a certain impact on the results, resulting in blurred images.
Model | pix2pix | pixG1 | pixG6 | pixG7 | pixG7+G13 | pixG1+G13 |
Time | 8.72 | 9.00 | 8.76 | 8.43 | 15.27 | 16.52 |
Figure 8 shows the performance on whether or not to add a penalty item in the discriminator based on the pixG1 model. Figure 8(a) is the effect of not adding penalty items, and Figure 8(b) is the effect of adding penalty items. Obviously, Figure 8(b) has less error in detail and better visual effect. Penalty term, that is, gradient punishment is carried out by interpolation method to make the model satisfy Lipschitz constraint. The addition of punishment terms similar to WGAN_GP basically solves the problems of training instability and model collapse in the GAN model and ensures the diversity of generated samples.
Tables 6 and 7 show the evaluation indexes whether or not to add a penalty item. With the addition of penalty term, the LSpix_GP model achieved the highest score in the minimum PSNR, which was 0.904dB higher than that of the original pix2pix model. Evidently, in the texture map extracted based on Gabor filter, the image with scale of 7 and direction of 0° has the best training effect. Furthermore, when the objective function is least squares loss, the average SSIM and performance are improved. When penalty term is added, the score of maximum and average SSIM is the highest, which is 1.753% and 1.083% higher than that of the pix2pix model. Therefore, the image rendered by the LSpixG1_GP model is better than that of the original model.
Network | MAX PSNR | MIN PSNR | AVE PSNR |
pix2pix | 29.204 | 11.126 | 23.024 |
pixG1 | 28.225 | 10.477 | 19.981 |
pixG6 | 27.874 | 9.883 | 20.012 |
LSpix | 32.795 | 11.003 | 24.107 |
LSpixG1 | 32.524 | 11.238 | 21.354 |
LSpix_GP | 31.859 | 12.030 | 24.019 |
LSpixG1_GP | 32.342 | 11.514 | 21.290 |
LSpixG6 | 32.616 | 10.632 | 21.409 |
LSpixG1_GP | 32.113 | 11.067 | 21.384 |
Note: Bold font is the best value for each column. |
Network | MAX SSIM | MIN SSIM | AVE SSIM |
pix2pix | 92.888 | 52.474 | 82.163 |
pixG1 | 86.101 | 36.592 | 69.145 |
pixG6 | 85.625 | 33.117 | 68.845 |
LSpix | 94.506 | 68.123 | 86.011 |
LSpixG1 | 91.757 | 54.785 | 78.485 |
LSpix_GP | 94.641 | 67.250 | 85.967 |
LSpixG1_GP | 90.772 | 54.308 | 78.067 |
LSpixG6 | 91.312 | 56.897 | 78.387 |
LSpixG6_GP | 90.941 | 52.740 | 78.236 |
Note: Bold font is the best value for each column. |
To compare the operating efficiency of different objective functions given as input and increase the punishment items, the running time is listed in Table 8 in hours. For example, LSpixG6_GP represents using the least squares loss, adding the penalty item, the direction is 0° and the scale is 17. Regardless of whether Gabor filter was used, which texture map was input, whether Vanilla GAN loss or least square loss was the target function, the training time was approximately 9 h. Although the algorithm efficiency of adding the filter alone is basically the same, the time of using the filter after adding the penalty term will be increased by 2–3 h. Therefore, the algorithm in this study adopts LSpixG1_GP model, namely Gabor texture map with model input scale of 7 and direction of 0°, least squares loss and penalty term.
Model | pix2pix | LSpix | LSpixG1 | LSpix_GP | LSpixG1_GP | LSpixG6 | LSpixG6_GP |
Time | 8.72 | 8.66 | 8.66 | 8.72 | 11.17 | 8.72 | 11.72 |
Note: Bold font is the best value for each row. |
In order to evaluate the robustness of the model for rendering robust image, the rendering effect of low-quality images was tested by adding noise and dimming the image brightness, as shown in Figure 9. When testing the noise image, the Gaussian noise image with mean value of 0 and variance of 10 is added. When testing low-illumination images, power operation is performed on the pixels of the image, and the power is set to 2.5 to generate low-illumination images.
We use PSNR evaluation metric to evaluate the rendering results of each model for low-quality images. As shown in Table 9, the image rendered by the LSpix model is of higher quality when rendering noisy images. As shown in Table 10, images rendered by Gabor filter models are generally of good quality for low-illumination images. After the Gabor filter, the objective function is least square loss and the penalty term is added, the image quality of the LSpixG1_GP model is higher than that of the original model. This is because the method in this paper uses Gabor filter to avoid the interference of noise to the image to a certain extent. And when extracting features, the depth information of the image can be extracted to avoid the influence of light on the image. Clearly, the proposed method in this paper is robust to color rendering of low-quality images.
Network | MAX PSNR | MIN PSNR | AVE PSNR |
pix2pix | 29.489 | 10.770 | 22.528 |
pixG1 | 25.657 | 10.562 | 18.665 |
LSpix | 29.655 | 11.805 | 22.528 |
LSpixG1 | 27.650 | 12.409 | 19.942 |
LSpix_GP | 29.516 | 11.950 | 22.504 |
LSpixG1_GP | 27.306 | 11.548 | 19.966 |
Note: Bold font is the best value for each column. |
Network | MAX PSNR | MIN PSNR | AVE PSNR |
pix2pix | 21.977 | 8.723 | 12.535 |
pixG1 | 26.158 | 7.864 | 14.441 |
LSpix | 21.457 | 9.119 | 12.579 |
LSpixG1 | 24.565 | 7.946 | 14.171 |
LSpix_GP | 21.948 | 9.334 | 12.563 |
LSpixG1_GP | 24.337 | 7.886 | 14.127 |
Note: Bold font is the best value for each column. |
We proposed a novel image color rendering method based on using Gabor filter based improved pix2pix for robust image and demonstrate its feasibility and superiority for a variety of tasks. It enables automatically render robust images and has good robustness with low-quality image rendering. The experimental results on summer dataset demonstrate that the proposed method can achieve high-quality performance with image color rendering. At present, the image resolution of image processing based on deep learning is limited, which leads to the limitation in the practical application of rendering method. In the future, we will focus on improving the resolution of network model input images.
This work were partially supported by the National Natural Science Foundation of China (No. 62002285 and No. 61902311).
The authors declare there is no conflict of interest.
[1] | Luca S, Mihaescu T (2013) History of BCG vaccine. Maedica (Buchar) 8: 53-58. |
[2] |
Benévolo-de-Andrade TC, Monteiro-Maia R, Cosgrove C, et al. (2005) BCG Moreau Rio de Janeiro-An oral vaccine against tuberculosis-review. Mem Inst Oswaldo Cruz 100: 459-465. doi: 10.1590/S0074-02762005000500002
![]() |
[3] | Dagg B, Hockley J, Rigsby P, et al. (2014) The establishment of sub-strain specific WHO Reference Reagents for BCG vaccine 32: 6390-6395. |
[4] |
Comstock G (1994) The International tuberculosis campaign: a pioneering venture in mass vaccination and research. Clin Infect Dis 19: 528-540. doi: 10.1093/clinids/19.3.528
![]() |
[5] |
Zwerling A, Behr MA, Verma A, et al. (2011) The BCG World Atlas: A database of global BCG vaccination policies and practices. PLoS Med 8: e1001012. doi: 10.1371/journal.pmed.1001012
![]() |
[6] |
Trunz BB, Fine P, Dye C (2006) Effect of BCG vaccination on childhood tuberculous meningitis and miliary tuberculosis worldwide: a meta-analysis and assessment of cost-effectiveness. Lancet 367: 1173-1180. doi: 10.1016/S0140-6736(06)68507-3
![]() |
[7] | Awasthi S, Moin S (1999) Effectiveness of BCG vaccination against tuberculous meningitis. Indian Pediatr 36: 455-460. |
[8] |
Mangtani P, Abubakar I, Ariti C, et al. (2013) Protection by BCG vaccine against Tuberculosis: a systematic review of randomized controlled trials. Clin Infect Dis 58: 470-80. doi: 10.1093/cid/cit790
![]() |
[9] |
Setia MS, Steinmaus C, Ho CS, et al. (2006) The role of BCG in prevention of leprosy: A meta-analysis. Lancet Infect Dis 6: 162-170. doi: 10.1016/S1473-3099(06)70412-1
![]() |
[10] |
Shann F (2013) Nonspecific effects of vaccines and the reduction of mortality in children. Clin Ther 35: 109-114. doi: 10.1016/j.clinthera.2013.01.007
![]() |
[11] |
Shann F (2010) The non-specific effects of vaccines. Arch Dis Child 95: 662-667. doi: 10.1136/adc.2009.157537
![]() |
[12] |
Kristensen I, Aaby P, Jensen H (2000) Routine vaccinations and child survival: Follow up study in Guinea-Bissau, West Africa. Br Med J 321: 1435-1439. doi: 10.1136/bmj.321.7274.1435
![]() |
[13] |
Roth A, Gustafson P, Nhaga A, et al. (2005) BCG vaccination scar associated with better childhood survival in Guinea-Bissau. Int J Epidemiol 34: 540-547. doi: 10.1093/ije/dyh392
![]() |
[14] |
Garly ML, Martins CL, Balé C, et al. (2003) BCG scar and positive tuberculin reaction associated with reduced child mortality in West Africa: A non-specific beneficial effect of BCG? Vaccine 21: 2782-2790. doi: 10.1016/S0264-410X(03)00181-6
![]() |
[15] |
Biering-Sørensen S, Aaby P, Lund N, et al. (2017) Early BCG-Denmark and neonatal mortality among infants weighing <2500 g: A randomized controlled trial. Clin Infect Dis 65: 1183-1190. doi: 10.1093/cid/cix525
![]() |
[16] |
Nankabirwa V, Tumwine JK, Mugaba PM, et al. (2015) Child survival and BCG vaccination: A community based prospective cohort study in Uganda. BMC Public Health 15: 175. doi: 10.1186/s12889-015-1497-8
![]() |
[17] |
Zimmermann P, Finn A, Curtis N (2018) Does BCG vaccination protect against nontuberculous mycobacterial infection? A systematic review and meta-analysis. J Infect Dis 218: 679-687. doi: 10.1093/infdis/jiy207
![]() |
[18] |
Aaby P, Roth A, Ravn H, et al. (2011) Randomized Trial of BCG vaccination at birth to low-birth-weight children: beneficial nonspecific effects in the neonatal period? J Infect Dis 204: 245-252. doi: 10.1093/infdis/jir240
![]() |
[19] |
Biering-Sørensen S, Aaby P, Napirna BM, et al. (2012) Small randomized trial among low-birth-weight children receiving bacillus Calmette-Guéerin vaccination at first health center contact. Pediatr Infect Dis J 31: 306-308. doi: 10.1097/INF.0b013e3182458289
![]() |
[20] |
Stensballe LG, Sørup S, Aaby P, et al. (2017) BCG vaccination at birth and early childhood hospitalisation: A randomised clinical multicentre trial. Arch Dis Child 102: 224-231. doi: 10.1136/archdischild-2016-310760
![]() |
[21] |
Stensballe LG, Nante E, Jensen IP, et al. (2005) Acute lower respiratory tract infections and respiratory syncytial virus in infants in Guinea-Bissau: A beneficial effect of BCG vaccination for girls: Community based case-control study. Vaccine 23: 1251-1257. doi: 10.1016/j.vaccine.2004.09.006
![]() |
[22] | Wardhana, Datau E, Sultana A, et al. (2011) The efficacy of Bacillus Calmette-Guérin vaccinations for the prevention of acute upper respiratory tract infection in the elderly. Acta Med Indones 43: 185-190. |
[23] |
Ohrui T, Nakayama K, Fukushima T, et al. (2005) Prevention of elderly pneumonia by pneumococcal, influenza and BCG vaccinations. Japanese J Geriatr 42: 34-36. doi: 10.3143/geriatrics.42.34
![]() |
[24] |
Salem A, Nofal A, Hosny D (2013) Treatment of common and plane warts in children with topical viable bacillus calmette-guerin. Pediatr Dermatol 30: 60-63. doi: 10.1111/j.1525-1470.2012.01848.x
![]() |
[25] |
Podder I, Bhattacharya S, Mishra V, et al. (2017) Immunotherapy in viral warts with intradermal Bacillus Calmette-Guerin vaccine versus intradermal tuberculin purified protein derivative: A double-blind, randomized controlled trial comparing effectiveness and safety in a tertiary care center in Eastern India. Indian J Dermatol Venereol Leprol 83: 411. doi: 10.4103/0378-6323.188651
![]() |
[26] |
Daulatabad D, Pandhi D, Singal A (2016) BCG vaccine for immunotherapy in warts: is it really safe in a tuberculosis endemic area? Dermatol Ther 29: 168-172. doi: 10.1111/dth.12336
![]() |
[27] |
Morales A, Eidinger D, Bruce AW (1976) Intracavitary Bacillus Calmette Guerin in the treatment of superficial bladder tumors. J Urol 116: 180-182. doi: 10.1016/S0022-5347(17)58737-6
![]() |
[28] | Jackson A, James K (1994) Understanding the most successful immunotherapy for cancer. Immunol 2: 208-215. |
[29] |
Prescott S, James K, Busuttil A, et al. (1989) HLA—DR Expression by High Grade Superficial Bladder Cancer Treated with BCG. Br J Urol 63: 264-269. doi: 10.1111/j.1464-410X.1989.tb05187.x
![]() |
[30] |
Jackson AM, Alexandroff AB, McIntyre M, et al. (1994) Induction of ICAM 1 expression on bladder tumours by BCG immunotherapy. J Clin Pathol 47: 309-312. doi: 10.1136/jcp.47.4.309
![]() |
[31] | Meyer J-P, Persad R (2002) Use of bacille Calmette-Guérin in superficial bladder cancer. Postgrad Med J 78. |
[32] |
Leentjens J, Kox M, Stokman R, et al. (2015) BCG vaccination enhances the immunogenicity of subsequent influenza vaccination in healthy volunteers: A randomized, placebo-controlled pilot study. J Infect Dis 212: 1930-1938. doi: 10.1093/infdis/jiv332
![]() |
[33] |
Ritz N, Mui M, Balloch A, et al. (2013) Non-specific effect of Bacille Calmette-Guérin vaccine on the immune response to routine immunisations. Vaccine 31: 3098-3103. doi: 10.1016/j.vaccine.2013.03.059
![]() |
[34] |
Ota MOC, Vekemans J, Schlegel-Haueter SE, et al. (2002) Influence of Mycobacterium bovis Bacillus Calmette-Guérin on antibody and cytokine responses to human neonatal vaccination. J Immunol 168: 919-925. doi: 10.4049/jimmunol.168.2.919
![]() |
[35] |
Scheid A, Borriello F, Pietrasanta C, et al. (2018) Adjuvant effect of Bacille Calmette-Guérin on hepatitis B vaccine immunogenicity in the preterm and term newborn. Front Immunol 9: 29. doi: 10.3389/fimmu.2018.00029
![]() |
[36] |
Arts RJW, Moorlag SJCFM, Novakovic B, et al. (2018) BCG vaccination protects against experimental viral infection in humans through the induction of cytokines associated with trained immunity. Cell Host Microbe 23: 89-100. doi: 10.1016/j.chom.2017.12.010
![]() |
[37] | FD A, RN U, CL L (1974) Recurrent herpes genitalis. Treatment with Mycobacterium bovis (BCG). Obs Gynecol 43: 797-805. |
[38] | Hippmann G, Wekkeli M, Rosenkranz A, et al. (1992) Nonspecific immune stimulation with BCG in Herpes simplex recidivans. Follow-up 5 to 10 years after BCG vaccination. Wien Klin Wochenschr 104: 200-204. |
[39] |
Ahmed SS, Volkmuth W, Duca J, et al. (2015) Antibodies to influenza nucleoprotein cross-react with human hypocretin receptor 2. Sci Transl Med 7. doi: 10.1126/scitranslmed.aab2354
![]() |
[40] |
Su LF, Kidd BA, Han A, et al. (2013) Virus-Specific CD4+ Memory-Phenotype T Cells Are Abundant in Unexposed Adults. Immunity 38: 373-383. doi: 10.1016/j.immuni.2012.10.021
![]() |
[41] |
Bernasconi NL, Traggiai E, Lanzavecchia A (2002) Maintenance of serological memory by polyclonal activation of human memory B cells. Science 298: 2199-2202. doi: 10.1126/science.1076071
![]() |
[42] |
Kleinnijenhuis J, Quintin J, Preijers F, et al. (2014) Long-lasting effects of bcg vaccination on both heterologous th1/th17 responses and innate trained immunity. J Innate Immun 6: 152-158. doi: 10.1159/000355628
![]() |
[43] |
Vetskova EK, Muhtarova MN, Avramov TI, et al. (2013) Immunomodulatory effects of BCG in patients with recurrent respiratory papillomatosis. Folia Med 55: 49-54. doi: 10.2478/folmed-2013-0005
![]() |
[44] |
Mathurin KS, Martens GW, Kornfeld H, et al. (2009) CD4 T-Cell-Mediated Heterologous Immunity between Mycobacteria and Poxviruses. J Virol 83: 3528-39. doi: 10.1128/JVI.02393-08
![]() |
[45] | Kandasamy R, Voysey M, McQuaid F, et al. (2016) Non-specific immunological effects of selected routine childhood immunisations: Systematic review. BMJ 355. |
[46] |
Netea MG, Joosten LAB, Latz E, et al. (2016) Trained immunity: A program of innate immune memory in health and disease. Science 352: 427. doi: 10.1126/science.aaf1098
![]() |
[47] |
Kleinnijenhuis J, Quintin J, Preijers F, et al. (2012) Bacille Calmette-Guérin induces NOD2-dependent nonspecific protection from reinfection via epigenetic reprogramming of monocytes. Proc Natl Acad Sci USA 109: 17537-17542. doi: 10.1073/pnas.1202870109
![]() |
[48] |
Rathinam VAK, Fitzgerald KA (2010) Inflammasomes and anti-viral immunity. J Clin Immunol 30: 632-637. doi: 10.1007/s10875-010-9431-4
![]() |
[49] |
Allen IC, Scull MA, Moore CB, et al. (2009) The NLRP3 inflammasome mediates in vivo innate immunity to influenza a virus through recognition of viral RNA. Immunity 30: 556-565. doi: 10.1016/j.immuni.2009.02.005
![]() |
[50] |
Thomas PG, Dash P, Aldridge JR, et al. (2009) The intracellular sensor NLRP3 mediates key innate and healing responses to influenza a virus via the regulation of Caspase-1. Immunity 30: 566-575. doi: 10.1016/j.immuni.2009.02.006
![]() |
[51] |
Wu F, Zhao S, Yu B, et al. (2020) A new coronavirus associated with human respiratory disease in China. Nature 579: 265-269. doi: 10.1038/s41586-020-2008-3
![]() |
[52] | Miller A, Reandelar MJ, Fasciglione K, et al. Correlation between universal BCG vaccination policy and reduced morbidity and mortality for COVID-19: an epidemiological study (2020) . |
[53] |
Scheid A, Borriello F, Pietrasanta C, et al. (2018) Adjuvant effect of Bacille Calmette-Guérin on hepatitis B vaccine immunogenicity in the preterm and term newborn. Front Immunol 9: 24. doi: 10.3389/fimmu.2018.00029
![]() |
[54] |
Sato H, Jing C, Isshiki M, et al. (2013) Immunogenicity and safety of the vaccinia virus LC16m8δ vector expressing SIV Gag under a strong or moderate promoter in a recombinant BCG prime-recombinant vaccinia virus boost protocol. Vaccine 31: 3549-3557. doi: 10.1016/j.vaccine.2013.05.071
![]() |
[55] |
Yazdanian M, Memarnejadian A, Mahdavi M, et al. (2013) Immunization of mice by BCG formulated HCV core protein elicited higher th1-oriented responses compared to pluronic-F127 copolymer. Hepat Mon 13. doi: 10.5812/hepatmon.14178
![]() |
[56] |
Kim BJ, Kim BR, Kook YH, et al. (2018) Development of a live recombinant BCG expressing human immunodeficiency virus type 1 (HIV-1) gag using a pMyong2 vector system: Potential use as a novel HIV-1 vaccine. Front Immunol 9: 27. doi: 10.3389/fimmu.2018.00027
![]() |
[57] |
Zhou P, Yang X Lou, Wang XG, et al. (2020) A pneumonia outbreak associated with a new coronavirus of probable bat origin. Nature 579: 270-273. doi: 10.1038/s41586-020-2012-7
![]() |
[58] |
Adesanya OA, Adewale BA, Ebengho IG, et al. (2020) Current knowledge on the pathogenesis of and therapeutic options against SARS-CoV-2: An extensive review of the available evidence. Int J Pathog Res 4: 16-36. doi: 10.9734/ijpr/2020/v4i230108
![]() |
[59] | Johns Hopkins COVID-19 Map-Johns Hopkins Coronavirus Resource Center (2020) .Available from: https://coronavirus.jhu.edu/map.html. |
[60] | Hegarty PK, Service NH, Kamat AM, et al. (2020) BCG vaccination may be protective against Covid-19. medRxiv (March). |
[61] | Dayal D, Gupta S (2020) Connecting BCG Vaccination and COVID-19: Additional Data. medRxiv (April). |
[62] | Dolgikh S (2020) Further Evidence of a Possible Correlation between the Severity of Covid-19 and BCG Immunization. J Infect Dis Epidemiol 6. |
[63] |
Sharma A, Kumar Sharma S, Shi Y, et al. (2020) BCG vaccination policy and preventive chloroquine usage: do they have an impact on COVID-19 pandemic? Cell Death Dis 11: 516. doi: 10.1038/s41419-020-2720-9
![]() |
[64] |
Ebina-Shibuya R, Horita N, Namkoong H, et al. (2020) National policies for paediatric universal BCG vaccination were associated with decreased mortality due to COVID-19. Respirology 25: 898-899. doi: 10.1111/resp.13885
![]() |
[65] |
Kinoshita M, Tanaka M (2020) Impact of Routine Infant BCG Vaccination on COVID-19. J Infect 81: 625-633. doi: 10.1016/j.jinf.2020.08.013
![]() |
[66] |
Klinger D, Blass I, Rappoport N, et al. (2020) Significantly improved COVID-19 outcomes in countries with higher bcg vaccination coverage: A multivariable analysis. Vaccines 8: 1-14. doi: 10.3390/vaccines8030378
![]() |
[67] |
Weng CH, Saal A, Butt WWW, et al. (2020) Bacillus Calmette-Guérin vaccination and clinical characteristics and outcomes of COVID-19 in Rhode Island, United States: A cohort study. Epidemiol Infect 148: e140. doi: 10.1017/S0950268820001569
![]() |
[68] | Kirov S (2020) Association Between BCG Policy is Significantly Confounded by Age and is Unlikely to Alter Infection or Mortality Rates. medRxiv (April). |
[69] | Hensel J, McGrail D, McAndrews K, et al. (2020) Exercising caution in correlating COVID-19 incidence and mortality rates with BCG vaccination policies due to variable rates of SARS CoV-2 testing. medRxiv . |
[70] | (2020) ClinicalTrials.gov.BCG Vaccination to Protect Healthcare Workers Against COVID-19.National Library of Medicine (U.S.). |
[71] | (2020) ClinicalTrials.gov.Reducing Health Care Workers Absenteeism in Covid-19 Pandemic Through BCG Vaccine.National Library of Medicine (U.S.). |
[72] | ClinicalTrials.gov. BCG Vaccine for Health Care Workers as Defense Against COVID 19 (2020) . |
[73] | (2020) ClinicalTrials.gov.BCG Vaccine in Reducing Morbidity and Mortality in Elderly Individuals in COVID-19 Hotspots.National Library of Medicine (U.S.). |
[74] | Adesanya O, Ebengho I (2020) Possible Correlation between Bacillus Calmette Guérin (BCG) Vaccination Policy and SARS-Cov-2 Transmission, Morbidity and Mortality Rates: Implications for the African Continent. J Infect Dis Epidemiol 6: 137. |
1. | Hongan Li, Guanyi Wang, Qiaozhi Hua, Zheng Wen, Zhanli Li, Ting Lei, An image watermark removal method for secure internet of things applications based on federated learning, 2022, 0266-4720, 10.1111/exsy.13036 | |
2. | Hong-an Li, Guanyi Wang, Kun Gao, Haipeng Li, A Gated Convolution and Self-Attention-Based Pyramid Image Inpainting Network, 2022, 31, 0218-1266, 10.1142/S0218126622502085 | |
3. | Peng Zhao, Yongxin Zhang, Qiaozhi Hua, Haipeng Li, Zheng Wen, Bio-Inspired Optimal Dispatching of Wind Power Consumption Considering Multi-Time Scale Demand Response and High-Energy Load Participation, 2023, 134, 1526-1506, 957, 10.32604/cmes.2022.021783 | |
4. | Hong’an Li, Min Zhang, Dufeng Chen, Jing Zhang, Meng Yang, Zhanli Li, Image Color Rendering Based on Hinge-Cross-Entropy GAN in Internet of Medical Things, 2023, 135, 1526-1506, 779, 10.32604/cmes.2022.022369 | |
5. | Xiaonan Shi, Jian Huang, Bo Huang, An Underground Abnormal Behavior Recognition Method Based on an Optimized Alphapose-ST-GCN, 2022, 31, 0218-1266, 10.1142/S0218126622502140 | |
6. | Zhi-Hua zhao, Li Chen, Bifurcation Fusion Network for RGB-D Salient Object Detection, 2022, 31, 0218-1266, 10.1142/S0218126622502152 | |
7. | Hong-an Li, Liuqing Hu, Qiaozhi Hua, Meng Yang, Xinpeng Li, Image Inpainting Based on Contextual Coherent Attention GAN, 2022, 31, 0218-1266, 10.1142/S0218126622502097 | |
8. | Mei Gao, Baosheng Kang, Blind Image Inpainting Using Low-Dimensional Manifold Regularization, 2022, 31, 0218-1266, 10.1142/S0218126622502115 | |
9. | Qianqian Liu, Xiaoyan Zhang, Qiaozhi Hua, Zheng Wen, Haipeng Li, Yan Huo, Adaptive Differential Evolution Algorithm with Simulated Annealing for Security of IoT Ecosystems, 2022, 2022, 1530-8677, 1, 10.1155/2022/6951849 | |
10. | Hong’an Li, Jiangwen Fan, Qiaozhi Hua, Xinpeng Li, Zheng Wen, Meng Yang, Biomedical sensor image segmentation algorithm based on improved fully convolutional network, 2022, 197, 02632241, 111307, 10.1016/j.measurement.2022.111307 | |
11. | Youzhong Ma, Qiaozhi Hua, Zheng Wen, Ruiling Zhang, Yongxin Zhang, Haipeng Li, Thippa Reddy G, k Nearest Neighbor Similarity Join Algorithm on High-Dimensional Data Using Novel Partitioning Strategy, 2022, 2022, 1939-0122, 1, 10.1155/2022/1249393 | |
12. | Wenchao Ren, Liangfu Li, Shiyi Wen, Lingmei Ai, APE-GAN: A colorization method for focal areas of infrared images guided by an improved attention mask mechanism, 2024, 124, 00978493, 104086, 10.1016/j.cag.2024.104086 | |
13. | Thavavel Vaiyapuri, Jaiganesh Mahalingam, Sultan Ahmad, Hikmat A. M. Abdeljaber, Eunmok Yang, Soo-Yong Jeong, Ensemble Learning Driven Computer-Aided Diagnosis Model for Brain Tumor Classification on Magnetic Resonance Imaging, 2023, 11, 2169-3536, 91398, 10.1109/ACCESS.2023.3306961 | |
14. | Samarth Deshpande, Siddhi Deshmukh, Atharva Deshpande, Devyani Manmode, Sakshi Dhamne, Abha Marathe, 2023, Pseudo Coloring Using Deep Learning Approach, 979-8-3503-4805-7, 218, 10.1109/ICAECIS58353.2023.10170408 | |
15. | Ali Salim Rasheed, Marwa Jabberi, Tarek M. Hamdani, Adel M. Alimi, PIXGAN-Drone: 3D Avatar of Human Body Reconstruction From Multi-View 2D Images, 2024, 12, 2169-3536, 74762, 10.1109/ACCESS.2024.3404554 | |
16. | Mamoona Jamil, Mubashar Sarfraz, Sajjad A. Ghauri, Muhammad Asghar Khan, Mohamed Marey, Khaled Mohamad Almustafa, Hala Mostafa, Optimized Classification of Intelligent Reflecting Surface (IRS)-Enabled GEO Satellite Signals, 2023, 23, 1424-8220, 4173, 10.3390/s23084173 | |
17. | Muhammad Asif Khan, Hamid Menouar, Ridha Hamila, 2023, Crowd Counting in Harsh Weather using Image Denoising with Pix2Pix GANs, 979-8-3503-7051-5, 1, 10.1109/IVCNZ61134.2023.10343548 | |
18. | Huda F. AL-Shahad, Razali Yaakob, Nurfadhlina Mohd Sharef, Hazlina Hamdan, Hasyma Abu Hassan, An Improved Pix2pix Generative Adversarial Network Model to Enhance Thyroid Nodule Segmentation, 2025, 16, 17982340, 37, 10.12720/jait.16.1.37-48 | |
19. | Qizhi Zou, Binghua Wang, Zhaofei Jiang, Qian Wu, Jian Liu, Xinting Ji, Dynamic style transfer for interior design: An IoT-driven approach with DMV-CycleNet, 2025, 117, 11100168, 662, 10.1016/j.aej.2024.12.030 | |
20. | Alvan Reyhanza Vittorino, Giovanus Immanuel, Simeon Yuda Prasetyo, Eko Setyo Purwanto, 2024, Machine Learning Approaches for Diabetic Retinopathy Classification Utilizing Gabor, LBP, and HOG Feature Extraction, 979-8-3315-0857-9, 697, 10.1109/BTS-I2C63534.2024.10942171 |
Network | MAX PSNR | MIN PSNR | AVE PSNR |
pix2pix | 29.204 | 11.126 | 23.024 |
pixG1 | 28.225 | 10.477 | 19.981 |
LSpix | 32.795 | 11.003 | 24.107 |
pixG6 | 27.874 | 9.883 | 20.012 |
LSpixG6 | 32.616 | 10.632 | 21.409 |
LSpixG1 | 32.524 | 11.238 | 21.354 |
Note: Bold font is the best value for each column. |
Network | MAX SSIM | MIN SSIM | AVE SSIM |
pix2pix | 92.888 | 52.474 | 82.163 |
pixG1 | 86.101 | 36.592 | 69.145 |
LSpix | 94.506 | 68.123 | 86.011 |
pixG6 | 85.625 | 33.117 | 68.845 |
LSpixG6 | 91.312 | 56.897 | 78.387 |
LSpixG1 | 91.757 | 54.785 | 78.485 |
Note: Bold font is the best value for each column. |
Network | MAX PSNR | MIN PSNR | AVE PSNR |
pixG1 | 28.225 | 10.477 | 19.981 |
pixG6 | 27.874 | 9.883 | 20.012 |
pixG7 | 27.565 | 9.232 | 17.600 |
pixG1+G13 | 28.960 | 9.947 | 20.682 |
Note: Bold font is the best value for each column. |
Network | MAX SSIM | MIN SSIM | AVE SSIM |
pixG1 | 86.101 | 36.592 | 69.145 |
pixG6 | 85.625 | 33.117 | 68.845 |
pixG7 | 91.188 | 4.964 | 39.057 |
pixG1+G13 | 87.615 | 42.562 | 71.630 |
Note: Bold font is the best value for each column. |
Model | pix2pix | pixG1 | pixG6 | pixG7 | pixG7+G13 | pixG1+G13 |
Time | 8.72 | 9.00 | 8.76 | 8.43 | 15.27 | 16.52 |
Network | MAX PSNR | MIN PSNR | AVE PSNR |
pix2pix | 29.204 | 11.126 | 23.024 |
pixG1 | 28.225 | 10.477 | 19.981 |
pixG6 | 27.874 | 9.883 | 20.012 |
LSpix | 32.795 | 11.003 | 24.107 |
LSpixG1 | 32.524 | 11.238 | 21.354 |
LSpix_GP | 31.859 | 12.030 | 24.019 |
LSpixG1_GP | 32.342 | 11.514 | 21.290 |
LSpixG6 | 32.616 | 10.632 | 21.409 |
LSpixG1_GP | 32.113 | 11.067 | 21.384 |
Note: Bold font is the best value for each column. |
Network | MAX SSIM | MIN SSIM | AVE SSIM |
pix2pix | 92.888 | 52.474 | 82.163 |
pixG1 | 86.101 | 36.592 | 69.145 |
pixG6 | 85.625 | 33.117 | 68.845 |
LSpix | 94.506 | 68.123 | 86.011 |
LSpixG1 | 91.757 | 54.785 | 78.485 |
LSpix_GP | 94.641 | 67.250 | 85.967 |
LSpixG1_GP | 90.772 | 54.308 | 78.067 |
LSpixG6 | 91.312 | 56.897 | 78.387 |
LSpixG6_GP | 90.941 | 52.740 | 78.236 |
Note: Bold font is the best value for each column. |
Model | pix2pix | LSpix | LSpixG1 | LSpix_GP | LSpixG1_GP | LSpixG6 | LSpixG6_GP |
Time | 8.72 | 8.66 | 8.66 | 8.72 | 11.17 | 8.72 | 11.72 |
Note: Bold font is the best value for each row. |
Network | MAX PSNR | MIN PSNR | AVE PSNR |
pix2pix | 29.489 | 10.770 | 22.528 |
pixG1 | 25.657 | 10.562 | 18.665 |
LSpix | 29.655 | 11.805 | 22.528 |
LSpixG1 | 27.650 | 12.409 | 19.942 |
LSpix_GP | 29.516 | 11.950 | 22.504 |
LSpixG1_GP | 27.306 | 11.548 | 19.966 |
Note: Bold font is the best value for each column. |
Network | MAX PSNR | MIN PSNR | AVE PSNR |
pix2pix | 21.977 | 8.723 | 12.535 |
pixG1 | 26.158 | 7.864 | 14.441 |
LSpix | 21.457 | 9.119 | 12.579 |
LSpixG1 | 24.565 | 7.946 | 14.171 |
LSpix_GP | 21.948 | 9.334 | 12.563 |
LSpixG1_GP | 24.337 | 7.886 | 14.127 |
Note: Bold font is the best value for each column. |
Network | MAX PSNR | MIN PSNR | AVE PSNR |
pix2pix | 29.204 | 11.126 | 23.024 |
pixG1 | 28.225 | 10.477 | 19.981 |
LSpix | 32.795 | 11.003 | 24.107 |
pixG6 | 27.874 | 9.883 | 20.012 |
LSpixG6 | 32.616 | 10.632 | 21.409 |
LSpixG1 | 32.524 | 11.238 | 21.354 |
Note: Bold font is the best value for each column. |
Network | MAX SSIM | MIN SSIM | AVE SSIM |
pix2pix | 92.888 | 52.474 | 82.163 |
pixG1 | 86.101 | 36.592 | 69.145 |
LSpix | 94.506 | 68.123 | 86.011 |
pixG6 | 85.625 | 33.117 | 68.845 |
LSpixG6 | 91.312 | 56.897 | 78.387 |
LSpixG1 | 91.757 | 54.785 | 78.485 |
Note: Bold font is the best value for each column. |
Network | MAX PSNR | MIN PSNR | AVE PSNR |
pixG1 | 28.225 | 10.477 | 19.981 |
pixG6 | 27.874 | 9.883 | 20.012 |
pixG7 | 27.565 | 9.232 | 17.600 |
pixG1+G13 | 28.960 | 9.947 | 20.682 |
Note: Bold font is the best value for each column. |
Network | MAX SSIM | MIN SSIM | AVE SSIM |
pixG1 | 86.101 | 36.592 | 69.145 |
pixG6 | 85.625 | 33.117 | 68.845 |
pixG7 | 91.188 | 4.964 | 39.057 |
pixG1+G13 | 87.615 | 42.562 | 71.630 |
Note: Bold font is the best value for each column. |
Model | pix2pix | pixG1 | pixG6 | pixG7 | pixG7+G13 | pixG1+G13 |
Time | 8.72 | 9.00 | 8.76 | 8.43 | 15.27 | 16.52 |
Network | MAX PSNR | MIN PSNR | AVE PSNR |
pix2pix | 29.204 | 11.126 | 23.024 |
pixG1 | 28.225 | 10.477 | 19.981 |
pixG6 | 27.874 | 9.883 | 20.012 |
LSpix | 32.795 | 11.003 | 24.107 |
LSpixG1 | 32.524 | 11.238 | 21.354 |
LSpix_GP | 31.859 | 12.030 | 24.019 |
LSpixG1_GP | 32.342 | 11.514 | 21.290 |
LSpixG6 | 32.616 | 10.632 | 21.409 |
LSpixG1_GP | 32.113 | 11.067 | 21.384 |
Note: Bold font is the best value for each column. |
Network | MAX SSIM | MIN SSIM | AVE SSIM |
pix2pix | 92.888 | 52.474 | 82.163 |
pixG1 | 86.101 | 36.592 | 69.145 |
pixG6 | 85.625 | 33.117 | 68.845 |
LSpix | 94.506 | 68.123 | 86.011 |
LSpixG1 | 91.757 | 54.785 | 78.485 |
LSpix_GP | 94.641 | 67.250 | 85.967 |
LSpixG1_GP | 90.772 | 54.308 | 78.067 |
LSpixG6 | 91.312 | 56.897 | 78.387 |
LSpixG6_GP | 90.941 | 52.740 | 78.236 |
Note: Bold font is the best value for each column. |
Model | pix2pix | LSpix | LSpixG1 | LSpix_GP | LSpixG1_GP | LSpixG6 | LSpixG6_GP |
Time | 8.72 | 8.66 | 8.66 | 8.72 | 11.17 | 8.72 | 11.72 |
Note: Bold font is the best value for each row. |
Network | MAX PSNR | MIN PSNR | AVE PSNR |
pix2pix | 29.489 | 10.770 | 22.528 |
pixG1 | 25.657 | 10.562 | 18.665 |
LSpix | 29.655 | 11.805 | 22.528 |
LSpixG1 | 27.650 | 12.409 | 19.942 |
LSpix_GP | 29.516 | 11.950 | 22.504 |
LSpixG1_GP | 27.306 | 11.548 | 19.966 |
Note: Bold font is the best value for each column. |
Network | MAX PSNR | MIN PSNR | AVE PSNR |
pix2pix | 21.977 | 8.723 | 12.535 |
pixG1 | 26.158 | 7.864 | 14.441 |
LSpix | 21.457 | 9.119 | 12.579 |
LSpixG1 | 24.565 | 7.946 | 14.171 |
LSpix_GP | 21.948 | 9.334 | 12.563 |
LSpixG1_GP | 24.337 | 7.886 | 14.127 |
Note: Bold font is the best value for each column. |