Underwater image super-resolution reconstruction method based on deep residual learning

Abstract

The invention aims to provide an underwater image super-resolution reconstruction method based on deep residual learning, relates to the technical field of underwater image super-resolution reconstruction, and provides an underwater image super-resolution reconstruction method based on a deep residual learning network (VDSR), which introduces a residual block structure and combines an alternate connection strategy to deepen the depth of the network. Not only is the characterization capability of the network to a complex underwater scene enhanced, but also the common problems of gradient disappearance and model degradation of the deep network are effectively relieved by introducing a residual learning mechanism. The alternative configuration of the identical residual connection blocks and the convolution residual connection blocks not only retains the stability advantage of the identical connection, but also promotes the effective multiplexing and transmission of the features, and significantly reduces the phenomenon of low-level feature loss in the deep network. Experimental results show that the proposed method is generally superior to a comparison algorithm in subjective and objective evaluation, and the texture, edge and other information of the reconstructed underwater image are greatly reserved.

Description

Technical Field

[0001] The present invention relates to the technical field related to underwater image super-resolution reconstruction, and in particular to an underwater image super-resolution reconstruction method based on deep residual learning. Background Art

[0002] As the consumption of resources on land accelerates day by day, people pay more and more attention to marine resources, and the development of underwater space and marine resources becomes very urgent. However, there are a large number of suspended particles in the water, which causes light to disperse and absorb when propagating in the water, resulting in color distortion, blurred background, low resolution and other problems in the underwater images obtained by the image acquisition equipment, which makes it extremely difficult to detect and identify underwater targets.

[0003] At present, since the super-resolution reconstruction algorithm can restore the missing high-frequency information while enlarging the size of the low-resolution image, it can effectively improve the resolution of underwater images, enhance the image details, and thus improve the image quality. Therefore, underwater image super-resolution reconstruction has attracted widespread attention from researchers.

[0004] Image super-resolution reconstruction algorithms are mainly divided into three categories: interpolation-based algorithms, reconstruction-based algorithms, and deep learning-based algorithms. Based on the traditional sparse coding method, the image super-resolution reconstruction algorithm is combined with deep learning, and SRCNN (Super-Resolution Convolutional Neural Networks, SRCNN) is proposed to solve the super-resolution problem, which greatly improves the quality of reconstructed images, but its shallow network structure limits the ability to reconstruct inferences; the multi-scale dense feature fusion network is used to extract features from images, effectively extracting rich image features. The deeper the neural network, the better the network performance, but too deep neural networks are prone to gradient disappearance during training; through the deep residual learning network (Deep Residual Learning Network, ResNet), the output obtained by cascading multiple convolutional layers and the input image are added and activated with the ReLU function to obtain a high-resolution image; VDSR and EDSR algorithms are also proposed in the prior art, which respectively introduce gradient clipping and delete batch normalization (BN) layers to the residual network to achieve better image restoration effects. However, these traditional deep learning algorithms assume that all extracted features are equally important and cannot suppress useless information.

[0005] In order to strengthen the learning of important features, this paper mainly optimizes the deep residual learning network, and uses the trained network to super-reconstruct the input underwater low-resolution image, that is, expands the size of the low-resolution image while supplementing its high-frequency information, and proposes a very deep residual super-resolution network (VDRSRN). Summary of the invention

[0006] In view of the deficiencies of the prior art, the present invention provides the following technical solutions:

[0007] A method for underwater image super-resolution reconstruction based on deep residual learning is proposed. By superimposing deep structures and incorporating an alternating residual connection strategy, a VDRSRN network is constructed to reconstruct underwater image super-resolution. The VDRSRN network includes image input, residual convolution and image reconstruction. The specific method includes the following steps:

[0008] S1, image preprocessing: convert the image to YCbCr color space, and train the VDRSRN network for the three channels respectively, where the brightness (Y) channel plays a major role in reconstruction, and the two chrominance channels (Cb) and (Cr) play auxiliary roles;

[0009] S2, image input: input the preprocessed image into the convolution layer for feature extraction to obtain a shallow feature map;

[0010] S3, residual convolution: The middle part of the VDRSRN network is composed of multiple residual blocks connected together, and the dilated convolution layer is used to replace the ordinary convolution layer in the residual block to deepen the network depth; the residual block includes an identical residual connection block and a convolution residual connection block, and the identical residual connection block and the convolution residual connection block are alternately connected, that is, each residual block is composed of an identical residual connection block and a convolution residual connection block;

[0011] S4, Image reconstruction: The VDRSRN network learns the residual information between the input image and the target image, and reconstructs the feature map into the final output residual image I through a common convolutional layer residual ;

[0012] The residual image I residual With the upsampled low-resolution image I up Combined, we get the reconstructed image I SR , the reconstructed image I SR Convert back to RGB space to get the color super-resolution reconstructed image I highers ;

[0013] S5, selects the root mean square error as the network loss function, which defines the difference between the model prediction and the actual label and guides the model to adjust its parameters to minimize the difference.

[0014] Furthermore, in S2, the preprocessed image is input into a convolution layer with a convolution kernel size of 5×5 and 64 channels for feature extraction, and the shallow feature map α0 obtained is as follows:

[0015] α0=δ(BN(F 5*5 (I LR )))

[0016] Where δ is the ReLU activation function, BN is the batch normalization layer, and F 5*5 is the convolutional layer, I LR is the input image.

[0017] Furthermore, in S3, the VDRSRN network uses the clipped gradient method, which clips the gradient interval to Avoid slow convergence and gradient explosion after deepening the network depth.

[0018] Furthermore, in S4, the residual information between the input image and the target image is learned, that is, learning to establish a high-resolution image I highers With low resolution image I lowers The mapping between residual :

[0019] I residual =I highers -I lowers

[0020] The image reconstruction part uses a common convolution layer with a convolution kernel size of 3×3 and 64 channels to reconstruct the feature map into the final output residual image I residual .

[0021] Furthermore, in S5, the root mean square error is used as the network loss function, which is defined as follows:

[0022]

[0023] Where L i is the loss function, f(r) is the residual image predicted by the network, I h For high-resolution images.

[0024] Compared with the prior art, the technical solution of this application has the following beneficial effects:

[0025] The present invention constructs a VDRSRN network through creative expansion and deep optimization of the classic VDSR architecture. The network deepens the network depth by superimposing deep structures and incorporating alternating residual connection strategies, which not only enhances the network's ability to characterize complex underwater scenes, but also effectively alleviates the common gradient vanishing and model degradation problems of deep networks by introducing a residual learning mechanism, significantly enhancing the model's ability to deeply fit image features and learn complex mapping relationships, effectively maintaining low-level features of images and improving the extraction accuracy of high-level features. Compared with traditional methods, VDRSRN shows significant performance improvement in super-resolution reconstruction of underwater images. The network provides an innovative solution for the processing of similar images and provides strong technical support for the fields of smart ocean and underwater exploration. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 This is a VDRSRN network structure diagram of the present invention;

[0027] Figure 2 This is a schematic diagram of the constant residual connection block of the present invention;

[0028] Figure 3 Schematic diagram of the convolution residual connection block of the present invention;

[0029] Figure 4 It is a schematic diagram of image reconstruction of the present invention;

[0030] Figure 5 Schematic diagram of the impact of different numbers of convolutions on the training loss function;

[0031] Figure 6 This is a schematic diagram for comparing subjective reconstruction effects. There are 7 rows from top to bottom in the figure, which are the original underwater image, NNI, Bicubic, SRCNN, VDSR, the reconstruction result of the algorithm in this paper, and HR image. DETAILED DESCRIPTION

[0032] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0033] Embodiment 1:

[0034] The two most important parts of a deep learning network are the network structure and the loss function. Inspired by VDSR, a global residual learning method is adopted to learn only the residual information between the input image and the target image, rather than directly learning the complete mapping from input to target. This greatly simplifies the learning task and improves network performance. At the same time, the network is improved by referring to EDSR and ResNet, and the VDRSRN network structure is proposed, such as Figure 1 As shown in the figure. VDRSRN consists of three parts: image input, residual convolution, and image reconstruction. In the network structure, Convolutional Layer represents the ordinary convolution layer, Residual Block represents the residual module, Batch Normalization Layer represents the batch normalization layer, which is used to standardize the input data in batches, ReLU (Rectified Liner Unit) Layer represents the rectified linear unit layer, and the activation function is ReLU.

[0035] The optimization of VDRSRN in network structure is mainly divided into the optimization of the entire network and the optimization of the residual convolution part.

[0036] Optimization of the overall network structure: In order to improve the learning efficiency and stability of the model, a batch normalization layer is introduced into the overall network architecture. This layer follows the convolutional layer and effectively speeds up the convergence of the network by normalizing the input data, and enhances the generalization ability of the model. This strategy not only helps reduce the risk of overfitting, but also ensures the stability of the network during training.

[0037] Optimization of the residual convolution part: In the residual convolution part, a number of measures are taken to optimize the network performance. First, by introducing a local residual learning mechanism, a unique residual block is designed, and residual connections are established between different layers, which effectively alleviates the gradient vanishing problem and improves the training speed and convergence effect of the model. Secondly, the ordinary convolution layer is replaced by the dilated convolution layer in the residual block to increase the receptive field of the convolution kernel, so that the network can capture richer contextual information while maintaining parameter efficiency. Finally, the strategy of alternating the connection of the identity residual connection block and the convolution residual connection block is adopted to further deepen the network structure, while avoiding the performance degradation problem caused by the increase in network depth, ensuring the stability and performance improvement of the model during the deepening process.

[0038] A method for underwater image super-resolution reconstruction based on deep residual learning, the specific steps are as follows:

[0039] 1.1 Image preprocessing and image input:

[0040] In the image preprocessing stage, because the human eye is significantly more sensitive to brightness information than to chrominance information, and due to the discontinuity of the HSV color space, artifacts will be generated when upsampling the image, so the image is converted to the YCbCr color space, and the VDRSRN network is trained for the three channels respectively, among which the brightness Y channel plays a major role in the reconstruction, and the two chrominance channels Cb and Cr play an auxiliary role, because the brightness Y is more in line with the human eye;

[0041] The preprocessed image is then input into a convolutional layer with a convolution kernel size of 5×5 and 64 channels for feature extraction to obtain a shallow feature map α0.

[0042] α0=δ(BN(F 5*5 (I LR )))

[0043] Where δ is the ReLU activation function, BN is the batch normalization layer, and F 5*5 is the convolutional layer, I LR is the input image.

[0044] 1.2. Residual Convolution:

[0045] The middle part of the network, which is also the most important part, is composed of multiple residual blocks connected together, each of which is composed of an identical residual connection block (such as Figure 2 ) and a convolutional residual connection block (such as Figure 3 ). This residual connection method can not only greatly reduce the risk of network gradient disappearance, but also transfer shallow layer features to deep layers. This further deepens the network depth to improve the network's receptive field size and learning generalization ability.

[0046] The ordinary convolution layer is replaced by the dilated convolution layer. Compared with the ordinary convolution layer, the dilated convolution introduces a new parameter called "dilation rate" to define the spacing between the values ​​when the convolution kernel processes data. This allows the dilated convolution kernel to have a larger receptive field than the ordinary convolution under the same computing conditions, thereby being able to capture a wider range of information. At the same time, since the dilated convolution kernel can increase the receptive field without increasing the amount of calculation, it can avoid increasing the receptive field by downsampling like ordinary convolution, thereby retaining the spatial resolution of the input feature map, which is very suitable for tasks such as image super-resolution that require retaining spatial information. As shown in Table 1, the difference in the receptive field of each layer under the ordinary convolution kernel and the dilated convolution with different dilation rates is shown when the convolution kernel is 3×3, the step size is 1, and the padding method is SAME.

[0047] Table 1 Receptive field table of ordinary convolution kernel and dilated convolution kernel

[0048]

[0049] The convolution kernel with a dilation rate of 1 is also a common convolution kernel. It can be seen that the receptive field of the network increases significantly with the increase of the dilation rate of the dilated convolution kernel compared to the common convolution kernel. Therefore, the feature semantics extracted by the dilated convolution kernel are richer, which greatly improves the effect of subsequent reconstruction.

[0050] In order to avoid the slow convergence and gradient explosion after deepening the network depth, VDRSRN is different from the traditional method of defining the gradient within [-θ, θ]. Instead, it adopts the clipping gradient method and clips the gradient interval to

[0051] 1.3 Image reconstruction:

[0052] The last part of the network is image reconstruction. In order to reduce the amount of calculation and improve the efficiency of calculation, the network only learns the residual information between the input image and the target image, rather than directly learning the complete mapping from the input to the target, that is, learning to build a high-resolution image I highers With low resolution image I lowers The mapping between residual :

[0053] I residual =I highers -I lowers

[0054] The image reconstruction part uses a common convolution layer with a convolution kernel size of 3×3 and 64 channels to reconstruct the feature map into the final output residual image I residual . Then the residual image is combined with the upsampled low-resolution image I up Combined, we get the reconstructed image I SR ,like Figure 4 As shown. Finally, the image is converted back to RGB space to obtain the color super-resolution reconstructed image I highers ;

[0055] 1.4. Loss function:

[0056] The loss function defines the difference between the model prediction and the actual label, and guides the model to adjust its parameters to minimize this difference. Therefore, a suitable loss function is crucial to building a high-performance neural network. The L1 loss function is one of the commonly used loss functions, but considering that it is not differentiable at zero and may cause unstable model training in some cases, the root mean square error (RMSE) is selected as the network loss function, which is defined as follows:

[0057]

[0058] Where L iis the loss function, f(r) is the residual image predicted by the network, I h For high-resolution images.

[0059] Embodiment 2:

[0060] According to Example 1, this example is an experimental result and analysis of an underwater image super-resolution reconstruction method based on deep residual learning:

[0061] 2.1 Evaluation indicators:

[0062] Image evaluation indexes are important criteria for measuring super-resolution reconstruction algorithms. These indicators not only reflect the performance of the algorithm on objective data, but also provide researchers with a reference for optimizing the algorithm. This paper uses subjective evaluation indicators and objective evaluation indicators to evaluate the performance of the proposed algorithm. Subjective evaluation indicators are mainly based on human eye observation. By comparing the reconstructed image with the original image, the differences in visual effects, texture details, edge information, etc. are evaluated. Although the subjective evaluation method can intuitively reflect the visual quality of the image, its results are often affected by factors such as the observer's experience and environment. Therefore, it is necessary to combine objective evaluation indicators for comprehensive evaluation.

[0063] In the image super-resolution reconstruction task, the commonly used objective evaluation indicators are Peak Signal to Noise Ratio (PSNR) and Structural Similarity (SSIM).

[0064] PSNR evaluates the fidelity of a reconstructed image by calculating the pixel difference between the reconstructed image and the original image. The larger the value, the smaller the image distortion and the better the image quality. The calculation formula of PSNR is as follows:

[0065]

[0066] in is the maximum value of pixels in the image, MSE is the mean square error between the reconstructed image and the original image, the unit of PSNR is dB, and its value is (0, +∞).

[0067] SSIM is an indicator used to measure the structural similarity between two images. It is a full-reference image quality evaluation indicator that can measure the similarity between two images in terms of brightness, contrast, and structure. Its value range is usually between 0 and 1. When two images are exactly the same, the SSIM value is equal to 1, indicating the best image quality or the least distortion. The SSIM calculation formula is as follows:

[0068]

[0069] Where X is the reconstructed image, Y is the original image, l(X, Y) is the brightness, c(X, Y) is the contrast, s(X, Y) is the structure; μ X , μ Y represents the mean of X and Y, σ X , σ Y represents the standard deviation of X and Y, σ XY Represents the covariance of X and Y, C1, C2, C3 are constant terms to avoid the denominator being 0;

[0070] 2.2 Data set and network training parameter settings:

[0071] The training set images used in this experiment come from the UFO-120 static natural image set widely used in the field of super-resolution underwater. At the same time, some images in the TURBID training set are selected for data supplementation to increase the number of data sets and improve the generalization ability of the model. Data augmentation uses flipping the images by 90°, 180°, and 270°, and splitting the images into 41×41 patches to further increase the number of data sets. Low-resolution images are downsampled using bicubic interpolation according to the scaling factors (×2, ×3, ×4). The maximum epochs is set to 10. The SGDM optimizer with parameter β1=0.9 is used for training, the initial learning rate is 0.1, the learning rate change factor is 0.3 per round, the batch size is 128, and the input of each convolutional layer is padded with zeros.

[0072] In order to further illustrate the impact of introducing residual connections to increase network depth on network training results, a comparative experiment was conducted using different network depths. Since the network used in this article is a convolutional network, the network depth is measured by the number of convolutional layers (excluding the convolutional layers in the convolutional residual connection). The network loss with different numbers of convolutional layers is shown in Figure 2. Figure 5 shown.

[0073] Figure 5 The horizontal axis in the middle represents the iteration round of network training, and the vertical axis is the training loss function value. It can be seen from the figure that the training loss tends to stabilize when the round reaches about 3. The network loss of 20 convolutional layers finally converges to about 3.2; the network loss of 30 layers finally converges to about 2.8; the network loss of 40 finally converges to about 0.8. With the increase in the number of convolutional layers, that is, the deepening of the network depth, the training loss has a significant decrease, and there is no result of gradient disappearance or network degradation caused by deepening the number of network layers, which proves the feasibility of this paper for network structure optimization. Therefore, a 40-layer network structure is used for experiments.

[0074] In order to better evaluate the reconstruction effect of VDRSRN, the bicubic linear interpolation method (Bicubic), SRCNN, VDSR and VDRSRN in this paper are compared under different scaling factors. The test set data uses images from the UFO-120TEST test data set. The reconstruction effect is compared using two objective evaluation indicators: average peak signal-to-noise ratio (PSNR) and structural similarity (SSIM). The experimental results are shown in Table 2.

[0075] Table 2 Average PSNR / SSIM of each reconstruction algorithm under different scaling factors

[0076]

[0077] As can be seen from Table 2, as the scaling factor increases, the image quality will continue to decline, and the difficulty of image super-resolution reconstruction will continue to increase. Therefore, for all super-resolution models, PSNR and SSIM will decrease as the scaling factor increases. At the same time, under the same scaling factor, the VDRSRN model built in this paper is also higher than the bicubic interpolation method and the two deep learning super-resolution models in all evaluation indicators, reflecting the advantages of the model in this paper in terms of reconstruction effect.

[0078] 2.3 Comprehensive evaluation:

[0079] In order to verify the effectiveness of the VDRSRN model built in this paper in image super-resolution reconstruction, and to facilitate the human eye to evaluate the reconstruction effect from a subjective perspective, LR images, HR images, Nearest Neighbor Interpolation (NNI), Bicubic, SRCNN, VDSR and VDRSRN were compared under the condition of a scaling factor of 4. The results of the comparison experiment are shown in Figure 2. Figure 6 shown.

[0080] Table 3 Objective evaluation comparison

[0081]

[0082] The above comparison experiment also focuses on details such as texture and edges, using red bounding boxes for positioning and partial magnification. Figure 6It can be seen that NNI and Bicubic based on interpolation methods only rely on the image's own information for reconstruction, and their ability to restore high-frequency information is extremely limited. Therefore, they perform poorly in terms of subjective observation of the human eye and objective evaluation indicators PSNR and SSIM, and the effect is poor. Among the learning-based methods, SRCNN initially uses convolutional networks. Although the reconstruction effect and objective indicators have been significantly improved compared with the traditional interpolation-based methods, the reconstruction of edge details is still not clear enough, and there is still a certain gap with the original image. Compared with SRCNN, VDSR is one step closer and has better feature extraction capabilities, so some simple edge reconstructions are clearer, but under the influence of factors such as turbid water quality and light refraction, the reconstruction effect of some small features is not good. The VDRSRN network proposed in this paper has shown excellent reconstruction performance in the above examples after a series of model optimizations. The edges of its reconstructed images are smoother and clearer, and the reconstruction of complex textures has also achieved remarkable results. At the same time, in terms of objective evaluation indicators PSNR and SSIM, VDRSRN has also reached the highest level among many models.

[0083] The above experiments demonstrate the effectiveness of VDRSRN in image super-resolution from both objective evaluation indicators and subjective perspectives.

[0084] In order to evaluate the overall performance of the algorithm, experiments were conducted on 124 underwater images in UFO-120TEST, and the average PSNR and SSIM of the images processed by each reconstruction algorithm were counted. The results are shown in Table 4:

[0085] Table 4 Overall objective evaluation of image reconstruction effect

[0086]

[0087] As can be seen from Table 4, compared with NNI, Bicubic, SRCNN, and VDSR, the average PSNR of the reconstructed images of the proposed algorithm is improved by 6.6, 4.87, 2.23, and 2.46 dB, respectively, and the average SSIM is improved by 0.133, 0.110, 0.048, and 0.041, respectively, indicating that the proposed method has better performance.

[0088] This application aims to solve the problem of detail loss in underwater low-resolution images due to turbid water, light refraction and other problems. This paper proposes a super-resolution reconstruction method based on deep learning optimization. Through the creative expansion and deep optimization of the classic VDSR architecture, the VDRSRN network is constructed. By superimposing deep structures and incorporating alternating residual connection strategies, the model's ability to deeply fit image features and learn complex mapping relationships significantly enhances the model's ability to effectively maintain low-level features of the image and improve the accuracy of high-level feature extraction. Compared with traditional methods, VDRSRN has shown significant performance improvements in super-resolution reconstruction of underwater images. The network provides an innovative solution for the processing of similar images and provides strong technical support for the fields of smart oceans and underwater exploration.

[0089] The preferred embodiments of the present invention disclosed above are only used to help illustrate the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to the specific implementation methods described. Obviously, many modifications and changes can be made according to the content of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can understand and use the present invention well. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. An underwater image super-resolution reconstruction method based on deep residual learning, characterized in that: By stacking deep structures and incorporating an alternating residual connection strategy, a VDRSRN network is constructed to reconstruct underwater image super-resolution. The VDRSRN network includes image input, residual convolution, and image reconstruction. The specific method includes the following steps: S1, image preprocessing: convert the image to YCbCr color space; S2, image input: input the preprocessed image into the convolution layer for feature extraction to obtain the shallow feature map α0; S3, residual convolution: The middle part of the VDRSRN network is composed of multiple residual blocks connected together, and the dilated convolution layer is used to replace the ordinary convolution layer in the residual block to deepen the network depth; the residual block includes an equal residual connection block and a convolution residual connection block, and the equal residual connection block and the convolution residual connection block are alternately connected; S4, Image reconstruction: The VDRSRN network learns the residual information between the input image and the target image, and reconstructs the feature map into the final output residual image I through a common convolutional layer residual ; The residual image I residual With the upsampled low-resolution image I up Combined, we get the reconstructed image I SR , the reconstructed image I SR Convert back to RGB space to get the color super-resolution reconstructed image I highers ; S5, selects the root mean square error as the network loss function, which defines the difference between the model prediction and the actual label and guides the model to adjust its parameters to minimize the difference.

2. According to claim 1, the underwater image super-resolution reconstruction method based on deep residual learning is characterized in that: In S2, the preprocessed image is input into a convolution layer with a convolution kernel size of 5×5 and 64 channels for feature extraction, and the shallow feature map α0 obtained is as follows: α0=δ(BN(F 5*5 (I LR ))) Where δ is the ReLU activation function, BN is the batch normalization layer, and F 5*5 is the convolutional layer, I LR is the input image.

3. The underwater image super-resolution reconstruction method based on deep residual learning according to claim 1, characterized in that: In S3, the VDRSRN network uses the clipped gradient method, which clips the gradient interval to Avoid slow convergence and gradient explosion after deepening the network depth.

4. The underwater image super-resolution reconstruction method based on deep residual learning according to claim 1, characterized in that: In S4, the residual information between the input image and the target image is learned, that is, learning to establish a high-resolution image I highers With low resolution image I lowers The mapping between residual : I residual =I highers -I lowers The image reconstruction part uses a common convolution layer with a convolution kernel size of 3×3 and 64 channels to reconstruct the feature map into the final output residual image I residual .

5. The underwater image super-resolution reconstruction method based on deep residual learning according to claim 1, characterized in that: In S5, the root mean square error is used as the network loss function, which is defined as follows: Where L i is the loss function, f(r) is the residual image predicted by the network, I h For high-resolution images.

Images