SAR (Synthetic Aperture Radar) image change detection method for adaptively eliminating noise features

Abstract

The invention relates to the technical field of image processing, and discloses an SAR (Synthetic Aperture Radar) image change detection method for adaptively eliminating noise features, which comprises the following steps: constructing an SAR image change detection model; performing data processing on the two to-be-detected original images by adopting a logarithm ratio operator and a two-stage fuzzy C-means clustering method to generate a double-time-phase image block; inputting the dual-time-phase image blocks into a multi-scale feature module for multi-scale feature extraction, and generating a fusion feature map; inputting the fused feature map into a first convolution down-sampling module for down-sampling to generate a first down-sampling feature map; inputting the first down-sampling characteristic pattern into a self-adaptive noise characteristic elimination module for noise elimination, and generating a noise elimination characteristic pattern; sequentially inputting the noise-eliminated feature map into a second convolution down-sampling module, a convolution module and a classifier for down-sampling, convolution and classification operation to obtain the classification probability of each category; according to the invention, the precision of image change detection is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, in particular to a SAR image change detection method for adaptively eliminating noise features. BACKGROUND

[0002] Due to the fact that SAR is not affected by heavy fog, light, rainy weather and day and night, it can work all day and all weather, and can effectively penetrate and identify cover under very low visibility weather conditions, obtaining rich reflection, polarization and scattering characteristics of ground object image data. Therefore, SAR image reflects more structural information than optical image, and becomes a reliable signal source in change detection. However, like other coherent imaging systems, the random scattering wave of the ground object surface will cause a large number of speckle noise in the image. It seriously restricts the improvement of SAR image change detection accuracy. Relying on manual interpretation or traditional methods to obtain SAR image change detection results cannot meet the requirements of high efficiency, automation and high precision.

[0003] In the traditional change detection method, the detection result is often obtained by analyzing the difference map. Although there are many works to suppress the noise influence in the difference map. However, these methods are highly dependent on manual experience and cannot meet the automation requirements.

[0004] Compared with the traditional change detection method, the deep network can automatically extract the multi-level features of the original image, and the features obtained by training have robustness, which can obtain better change detection results. With the successful application of convolutional neural network in various fields, SAR image change detection based on CNN has gradually attracted people's attention. However, the existing method makes more use of the powerful feature extraction ability of deep learning model to extract robust features, and lacks effective processing method for noise features in SAR image. Although the convolutional wavelet neural network replaces the pooling layer of CNN with wavelet transform to make the model have stronger noise suppression ability, the threshold in wavelet transform is fixed and cannot be learned, and the fixed threshold is difficult to adaptively eliminate noise features in different samples. Especially for SAR image data set, the information contained in different data sets is different (water area, forest, mountain, city, etc.), and a large number of different objects are also contained in the same data set. SUMMARY

[0005] In view of the above problems in the prior art, the present application provides a SAR image change detection method for adaptively eliminating noise features, which can effectively process noise features in the image, thereby solving the problem of low image change detection accuracy.

[0006] In order to achieve the above application purpose, the technical scheme adopted by the present application is: A SAR image change detection method for adaptively eliminating noise features, comprising the following steps: Collect two original images to be detected, and use a logarithmic ratio operator and a two-stage fuzzy C-means clustering method to process data to generate double-time image blocks; A SAR image change detection model is constructed, which includes a multi-scale feature module, a first convolution downsampling module, an adaptive noise feature elimination module, a second convolution downsampling module, a convolution module, and a classifier; The double-time image blocks are input into the multi-scale feature module to extract multi-scale features, and the extracted multi-scale feature maps are fused to generate a fused feature map; The fused feature map is input into the first convolution downsampling module for downsampling to generate a first downsampled feature map; The first downsampled feature map is input into the adaptive noise feature elimination module to eliminate noise through soft thresholding to generate a noise-eliminated feature map; The noise-eliminated feature map is input into the second convolution downsampling module for downsampling to generate a second downsampled feature map; The second downsampled feature map is input into the convolution module for two-dimensional convolution operation to generate a two-dimensional convolution feature map; The two-dimensional convolution feature map is input into the classifier, and the normalized exponential activation function is used for mapping to obtain the classification probability of each class.

[0007] The present application has the following advantages: The SAR image change detection method for adaptively eliminating noise features uses a logarithmic ratio operator and a two-stage fuzzy C-means clustering method to process training data, avoids manual class label annotation, is beneficial to selecting high-quality training samples, and thus improves the precision of the SAR image change detection model in image change detection; Meanwhile, four different scale convolution kernels are introduced to extract features from the input image, which not only better expresses the complex change information in the image from multiple scales, but also retains features that are beneficial to improving the image change detection precision, thereby improving the image class prediction probability, i.e., improving the image change detection precision; In addition, by calculating an adaptive threshold and through residual connection, not only the noise and redundant features in the feature map are eliminated, but also the precision of the feature map is improved, and finally the precision of the image change detection is improved. BRIEF DESCRIPTION OF DRAWINGS

[0008] Figure 1 The flowchart of the SAR image change detection method for adaptively eliminating noise features is shown in the figure; Figure 2 The figure shows the pseudo-labels and true labels of the Ottawa data set in the embodiment; Figure 3A structural schematic diagram of the SAR image change detection model in the embodiment; Figure 4 A structural schematic diagram of the multi-scale feature module in the embodiment; Figure 5 A structural schematic diagram of the adaptive noise feature elimination module in the embodiment. DETAILED DESCRIPTION

[0009] The specific embodiments of the present application are described below to facilitate the understanding of the present application by those skilled in the art, but it should be clear that the present application is not limited to the scope of the specific embodiments, and for those skilled in the art, it is obvious that various changes are within the spirit and scope of the present application defined and determined by the appended claims, and all the inventions utilizing the concept of the present application are within the scope of protection.

[0010] As shown in the figure, a SAR image change detection method for adaptively eliminating noise features comprises the following steps: Figure 1 Collect two original images to be detected, and use a logarithmic ratio operator and a two-stage fuzzy C-means clustering method to process data to generate a dual-time image block.

[0011] In this embodiment, a logarithmic ratio operator and a two-stage fuzzy C-means clustering method are used for data processing to obtain high-precision pseudo labels, so as to select high-quality training samples, and the operation process is as follows: Collect two original images to be detected.

[0012] The two original images to be detected are processed by using a logarithmic ratio operator method to generate a difference image, that is:

[0013] Among them, represents the difference image, represents a logarithmic function, , respectively represent the two original images to be detected.

[0014] The difference image is subjected to Gabor wavelet transform to extract the local spatial context features of the difference image, and the Gabor features of the difference image are generated, that is:

[0015] Among them, represents the Gabor features of the difference image, represents convolution calculation, represents a Gabor convolution kernel.

[0016] ​In this embodiment, in order to extract features conducive to clustering, so as to provide as much discriminative information as possible to facilitate change detection, the Gabor wavelet transform method is used to extract the local spatial context features of the difference map.

[0017] The Gabor features of the difference map are pre-classified by using a two-stage fuzzy C-means clustering method, to generate double-phase image blocks.

[0018] In this embodiment, the two-stage fuzzy C-means clustering method (two-stage FCM) is used to achieve three classification from coarse to fine, to ensure the accuracy and diversity of the samples; the operation process is as follows: Based on the Gabor features of the difference map, the fuzzy C-means clustering method is used to coarsely classify the difference map, to generate the change class and the unchanged class of the difference map.

[0019] The number of pixels in the change class is counted, and the upper bound of the pixels in the change class is calculated, that is:

[0020] wherein, represents the upper bound of the pixels in the change class, represents a constant value, represents the number of pixels in the change class.

[0021] In this embodiment, based on the Gabor features of the difference map, the two-stage fuzzy C-means clustering method is used to coarsely classify the difference map, which is divided into a change class and an unchanged class, denoted as , The number of pixels in the change class is denoted as , and the upper bound of the pixels in the change class is defined as In the experiment, the value is 1.25; therefore, the benefits of this step are: in the change detection task, the number of pixels in the change class is relatively small, coarse classification is performed to preferentially ensure that there are pixels in the change class, and at the same time, it is also necessary to avoid too many error class pixel quantities.

[0022] Based on the Gabor features of the difference map and the upper bound of the pixels in the change class, the fuzzy C-means clustering method is used to finely classify the difference map, which is divided into classes, and the classes are sorted in descending order according to the average value of the pixels in the classes, and the pixels in the classes sorted in descending order are counted, that is:

[0023]

[0024] wherein, , , These represent the first, second, and third classes sorted in descending order, respectively. Each class , , These represent the first, second, and third classes sorted in descending order, respectively. The number of pixels in each class; where, in the experiment The value is 10.

[0025] In this embodiment, based on the Gabor features of the difference map and the upper bound of pixels in the change class, the fuzzy C-means clustering method is executed again to divide the difference map into ten classes, namely... , … Furthermore, these ten classes are sorted in descending order of their average value, meaning that in the class... The pixels in the class have the highest average value. The pixels in the cluster have the smallest average value, and the larger the average value, the greater the probability of variation in the cluster; meanwhile, the number of pixels in the ten clusters is defined as follows: , … In order to put in the class Pixels in the prediction change class Therefore, the benefits of this step are: by more accurately classifying and statistically analyzing the pixels of the changing and unchanged classes, the accuracy of pseudo-labels is improved, and the selection of training samples is also guaranteed.

[0026] Traverse the sorted data in descending order. Each class assigns all pixels in the first class (sorted in descending order) to the predicted change class. In the middle, both the class loop variable and the pixel loop variable within the class are set, that is:

[0027]

[0028] in, Represents a loop variable. Represents a pixel loop variable within the class. Indicates the order of descending sorting. The number of pixels in each class.

[0029] Check if the pixel loop variable within the currently traversed class is less than the upper bound of the pixels in the changed class. If so, sort the pixels in the currently traversed class in descending order. Class All pixels are assigned to the predicted uncertainty class. Otherwise, sort the current loop in descending order. classes all pixels in the predicted unchanged class all pixels in the predicted unchanged class ; all pixels in the predicted unchanged class all pixels in the predicted unchanged class all pixels in the predicted unchanged class

[0030] In this embodiment, the classes are divided into ten classes by setting the class loop variable and the pixel loop variable in the class, and the pixels in the ten classes are sequentially assigned to the corresponding classes, so that the image with labels is output to represent the pre-classification change map, that is all pixels in the predicted unchanged class all pixels in the predicted unchanged class all pixels in the predicted unchanged class In addition, Figure 2 The pseudo labels and the true labels of the Ottawa dataset are shown, and it can be seen from Figure 2 that Figure 2 in (a), the white pixels belong to the predicted change class , the black pixels belong to the predicted unchanged class , and the gray pixels belong to the predicted uncertain class ; compared with the reference map in Figure 2 (b), the change class and the unchanged class in the pseudo labels are mostly correct, so that the reliability and diversity of the samples can be ensured by selecting the training samples according to the above method.

[0031] The SAR image change detection model is constructed, which includes a multi-scale feature module, a first convolutional down-sampling module, an adaptive noise feature elimination module, a second convolutional down-sampling module, a convolutional module and a classifier.

[0032] In this embodiment, the structure and connection relationship of the SAR image change detection model are as shown in Figure 3 The output of the multi-scale feature module is connected with the input of the first convolutional down-sampling module, the output of the first convolutional down-sampling module is connected with the input of the adaptive noise feature elimination module, the output of the adaptive noise feature elimination module is connected with the input of the second convolutional down-sampling module, the output of the second convolutional down-sampling module is connected with the input of the convolutional module, and the output of the convolutional module is connected with the input of the classifier. The structure and connection relationship of the multi-scale feature module are as shown in Figure 4As shown, the kernels include 1×1, 3×3, 5×5, and 7×7 convolutional kernels. In the above steps, the size of the resulting dual-temporal image patch is 16×16×2. This dual-temporal image patch is then input into four convolutional kernels of different scales: 1×1, 3×3, 5×5, and 7×7, for feature extraction at different scales. The number of convolutional kernels at each scale is 8 (i.e., the number of channels), and their stride is... s The image fill size is 1. p The values ​​are set to 0, 1, 2, and 3 respectively to obtain a feature map of size 16×16 and 8 channels at each scale. Finally, the feature maps obtained from the convolution operations at the four scales are superimposed (feature fusion) by extending the channel axis to obtain a multi-scale feature map; the feature fusion formula is: , Represents the fused feature map. , , , These represent feature maps at four different scales; The first convolutional downsampling module uses a 4×4 convolutional kernel with a stride of 2 and image padding of 1 to perform convolution operations. Specifically, it fuses the feature maps... The input is processed by the first convolutional downsampling module to generate the first downsampled feature map. And its size is , , These are the length and width of the feature map, respectively. Indicates the number of channels.

[0033] The structure and connection relationships of the adaptive noise feature cancellation module are as follows: Figure 5 As shown, it includes two 3×3 convolutional kernels and an adaptive thresholding module; the adaptive thresholding module consists of a two-layer perceptron (FC), the first layer of which introduces a leaky ReLU function (…). The second-layer perceptron introduces a sigmoid function ( ); at the same time, from Figure 5 It can also be seen that a residual connection structure is introduced. By adding the feature map with filtered noise and redundancy to the first downsampled feature map of the initial input through the residual connection, not only can the gradient vanishing effect be alleviated and the training efficiency of the model be improved, but the filtering of important features in the noise feature elimination process can also be compensated, thereby obtaining the noise-eliminated feature map.

[0034] Specifically, the process of noise removal by the adaptive noise feature removal module on the first downsampled feature map is as follows: The first down-sampling feature map is sequentially input into two 3*3 convolution kernels of the adaptive noise feature elimination module, and two convolution operations are performed to generate a first convolution feature map.

[0035] In this embodiment, the first down-sampling feature map is sequentially input into two 3*3 convolution kernels for two convolution operations, and the padding is 1 and the moving step is 1, and finally a first convolution feature map is generated. .

[0036] The first convolution feature map is input into the adaptive threshold module to calculate the adaptive threshold, specifically: After taking the absolute value of each channel map in the first convolution feature map, global average pooling is performed to generate a pooling vector of aggregated information , that is:

[0037] wherein, represents the aggregated information of the channel map , that is, the component of the pooling vector of aggregated information , , respectively represent the length and width of the first convolution feature map, represents the feature of the channel map with a length of and a width of .

[0038] According to the pooling vector of aggregated information , the adaptive threshold is calculated, that is:

[0039] wherein, represents the adaptive threshold, represents the S-shaped function, represents the first layer perceptron operation, represents the second layer perceptron operation, represents the relu function with leakage, represents element-wise multiplication.

[0040] In this embodiment, the process of inputting the first convolution feature map into the adaptive threshold module to obtain the adaptive threshold is as follows: first, taking the absolute value of each channel map in the first convolution feature map, global average pooling is performed to obtain the aggregated information of each channel map, and the aggregated information is multiplied element by element to generate a pooling vector of aggregated information, and then according to the pooling vector of aggregated information, the adaptive threshold is calculated, so that the subsequent step utilizes the calculated adaptive threshold to perform soft threshold activation on the first convolution feature map to filter noise and redundant features; the filtering process is as follows: Based on the calculated threshold, noise and redundant features are filtered from the first convolutional feature map to generate a filtered first convolutional feature map; the formula for filtering noise and redundant features from the first convolutional feature map is as follows:

[0041] in, This represents the soft thresholding function of the first convolutional feature map. This represents the first convolutional feature map. This indicates the condition for judgment.

[0042] The filtered first convolutional feature map and the first downsampled feature map are added together through a residual connection to obtain the noise-reduced feature map, i.e.:

[0043]

[0044] in, Represents the noise-reduced feature map. This represents the first convolutional feature map after filtering. This represents the first downsampled feature map. Represents a symbolic function. This represents the first convolutional feature map. This indicates taking the maximum value. This indicates the adaptive threshold calculation process.

[0045] In this embodiment, theoretically, the soft threshold function is: In actual operation, in order to utilize This whole thing expresses the theory through the actual operation of code. .

[0046] In this module, the second convolutional downsampling module uses 4×4 convolutional kernels with a stride of 2 and image padding of 1 to perform the convolution operation. Specifically, it de-noises the feature map... The input is processed by the second convolutional downsampling module to generate the second downsampled feature map. ; The convolution module includes two 2×2 convolution kernels, which process the second downsampled feature map obtained above. Perform two convolution operations to obtain a two-dimensional convolutional feature map. ; The classifier employs the normalized exponential activation function (softmax), which utilizes the normalized exponential activation function to apply the two-dimensional convolutional feature map. By performing mapping, we obtain the classification probabilities for each category, that is, the probability that a sample belongs to one of the two categories (the changed class and the unchanged class).p With 1-p .

[0047] In addition, the SAR image change detection model needs to be trained by the objective function through the back propagation method, and the cross-entropy is generally used as the objective function for the classification problem, that is:

[0048] Wherein, The cross-entropy loss function is represented by, The number of training samples is represented by, The class label of the th sample is represented by, the change class is 1, and the non-change class is 0, The logarithmic function is represented by, The predicted class probability of the th sample is represented by.

[0049] In summary, the SAR image change detection method for adaptive noise feature elimination proposed in the application adopts the logarithmic ratio operator and the two-level fuzzy C-means clustering method to process the training data, avoids manual class label annotation, is beneficial to selecting high-quality training samples, and thus improves the precision of the SAR image change detection model for image change detection; at the same time, four different scale convolution kernels are introduced to extract features from the input image, which better expresses the complex change information in the image from multiple scales and retains the features conducive to improving the image change detection precision, thereby improving the image class prediction probability, that is, improving the precision of the image change detection; in addition, the adaptive threshold is calculated, and the residual connection is used, which not only eliminates the noise and redundant features in the feature map, but also improves the precision of the feature map, and finally improves the precision of the image change detection.

[0050] The principles and implementation modes of the application are described in the specific embodiments, and the above embodiment descriptions are only used to help understand the method and core idea of the application; at the same time, for those skilled in the art, according to the idea of the application, the specific implementation modes and application ranges will be changed, and the above description should not be understood as limiting the application.

[0051] Those skilled in the art will realize that the embodiments described herein are for the purpose of helping the reader to understand the principles of the application and should be understood as not limiting the protection scope of the application to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations according to the technical inspirations disclosed in the application without departing from the essence of the application, and these modifications and combinations are still within the protection scope of the application.

Claims

1. A method for detecting changes in SAR images with adaptive noise reduction features, characterized in that, Includes the following steps: Two original images to be detected were collected, and data processing was performed using the log ratio operator and two-level fuzzy C-means clustering method to generate dual-temporal image patches; A SAR image change detection model is constructed, which includes a multi-scale feature module, a first convolutional downsampling module, an adaptive noise feature elimination module, a second convolutional downsampling module, a convolutional module, and a classifier; The dual-temporal image blocks are input into the multi-scale feature module for multi-scale feature extraction, and the extracted multi-scale feature maps are fused to generate a fused feature map. The fused feature map is input into the first convolutional downsampling module for downsampling to generate the first downsampled feature map; The first downsampled feature map is input into the adaptive noise feature elimination module, and noise is eliminated through soft thresholding to generate a noise-eliminated feature map. The noise-reduced feature map is input into the second convolutional downsampling module for downsampling to generate the second downsampled feature map. The second downsampled feature map is input into the convolution module to perform a two-dimensional convolution operation, generating a two-dimensional convolutional feature map; The two-dimensional convolutional feature map is input into the classifier and mapped through the normalized exponential activation function to obtain the classification probability of each category.

2. The SAR image change detection method with adaptive noise cancellation features according to claim 1, characterized in that, The process of collecting two original images to be detected and processing them using the logarithmic ratio operator and two-level fuzzy C-means clustering method to generate dual-temporal image patches is as follows: Collect two original images to be detected; The two original images to be detected are processed using the logarithmic ratio operator to generate a difference map, i.e.: in, Representing a difference graph, Represents the logarithmic function. , These represent two original images to be detected; The difference map is subjected to Gabor wavelet transform to extract local spatial context features, generating Gabor features of the difference map, i.e.: in, This represents the Gabor features of the difference plot. This represents convolution calculation. Indicates the Gabor convolution kernel; A two-level fuzzy C-means clustering method is used to pre-classify the Gabor features of the difference map and generate dual-temporal image patches.

3. The SAR image change detection method with adaptive noise cancellation features according to claim 2, characterized in that, The process of pre-classifying Gabor features of the difference map and generating dual-temporal image patches using a two-level fuzzy C-means clustering method is as follows: Based on the Gabor features of the difference map, the fuzzy C-means clustering method is used to coarsely classify the difference map, generating the changed class and the unchanged class of the difference map; Count the number of pixels in the change class, and calculate the upper bound of the pixels in the change class, i.e.: in, This represents the upper bound of pixels in the variation class. Indicates a constant value. Indicates the number of pixels in the change class; Based on the Gabor features of the difference map and the upper bound of pixels in the change class, the fuzzy C-means clustering method is used to further classify the difference map, dividing it into... There are several classes, and the values ​​of the pixels in each class are adjusted accordingly. Sort each class in descending order, and simultaneously calculate the results of the descending sort. The pixels in each class, namely: in, , , These represent the first, second, and third classes sorted in descending order, respectively. Each class , , These represent the first, second, and third classes sorted in descending order, respectively. The number of pixels in each class; Traverse the sorted data in descending order. Each class assigns all pixels in the first class (sorted in descending order) to the predicted change class. In the middle, both the class loop variable and the pixel loop variable within the class are set, that is: in, Represents a loop variable. Represents a pixel loop variable within the class. Indicates the order of descending sorting. The number of pixels in each class; Check if the pixel loop variable within the currently traversed class is less than the upper bound of the pixels in the changed class. If so, sort the pixels in the currently traversed class in descending order. Class All pixels are assigned to the predicted uncertainty class. Otherwise, sort the current loop in descending order. Class All pixels are assigned to the predicted unchanged class. In the end, a pre-classified change map with labels is obtained, namely: ； Based on the predicted change class Compared with the predicted unchanged class Centered on all pixels, training samples are selected to generate dual-temporal image patches.

4. The SAR image change detection method with adaptive noise cancellation features according to claim 1, characterized in that, The multi-scale feature modules include 1×1 convolution kernels, 3×3 convolution kernels, 5×5 convolution kernels, and 7×7 convolution kernels.

5. The SAR image change detection method with adaptive noise cancellation features according to claim 1, characterized in that, Both the first and second convolutional downsampling modules have 4×4 convolutional kernels.

6. The SAR image change detection method with adaptive noise cancellation features according to claim 1, characterized in that, The adaptive noise feature elimination module includes two 3×3 convolution kernels and an adaptive threshold module.

7. The SAR image change detection method with adaptive noise cancellation features according to claim 6, characterized in that, The first downsampled feature map is input into the adaptive noise feature cancellation module. Noise cancellation is performed through soft thresholding, and the process of generating a noise-cancelled feature map is as follows: The first downsampled feature map is sequentially input into two 3×3 convolution kernels of the adaptive noise feature elimination module for two convolution operations to generate the first convolution feature map; The first convolutional feature map is input into the adaptive thresholding module to calculate the adaptive threshold, specifically as follows: After taking the absolute values ​​of each channel map in the first convolutional feature map, global average pooling is performed to generate a pooling vector of aggregated information. ,Right now: in, Channel diagram The aggregated information, i.e., the pooling vector of the aggregated information. The amount, , Let the length and width of the first convolutional feature map be represented respectively. Indicates length is Width Channel diagram Features; Based on the pooling vector of aggregated information Calculate the adaptive threshold, i.e.: in, Indicates an adaptive threshold. Represents a sigmoid function. This represents the first-layer perceptron operation. This indicates the operation of the second-layer perceptron. This refers to a ReLU function with leakage. This indicates element-wise multiplication; Based on the calculated threshold, noise and redundant features are filtered from the first convolutional feature map to generate a filtered first convolutional feature map. The filtered first convolutional feature map and the first downsampled feature map are added together through residual connection to obtain the noise-reduced feature map.

8. The SAR image change detection method with adaptive noise cancellation features according to claim 7, characterized in that, The formula for filtering noise and redundant features from the first convolutional feature map is as follows: in, This represents the soft thresholding function of the first convolutional feature map. This represents the first convolutional feature map. This indicates the condition for judgment.

9. The SAR image change detection method with adaptive noise cancellation features according to claim 8, characterized in that, The formula for obtaining the noise-reduced feature map by adding the filtered first convolutional feature map and the first downsampled feature map through a residual connection is as follows: in, Represents the noise-reduced feature map. This represents the first convolutional feature map after filtering. This represents the first downsampled feature map. Represents a symbolic function. This represents the first convolutional feature map. This indicates taking the maximum value. This indicates the adaptive threshold calculation process.

10. The SAR image change detection method with adaptive noise cancellation features according to claim 1, characterized in that, The convolution module consists of two 2×2 convolution kernels.

Images