An image super-resolution method driven by remote sensing target recognition task

Abstract

The present invention belongs to the technical field of image enhancement and image recognition, and provides an image super-resolution method driven by remote sensing target recognition tasks. For remote sensing target recognition tasks, the quality of low-resolution images is improved by image enhancement technology. The method first obtains a low-resolution slice image of a remote sensing target, and then uses an SRGAN generative network for upsampling to generate a high-resolution image. To improve image clarity, a cascade refinement subnetwork is used to restore detail information. Fine-grained recognition branches and multi-task classification loss functions are added to the discriminative network to enhance the ability to distinguish small targets from the background, and guide the discriminative network to distinguish between super-resolution and true high-resolution images, as well as small targets and ground background interferences. The present invention uses the loss of remote sensing target recognition as the target perception loss to guide network training, so that super-resolution images pay more attention to target reconstruction, improve image quality and distinguishability features, and enhance remote sensing target recognition performance.

Description

Technical Field

[0001] The invention belongs to the technical field of image enhancement and image recognition, and in particular relates to an image super-resolution method driven by a remote sensing target recognition task. Background Art

[0002] Image super-resolution belongs to image restoration and enhancement algorithms, such as denoising and deblurring, and has been widely studied in the fields of computer vision and graphics. The recognition of small targets in low-resolution remote sensing images is a difficult problem in remote sensing image interpretation. The lack of visual information makes it difficult for small targets to extract strong distinguishing features based on their own information for classification and recognition with background or other targets. Image super-resolution is a feasible way to solve this problem. However, traditional super-resolution methods aim at the detailed texture features of natural scene images, attach importance to image quality that meets the human visual system, but ignore the strong discriminant information required for downstream tasks such as image recognition.

[0003] Secondly, compared with natural scene images, the objects in remote sensing images have arbitrary postures and lighting changes, and the super-resolution of remote sensing scene images is more challenging. Therefore, an image super-resolution network driven by target recognition tasks and aware of target content is needed to generate clear, sharp and recognizable high-resolution images to improve the performance of downstream image recognition tasks. Summary of the invention

[0004] In view of this, the present invention proposes an image super-resolution method driven by remote sensing target recognition tasks, which is used to solve the problem that the image super-resolution task ignores the strong discriminative information required by downstream tasks and the recognition of small targets in low-resolution remote sensing images. By organically combining the network architecture and the loss function, the super-resolution network is driven to generate super-resolution images with high resolution and strong discriminative information, and the low-resolution small target features are mapped into features similar to those of high-resolution targets, thereby improving the downstream recognition task's ability to distinguish small targets and background interferences. The technical solution of the present invention is:

[0005] An image super-resolution method driven by remote sensing target recognition task includes the following steps:

[0006] Step 1: Obtain at least one low-resolution remote sensing target slice image with an image size of , perform standard deviation normalization on the target slice image;

[0007] Step 2: Improve the generative network of SRGAN so that the improved generative network can upsample the input low-resolution target slice image and perform deblurring operations to finally obtain a clear high-resolution reconstructed image;

[0008] Step 3: Add a fine-grained recognition branch to the discriminative network of SRGAN. The fine-grained recognition branch is used to enhance the ability to distinguish small targets from ground background interferences;

[0009] Step 4: Design a multi-task classification loss function for joint learning, drive the super-resolution network to generate super-resolution images with high resolution and strong discriminant information, and make the super-resolution reconstruction and the target precise classification task complementary in the training process. The multi-task classification loss function consists of content loss function, adversarial loss function and classification loss function.

[0010] As a further preferred solution of the image super-resolution method driven by remote sensing target recognition task of the present invention, the step 2 specifically comprises the following steps:

[0011] Step 2.1, the BN layer is deleted in the residual module of the generative network G in SRGAN;

[0012] Step 2.2, because the pixels of the low-resolution small target itself are too small, resulting in a lack of detail information, coupled with the influence of the mean square error loss function itself, the super-resolution reconstructed image generated by the generative network often becomes blurred. This situation may affect the subsequent discriminative network's discrimination of the true high-resolution image and the super-resolution image, as well as the small target and the ground background interference. In order to solve this problem, the present invention cascades a refinement subnetwork in the generative network G for deblurring.

[0013] The main function of the refinement subnetwork is to perform a deblurring operation by comparing and analyzing the difference between the high-resolution reconstructed image and the true high-resolution image, thereby improving the quality of the reconstructed image. After removing the blur effect, the network can significantly improve the clarity of the target edge and the overall image, and further improve the discrimination ability of the subsequent discriminant network and the ability of the target precision classification network to remove ground background interference.

[0014] As a further preferred solution of the image super-resolution method driven by remote sensing target recognition task of the present invention, step 3 specifically comprises the following steps:

[0015] Step 3.1: To avoid losing important feature information of small objects after too much downsampling, the discriminant network of SRGAN The network removes the maximum pooling operation in the "Conv5" layer of the VGG16 network;

[0016] Step 3.2, for the discriminative network of SRGAN It has multi-task capabilities, replacing the three fully connected layers in the VGG16 network with a parallel branch consisting of a fully connected layer and bilinear pooling. The branch where the fully connected layer is located is responsible for distinguishing between real high-resolution images and super-resolution images, and the branch where the bilinear pooling is located is responsible for distinguishing between small targets and ground background interference.

[0017] As a further preferred solution of the image super-resolution method driven by remote sensing target recognition task of the present invention, step 4 specifically comprises the following steps:

[0018] Step 4.1, retain the pixel-by-pixel mean square error loss function in the content loss function of SRGAN as the content loss function, remove the VGG feature matching loss function of SRGAN to reduce the computational complexity, and the calculation method of the pixel-by-pixel mean square error loss is as follows:

[0019]

[0020] in, and Respectively represent A low-resolution blurred image and a true high-resolution image; express The image goes through the upsampling subnetwork The image generated after processing; Representing images After upsampling sub-network and refinement sub-network The super-resolution image generated later; and Represent the upsampling subnetwork and the refinement subnetwork respectively; parameters The number of input images; express Norm, used to obtain the mean square error loss;

[0021] Step 4.2, in order to reduce the visual difference between the real image and the super-resolution image, the adversarial loss function in the GAN network is added to the target loss function; the adversarial loss function is defined as follows:

[0022]

[0023] in, Generally, 2, the natural constant e, and 10 are taken as bases; represents the discriminative network; Represents super-resolution reconstructed image is the probability of a real high-resolution image; the generative network learns the distribution of real high-resolution image data, making the super-resolution reconstructed image generated by itself more realistic. The above process drives the generation of reconstructed images that are sufficiently similar to the real high-resolution image through the adversarial loss function;

[0024] Step 4.3, the discriminative network is a multi-task classification network that can not only distinguish high-resolution images and generated reconstructed images, but also identify whether the candidate area is an aircraft target; the classification loss function is defined as follows:

[0025]

[0026] in, Generally, 2, the natural constant e, and 10 are taken as bases; Representing images Corresponding label information; when When the image For small aircraft targets, When the image The classification loss function is used to guide the discriminative network to learn to distinguish between real high-resolution images and super-resolution reconstructed images, as well as small targets and ground background distractors, so as to promote the generative network to reconstruct clearer images for easier target classification.

[0027] Step 4.4, combine the above content loss function, adversarial loss function and classification loss function to design a multi-task classification loss function, which is defined as follows:

[0028]

[0029] in and are weight parameters, and their default values ​​are , .

[0030] The beneficial effects achieved by the present invention are:

[0031] The present invention aims at the problem that traditional super-resolution methods mainly focus on improving the image quality of blurred and low-resolution images, rather than constructing high-resolution images with strong discriminant information for downstream recognition tasks, and designs an image super-resolution method driven by remote sensing target recognition tasks. The method uses SRGAN as the benchmark model and includes two parts: a generative network and a discriminative network. Unlike SRGAN, the generative network includes two components: an upsampling subnetwork and a refinement subnetwork. The function of the upsampling subnetwork is to upsample the low-resolution image to obtain a high-resolution reconstructed image. The refinement subnetwork is based on comparing the difference between the high-resolution reconstructed image and the true high-resolution image, restoring the missing detail information in the upsampled image, thereby generating a clear high-resolution image for classification by the discriminative network. In addition, the present invention adds a fine-grained recognition branch and a multi-task classification loss function to the discriminative network of SRGAN.

[0032] In order to complement the information of super-resolution task and target recognition task, the present invention designs an image super-resolution method driven by remote sensing target recognition task. The advantages of the present invention are: through the organic combination of network architecture and loss function, the super-resolution image generated by driving the super-resolution network pays more attention to the reconstruction of the target itself, rather than the irrelevant background area, so that the reconstructed super-resolution image not only has clear image quality, but also contains distinguishable features, thereby improving the performance of remote sensing target recognition. In addition, the invention can also map the low-resolution small target features into features similar to high-resolution targets, thereby improving the classification and recognition capabilities of small targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 Shown is a block diagram of the overall idea of ​​the method of the present invention.

[0034] Figure 2 Shown is a schematic diagram of the generated network structure of the method of the present invention.

[0035] Figure 3 Shown is a schematic diagram of the residual module with the BN layer removed.

[0036] Figure 4 Shown is a schematic diagram of the residual module in the SRGAN network structure.

[0037] Figure 5 Shown is the discriminant network structure diagram of the method of the present invention.

[0038] Figure 6 Shown is the VGG16 network structure diagram. DETAILED DESCRIPTION

[0039] The technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the present invention. In addition, the forms of the various structures recorded in the following embodiments are merely illustrative, and the present invention is not limited to the various structures recorded in the following embodiments. All other embodiments obtained by ordinary technicians in this field without making creative work belong to the scope of protection of the present invention.

[0040] Reference Figure 1-Figure 6 The image super-resolution method based on remote sensing target recognition task driving of the present invention is specifically described. Figure 1-Figure 6 The structure of each network is described.

[0041] Figure 1 The figure shows the overall idea of ​​the method of the present invention. First, a pair of low-resolution images are used as input, and the pair of low-resolution images are respectively the aircraft target and the background slice, and then sent to the generative network for image reconstruction to generate a pair of high-resolution reconstructed images. Subsequently, the pair of high-resolution reconstructed images and the corresponding pair of real high-resolution images are sent to the discriminative network to discriminate between the real high-resolution image and the reconstructed image, and the background slice and the aircraft target.

[0042] Figure 2 The figure shows a schematic diagram of the generative network structure of the method of the present invention. The generative network consists of an upsampling subnetwork and a refinement subnetwork. In the upsampling subnetwork structure, from left to right are convolutional layer, 8 residual modules, convolutional layer, 2 deconvolutional layers and convolutional layer, where the residual modules are convolutional layer, PReLU layer, convolutional layer and point-by-point addition from left to right; the refinement subnetwork is similar to the upsampling subnetwork structure, which is convolutional layer, 8 residual modules, convolutional layer, convolutional layer, convolutional layer and convolutional layer from left to right. The network structure of the above residual module is the same as that of the upsampling subnetwork. Figure 3 Consistent.

[0043] Figure 3 The figure shows the residual module after removing the BN layer. The network structure from left to right is convolutional layer, PReLU layer, convolutional layer and point-by-point addition.

[0044] Figure 4 The figure shows the residual module in SRGAN. The network structure from left to right is convolutional layer, BN layer, PReLU layer, convolutional layer, BN layer and point-by-point addition.

[0045] Figure 5The figure shows the discriminant network structure diagram of the method of the present invention. The network structure takes the true image and the reconstructed image as input, and then successively enters the "Conv1" layer, "Conv2" layer, "Conv3" layer, "Conv4" layer, and "Conv5" layer for feature extraction. It is worth noting that the "Conv5" layer is composed of 5 convolutions with 512 channels and a size of 3x3, and there is no pooling operation. The "Conv5" layer then connects two branches, one of which is subjected to bilinear pooling for feature aggregation, and then sent to the sigmoid function for classification and discrimination, and finally obtains the classification of aircraft and background categories; the other branch is subjected to the full connection layer for classification and discrimination, and is sent to the sigmoid function for classification and discrimination, and finally obtains the classification and discrimination of the reconstructed image and the true image.

[0046] Figure 6 The figure shows the VGG16 network structure. From left to right, the VGG16 network structure is "Conv1", "Conv2", "Conv3", "Conv4", "Conv5", 3 fully connected layers, Softmax discrimination, and output probability. Among them, the "Conv5" layer is composed of 5 convolutions with 512 channels and a size of 3x3 and a Maxpool pooling operation.

[0047] The overall idea of ​​this method is as follows Figure 1 As shown in the figure, a pair of low-resolution images are first used as input, which are the aircraft target and the background slice. Then they are sent to the generative network for image reconstruction to generate a pair of high-resolution reconstructed images. Subsequently, the pair of high-resolution reconstructed images and the corresponding pair of real high-resolution images are sent to the discriminative network to discriminate between the real high-resolution image and the reconstructed image, and the background slice and the aircraft target.

[0048] The image super-resolution method based on remote sensing target recognition task-driven of the present invention specifically comprises the following steps:

[0049] Step 1: Obtain at least one low-resolution remote sensing target slice image with an image size of , perform standard deviation normalization on the target slice image;

[0050] Step 2: The generative network structure proposed by the present invention is as follows Figure 2As shown in the figure, it can be seen that the generative network consists of an upsampling subnetwork and a refinement subnetwork. In the upsampling subnetwork structure, from left to right are convolutional layers, 8 residual modules, convolutional layers, 2 deconvolutional layers, and convolutional layers. The residual modules are convolutional layers, PReLU layers, convolutional layers, and point-by-point addition from left to right. The refinement subnetwork is similar to the upsampling subnetwork structure, which is convolutional layers, 8 residual modules, convolutional layers, convolutional layers, convolutional layers, and convolutional layers from left to right. The network structure of the residual module is the same as that of the upsampling subnetwork. Figure 3 The purpose of adding the refinement sub-network is to enable the improved generative network to not only upsample the input low-resolution target slice image, but also perform deblurring operations, and finally obtain a clear high-resolution reconstructed image. The specific sub-steps are:

[0051] (2.1) The residual module in the SRGAN network is as follows Figure 4 As shown in the figure, the network structure from left to right is convolutional layer, BN layer, PReLU layer, convolutional layer, BN layer and point-by-point addition. In order to avoid artifacts caused by the BN layer during training under the GAN network and reduce the time complexity of the network, the present invention deletes the BN layer in the residual module of the generative network G in SRGAN. The residual module with the BN layer deleted is shown in Figure 3 As shown in the figure, the network structure from left to right is convolutional layer, PReLU layer, convolutional layer and point-by-point addition. Figure 4 compared to, Figure 3 Only the BN layer is missing, and the main network structure remains unchanged;

[0052] (2.2), because the pixels of the low-resolution small target itself are too small, resulting in a lack of detail information, coupled with the influence of the mean square error loss function itself, the super-resolution reconstructed image generated by the generative network often becomes blurred. This situation may affect the subsequent discriminative network's discrimination of the real high-resolution image and the super-resolution image, as well as the small target and the ground background interference. In order to solve this problem, the present invention cascades a refinement subnetwork in the generative network G for deblurring.

[0053] The main function of the refinement subnetwork is to perform a deblurring operation by comparing and analyzing the difference between the high-resolution reconstructed image and the true high-resolution image, thereby improving the quality of the reconstructed image. After removing the blur effect, the network can significantly improve the clarity of the target edge and the overall image, and further improve the discrimination ability of the subsequent discriminant network and the ability of the target precision classification network to remove ground background interference.

[0054] Table 1 Generative network parameter configuration

[0055]

[0056] Table 1 shows the specific parameters of the two sub-networks of upsampling and refinement in detail. In Table 1, “Conv” represents a convolutional layer, “Res×8” represents 8 residual modules, and “DeConv” represents a deconvolutional layer.

[0057] Step 3: In the classic SRGAN super-resolution network, the discriminant network It is a two-class classification model used to distinguish between real images and reconstructed images. The VGGNet network is used as the main framework. This is similar to the bilinear CNN network structure used for fine-grained image recognition tasks. Bilinear CNN captures the pairwise correlation between features by modeling the second-order statistical information of features and generates discriminative features, which makes bilinear CNN have certain advantages in small object recognition.

[0058] Inspired by the above, the present invention designs a new discriminant network , which combines the functions of the discriminative network and bilinear CNN in SRGAN. This means that the discriminative network can not only distinguish between super-resolution images and real images, but also distinguish between small targets and similar interference objects. Figure 5 As shown in the figure, its network structure takes the true image and the reconstructed image as input, and then successively enters the "Conv1" layer, "Conv2" layer, "Conv3" layer, "Conv4" layer, and "Conv5" layer for feature extraction. It is worth noting that the "Conv5" layer is composed of 5 convolutions with 512 channels and a size of 3x3, and there is no pooling operation. The "Conv5" layer then connects two branches, one of which is subjected to bilinear pooling for feature aggregation and then sent to the sigmoid function for classification and discrimination, and finally obtains the classification of aircraft and background categories; the other branch is subjected to the fully connected layer for classification and discrimination, and is sent to the sigmoid function for classification and discrimination, and finally obtains the classification and discrimination of the reconstructed image and the true image. From the above description, it can be seen that the discriminant network The network structure takes the VGG16 network as the baseline model and improves it.

[0059] Figure 6The network structure of VGG16 is shown in the figure. From left to right, the network structure of VGG16 is "Conv1", "Conv2", "Conv3", "Conv4", "Conv5", 3 fully connected layers, Softmax discrimination, and output probability. Among them, the "Conv5" layer is composed of 5 convolutions with 512 channels and a size of 3x3 and a Maxpool pooling operation. It can be seen that the output layer uses the Softmax function to convert the output of the fully connected layer into the probability value of each category, which is convenient for multi-category classification. The discriminant network The network structure is based on the VGG16 network with the following modifications:

[0060] (3.1) To avoid the loss of important feature information of small objects after too much downsampling, the discriminant network of SRGAN The network removes the maximum pooling operation in the "Conv5" layer of the VGG16 network;

[0061] (3.2), for the discriminative network of SRGAN It has multi-task capabilities, replacing the three fully connected layers in the VGG16 network with a parallel branch consisting of a fully connected layer and bilinear pooling. The branch where the fully connected layer is located is responsible for distinguishing between real high-resolution images and super-resolution images, and the branch where the bilinear pooling is located is responsible for distinguishing between small targets and ground background interference.

[0062] Step 4: Design a multi-task classification loss function for joint learning, drive the super-resolution network to generate super-resolution images with high resolution and strong discriminant information, and make the super-resolution reconstruction and the target precise classification task complementary in the training process. The multi-task classification loss function consists of content loss function, adversarial loss function and classification loss function.

[0063] (4.1), retain the pixel-by-pixel mean square error loss function in the content loss function of SRGAN as the content loss function, remove the VGG feature matching loss function of SRGAN to reduce the computational complexity, and the calculation method of the pixel-by-pixel mean square error loss is as follows:

[0064]

[0065] in, and Respectively represent A low-resolution blurred image and a true high-resolution image; express The image goes through the upsampling subnetwork The image generated after processing; Representing images After upsampling sub-network and refinement sub-network The super-resolution image generated later; and Represent the upsampling subnetwork and the refinement subnetwork respectively; parameters The number of input images; express Norm, used to obtain the mean square error loss;

[0066] (4.2), in order to reduce the visual difference between the real image and the super-resolution image, the adversarial loss function in the GAN network is added to the target loss function; the adversarial loss function is defined as follows:

[0067]

[0068] in, Generally, 2, the natural constant e, and 10 are taken as bases; represents the discriminative network; Represents super-resolution reconstructed image is the probability of a real high-resolution image; the generative network learns the distribution of real high-resolution image data, making the super-resolution reconstructed image generated by itself more realistic. The above process drives the generation of reconstructed images that are sufficiently similar to the real high-resolution image through the adversarial loss function;

[0069] (4.3), the discriminative network is a multi-task classification network that can not only distinguish high-resolution images and generated reconstructed images, but also identify whether the candidate area is an aircraft target; the classification loss function is defined as follows:

[0070]

[0071] in, Generally, 2, the natural constant e, and 10 are taken as bases; Representing images Corresponding label information; when When the image For small aircraft targets, When the image The classification loss function is used to guide the discriminative network to learn to distinguish between real high-resolution images and super-resolution reconstructed images, as well as small targets and ground background distractors, so as to promote the generative network to reconstruct clearer images for easier target classification.

[0072] (4.4), combining the above content loss function, adversarial loss function and classification loss function, we design a multi-task classification loss function, which is defined as follows:

[0073]

[0074] in and are weight parameters, and their default values ​​are , .

[0075] Through the above steps, high-precision recognition of small targets in low-resolution remote sensing images can be achieved. This method organically combines the network architecture and loss function to drive the super-resolution network to generate super-resolution images with high resolution and strong discriminant information, and maps the low-resolution small target features into features similar to those of high-resolution targets, thereby improving the classification and recognition model's ability to discriminate small targets, background interferences, and other targets.

[0076] The present invention provides a new method and technical approach for the super-resolution method of low-resolution remote sensing images, and has great practical application value.

[0077] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. An image super-resolution method driven by remote sensing target recognition task, characterized in that: It includes: Step 1: Obtain at least one low-resolution remote sensing target slice image with an image size of 32 cm × 32 cm, and perform standard deviation normalization on the target slice image; Step 2: Improve the generative network of SRGAN so that the improved generative network can upsample the input low-resolution target slice image and perform deblurring operations to finally obtain a clear high-resolution reconstructed image; Step 3: A fine-grained recognition branch is added to the discriminative network of SRGAN to enhance the ability to distinguish small targets from ground background interferences; Step 4: Design a multi-task classification loss function for joint learning, drive the super-resolution network to generate super-resolution images with high resolution and strong discriminant information, and make the super-resolution reconstruction and the target precise classification task complementary in the training process; the multi-task classification loss function consists of content loss function, adversarial loss function and classification loss function; The generative network improvement of SRGAN described in step 2 includes the following steps: Step 2.1, the BN layer is deleted in the residual module of the generative network G in SRGAN; Step 2.2, a refinement sub-network is cascaded in the generative network G to perform a deblurring operation by comparing and analyzing the difference between the high-resolution reconstructed image and the true high-resolution image; Step 4 includes the following steps: Step 4.1, retain the pixel-by-pixel mean square error loss function in the content loss function of SRGAN as the content loss function, remove the VGG feature matching loss function of SRGAN to reduce the computational complexity, and the calculation method of the pixel-by-pixel mean square error loss is as follows: in, and Respectively represent A low-resolution blurred image and a true high-resolution image; express The image goes through the upsampling subnetwork The image generated after processing; Representing images After upsampling sub-network and refinement sub-network The super-resolution image generated later; and Represent the upsampling subnetwork and the refinement subnetwork respectively; parameters The number of input images; express Norm, used to obtain the mean square error loss; Step 4.2, in order to reduce the visual difference between the real image and the super-resolution image, the adversarial loss function in the GAN network is added to the target loss function; the adversarial loss function is defined as follows: in, Take 2, the natural constant e and 10 as bases; represents the discriminative network; Represents super-resolution reconstructed image is the probability of a real high-resolution image; the generative network learns the distribution of real high-resolution image data, making the super-resolution reconstructed image generated by itself more realistic; the above process drives the generation of reconstructed images that are sufficiently similar to the real high-resolution image through the adversarial loss function; Step 4.3, the discriminative network is a multi-task classification network that can not only distinguish high-resolution images and generated reconstructed images, but also identify whether the candidate area is an aircraft target; the classification loss function is defined as follows: in, Take 2, the natural constant e and 10 as bases; Representing images Corresponding label information; when When the image For small aircraft targets, When the image The classification loss function is used to guide the discriminative network to learn to distinguish between real high-resolution images and super-resolution reconstructed images, as well as small targets and ground background distractors, so as to promote the generative network to reconstruct clearer images for easier target classification. Step 4.4, combine the above content loss function, adversarial loss function and classification loss function to design a multi-task classification loss function, which is defined as follows: in and are weight parameters, and their default values ​​are , .

2. The image super-resolution method based on remote sensing target recognition task drive according to claim 1 is characterized in that: Step 3 includes the following steps: Step 3.1: To avoid losing important feature information of small objects after too much downsampling, the discriminant network of SRGAN The maximum pooling operation in the "Conv5" layer of the VGG16 network is removed; In step 3.2, the three fully connected layers in the VGG16 network are replaced by a parallel branch consisting of a fully connected layer and bilinear pooling. The branch where the fully connected layer is located is responsible for distinguishing between real high-resolution images and super-resolution images, and the branch where the bilinear pooling is located is responsible for distinguishing between small targets and ground background interferences.

Images