Target detection enhancement method and device based on generative adversarial architecture and storage medium

Abstract

The application discloses a target detection enhancement method and device based on a generative adversarial architecture and a storage medium, and relates to the technical field of target detection in computer vision. The application proposes a new target detection framework, which performs adversarial training on a target detection network that is difficult to further improve performance, can further improve the performance of the trained target detection network, does not increase parameters, can quickly and effectively improve the trained target detection network, has fast training speed, consumes few computing resources and is efficient, does not increase inference time or training difficulty as a cost, and is almost a plug-and-play training mode.

Description

Technical Field

[0001] This invention relates to a target detection enhancement method, apparatus, and storage medium based on generative adversarial architecture, belonging to the field of target detection technology in computer vision. Background Technology

[0002] Object detection is one of the fundamental problems in computer vision, and its derivative tasks, such as pedestrian detection, face recognition, and video surveillance detection, are receiving increasing attention. In recent years, with the enhancement of machine performance and the expansion of application areas and depth, researchers have begun to use numerous deep and complex convolutional layers to improve the network's ability to extract features from dense details in images. This directly leads to a sharp increase in computational resource consumption and difficulties in controlling the feature extraction network during training. To address these issues, a series of techniques have been proposed to improve network performance while avoiding a massive increase in network parameters. These include anchor-based and anchor-free models, greedy non-maximum suppression, learnable non-maximum suppression, residual connection networks, and feature pyramids. Furthermore, the ability of generative adversarial networks (GANs) to learn the mapping from the latent space to the real distribution allows them to tackle object detection tasks from a completely new perspective, especially for the detection of small objects and low-resolution images.

[0003] For Generative Adversarial Networks (GANs), once they obtain the target distribution for a task, they can potentially learn a matching method that is "close to the correct answer"—that is, pursuing a Nash equilibrium in a non-convex game with continuous high-dimensional parameters between the generator and discriminator. Since GANs typically use gradient descent to handle the losses of the generator and discriminator, this can cause them to get stuck at local lows in the artificially designed loss function, rather than at the Nash equilibrium point of the non-convex game. To address this issue, some studies have proposed feature matching methods for image generation seeking Nash equilibrium and semi-supervised training methods for object detection based on this approach. Other studies have proposed adding a detector as a third party to the two-way game between the generator and discriminator, forming a "Triple-GAN" with a three-way game.

[0004] Although this method carefully designs the adversarial training process (treating the generator and detector as the same party during discriminator training, and eliminating the latter and former during generator and detector training respectively), the discriminator's rejection of the detector's classification results for real images, and the strong coupling between the generator and detector that this method strives to achieve, will obviously hinder the detector's performance to some extent. Therefore, extending the two-party game to a three-party game is not a method conducive to stable training. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a target detection enhancement method, device and storage medium based on generative adversarial architecture, which solves the problem of difficulty in controlling the feature extraction network when the computer performs target detection.

[0006] To achieve the above objectives, the present invention is implemented using the following technical solution:

[0007] In a first aspect, the present invention provides a target detection enhancement method based on a generative adversarial architecture, the method comprising:

[0008] The trained CenterNet is separated into a feature extraction network and a classification network;

[0009] Construct a discriminator and train it adversarially with the feature extraction network to form an adversarial structure.

[0010] After the adversarial training is completed, the feature extraction network and the classification network are recombined. The feature extraction network is fixed, and the classification network is trained to match the feature extraction network.

[0011] In conjunction with the first aspect, furthermore, the construction of a discriminator, and the adversarial training of the discriminator and the feature extraction network to form an adversarial structure, includes:

[0012] Before adversarial training, CenterNet was fixed and the discriminator was trained for 50 epochs based on the COCO2017 dataset.

[0013] During adversarial training, high-quality image features are used as the target distribution to further enhance the feature extraction capability of the feature extraction network.

[0014] After the adversarial training is completed, the detection network that did not participate in the adversarial training is updated.

[0015] In conjunction with the first aspect, further, the construction of a discriminator, and the adversarial training of the discriminator and the feature extraction network to form an adversarial structure, also includes:

[0016] During adversarial training, a portion of images is extracted from the dataset, and image processing techniques are used to augment these images. The image data augmented using image processing techniques is the processed data batch; the image data without image processing is the original data batch.

[0017] In conjunction with the first aspect, the target enhancement further includes applying Gaussian blur to the background portion of the image to separate the background, and sharpening the details of the target to highlight the details.

[0018] In conjunction with the first aspect, further, the processed data batch is used to feed into a feature extraction network with fixed parameters to obtain high-quality features as the target distribution; the original data batch is used to generate ordinary features through a trainable feature extraction network; the discriminator is used to distinguish the sources of high-quality features and ordinary features, and to provide gradients for the feature extraction network.

[0019] In conjunction with the first aspect, the ratio of the number of alternations between adversarial training and classification training is further set to 4:1.

[0020] In conjunction with the first aspect, further, the construction of a discriminator, and the adversarial training of the discriminator and the feature extraction network to form an adversarial structure, also includes:

[0021] During the adversarial training process, network losses are generated; these network losses include generative adversarial losses and detector losses; the generative adversarial losses include adversarial losses from the feature extraction network and the discriminator; constraints are added to the generative adversarial losses, as expressed below:

[0022] L GAN-total =L GAN +αL CenterNet ;

[0023] In the formula, L GAN-total To generate adversarial loss, L GAN For the adversarial training objective consisting of a feature extraction network and a discriminator, L CenterNet The constraints are formed by the adversarial loss of the feature extraction network and the discriminator, where α is the weighting coefficient of the CenterNet network loss.

[0024] Secondly, the present invention provides a target detection enhancement device based on a generative adversarial architecture, comprising:

[0025] Separation module: Used to separate the trained CenterNet into a feature extraction network and a classification network;

[0026] Adversarial training module: Used to build a discriminator, and then adversarially train the discriminator with the feature extraction network to form an adversarial structure;

[0027] Network matching module: Used to recombine the feature extraction network and classification network after adversarial training, fix the feature extraction network, and perform classification training on the classification network to match the feature extraction network.

[0028] Thirdly, the present invention provides a target detection enhancement device based on generative adversarial architecture, including a processor and a storage medium;

[0029] The storage medium is used to store instructions;

[0030] The processor is configured to operate according to the instructions to perform the steps of the method described in any of the first aspects.

[0031] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in any of the first aspects.

[0032] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0033] This invention proposes a novel object detection framework that uses adversarial training to improve the performance of object detection networks that are difficult to further enhance. This allows well-trained object detection networks to improve their performance without adding any parameters. It can quickly and effectively improve well-trained object detection networks, with fast training speed, low computational resource consumption, and high efficiency. It does not come at the cost of increased inference time or training difficulty and is almost a plug-and-play training method. Attached Figure Description

[0034] Figure 1 This is a flowchart of the target detection enhancement method based on generative adversarial architecture provided in an embodiment of the present invention;

[0035] Figure 2 This is a schematic diagram of the overall structure of the target detection enhancement method based on generative adversarial architecture provided in the embodiments of the present invention;

[0036] Figure 3 These are image comparisons of different information representations in the target detection enhancement method based on generative adversarial architecture provided in this embodiment of the invention.

[0037] Figure 4 This is a schematic diagram of the discriminator structure in the target detection enhancement method based on generative adversarial architecture provided in the embodiments of the present invention. Detailed Implementation

[0038] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments and specific features in the embodiments are detailed descriptions of the technical solution of the present application, rather than limitations thereof. In the absence of conflict, the embodiments and technical features in the embodiments can be combined with each other.

[0039] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0040] Example 1:

[0041] Figure 1 This is a flowchart illustrating a target detection enhancement method based on generative adversarial architecture according to Embodiment 1 of the present invention. This flowchart merely shows the logical order of the method described in this embodiment; however, in other possible embodiments of the present invention, different methods may be used, provided there are no conflicts. Figure 1 The steps shown or described are performed in the indicated sequence. The target detection enhancement method based on generative adversarial architecture provided in this embodiment can be applied to a terminal and can be executed by a target detection enhancement device based on generative adversarial architecture. This device can be implemented in software and / or hardware and can be integrated into the terminal, such as any smartphone, tablet, or computer device with communication capabilities. See also Figure 1 The method implemented in this way specifically includes the following steps:

[0042] Step 1: Separate the trained CenterNet to obtain the feature extraction network and the classification network;

[0043] The trained CenterNet is separated into a feature extraction network and a classification network (for generating keypoints, keypoint accuracy loss, and bounding box size). Object detection features are crucial for researchers to achieve better classification accuracy and use fewer computational resources. In this embodiment, CenterNet is an anchorless object detection architecture designed to generate keypoint heatmaps. This architecture transforms a labeled image into three matrices: a heatmap representing the category and object location, a center point bias representing the accuracy loss due to scaling of the statistical feature map, and the object size representing the bounding box size. CenterNet's method traverses every pixel of the features, cleverly avoiding the missed detections caused by insufficient anchor points in anchor-based methods. Furthermore, CenterNet is also not limited by non-maximum suppression (NMS) because its candidate box count is always less than 1 / 4 of the total number of pixels (high-confidence points suppress surrounding points). Therefore, it does not require NMS to reduce redundant candidate boxes to reduce machine load, which also avoids the erroneous suppression effect of NMS when candidate boxes overlap significantly. In this embodiment, in order to verify the feasibility of the method, a deep design based on the CenterNet network was carried out. Its anchorless structure and convolutional network-based classifier are very suitable for the framework of this embodiment.

[0044] Step 2: Construct a discriminator, and then conduct adversarial training between the discriminator and the feature extraction network to form a discriminator.

[0045] Adversarial structure;

[0046] Figure 2 This is a schematic diagram of the overall structure of a target detection enhancement method based on generative adversarial architecture provided in Embodiment 1 of the present invention. Figure 4 This is a schematic diagram of the discriminator structure in the target detection enhancement method based on generative adversarial architecture provided in the embodiments of the present invention.

[0047] The entire object detection architecture consists of three modules: the detection network C, the feature extraction network G, and the discriminator D. The first two form the object detection architecture CenterNet, while the latter two form the generative adversarial mode.

[0048] During the training of generative adversarial training, a subset of images x∈X is extracted from the dataset X, and image processing technique f(·) is used to augment these images, making x^=f(x). Specifically, the augmentation includes Gaussian blurring of the background to separate it, and sharpening of the target details to highlight them. The image data augmented using image processing techniques is the processed batch; the image data without image processing is the original batch. The processed batch... A feature extraction network G with fixed parameters is used for input. * To obtain high-quality features As the target distribution; the original data batch X generates common features Y through a trainable feature extraction network G; the discriminator D aims to distinguish the sources of high-quality features and common features, providing gradients for the feature extraction network.

[0049] like Figure 3 As shown, the three images listed above have undergone Gaussian blurring and sharpening to enhance their expressiveness, while the three images below are the original images. Furthermore, the same number of unprocessed images are extracted again from the dataset. Thus, two batches of data are constructed: the unprocessed batch X^ serves as "noise" for training the feature extraction network, while the processed batch X^, after feature extraction, serves as "real data" for learning the distribution. Compared to the unprocessed images, the processed images can extract richer feature information, and this difference is sufficient for the discriminator to detect them.

[0050] It is worth noting that the weights of the two feature extraction networks for data batches X and X^ are shared only until the end of each generative adversarial training session, and the latter's weights do not participate in the training. This is because the real data distribution should be deterministic during the training of the generative adversarial network; only in this way can the generator learn from the real distribution p. data The generating distribution p is learned in (x). g (x), otherwise the distance between the two distributions will be difficult to reduce, and training will become infeasible.

[0051] In the training process of Generative Adversarial Networks (GANs), the performance difference between the generator and the discriminator is a significant cause of network imbalance. In this embodiment, the generator is derived from the feature extraction network of CenterNet, which has undergone 150 training epochs, while the discriminator is newly built. To avoid imbalanced training in GANs, the generator is fixed before all training begins, and the discriminator is trained for 50 epochs based on the COCO2017 dataset before the alternating training of the GAN commences.

[0052] After several iterations of generative adversarial training, the detection network that does not participate in adversarial training is updated. Since the feature extraction network has far more parameters than the classification network, the former will be trained more carefully and will have more training iterations. After experimental testing, the ratio of alternating generative adversarial training to classification training is set to 4:1. The specific training process is shown in the following pseudo-algorithm.

[0053]

[0054] Step 3: Recombine the feature extraction network and classification network after the adversarial training is completed. Fix the feature extraction network and perform classification training on the classification network to match the feature extraction network.

[0055] After each adversarial training session, the feature extraction network G with fixed parameters... * The parameters of the trained network G will be updated to provide a better target distribution. After several rounds of generative adversarial training, the feature extraction network G will be recombined with the classification network C. The feature extraction network G will be fixed, and the classification network C will be trained and adjusted for classification. Adversarial training will be performed on the target detection network, which is difficult to improve further, using high-quality image features as the target distribution, thereby further enhancing the network's feature extraction capabilities. After training, the classification network C will be trained to match the feature extraction network G, thus completing the network adaptation.

[0056] During adversarial training, network loss occurs; the network loss consists of two parts: generative adversarial loss and detector loss.

[0057] Generative adversarial loss primarily comprises the adversarial losses of the feature extraction network and the discriminator. To avoid mode collapse and training imbalance issues in GANs, and to prevent GAN gradient instability from significantly impacting the performance of the feature extraction network, a constraint L needs to be added to the training loss. CenterNet It can, to some extent, prevent the degradation of the object detection network caused by the training of generative adversarial networks, as shown in the following expression:

[0058] L GAN-total =L GAN +αL CenterNet ;

[0059] In the formula, L GAN-total To generate adversarial loss, L GAN For the adversarial training objective consisting of a feature extraction network and a discriminator, L CenterNet The constraints are formed by the adversarial loss of the feature extraction network and the discriminator, where α is the weighting coefficient of the CenterNet network loss.

[0060] In the process of training the GAN, this embodiment has two training batches, one of which is a batch with stronger representational information capabilities. The other is the original data batch X. Data is drawn from the two training batches: ∈X, respectively fed into two feature extraction networks with fixed parameters and those used in training. and y as The output corresponding to x is represented as follows:

[0061]

[0062] y=G(x)P(a≤Y≤b)=∫ y f Y (y)dywherey=[a,b];

[0063] In the formula, G * (·) represents the extraction network that does not participate in training. It shares weights with G(·) before the end of each training session. and f Y (·) represent the weights respectively. The probability density function of y also represents the distribution fitting target and training object of the generative adversarial network.

[0064] For the discriminator, Wasserstein distance is used to measure the distance between distributions. During GAN training, gradient penalty using WGAN-GP can effectively improve the stability of generative adversarial training. The training loss of the generative adversarial network after adding weight suppression is as follows:

[0065]

[0066] In the formula, the first two terms on the right-hand side are the standard training loss for generative adversarial networks, and the third term is the two-sided constraint of the L2 norm of the input gradient, where f penalty Through f Y and The distribution obtained by random sampling from points between them.

[0067] Detector loss L CIt is built from the three output modules of the CenterNet network (heatmap, center point offset, and target size), and the expression is as follows:

[0068] L C =L k +λ size L size +λ off L off ;

[0069] In the formula, L k It's a loss in the heatmap module, L size It is the loss of the target size functional module, L off It is the loss of the center point offset module, λ size and λ off λ is the weighted coefficient of the loss from the target size functional module and the loss from the center point offset module. In general experiments, λ is taken as... size =0.1, λ off =1.

[0070] After each GAN network training iteration, the detection network C(·) is trained in combination with the feature extraction network G(·) to match the latter's state. During training, the feature extraction network G(·) has fixed weights, while the detection network is carefully adjusted. Therefore, the loss L of the heatmap module... k Loss L of the target size functional module size The loss L of the center point offset module off They are defined as follows:

[0071]

[0072]

[0073]

[0074] In the formula, G * (·) indicates that the weights of the feature extraction network are not updated during the training of C. k (·) represents the heatmap module for detecting network C(·), where C off (·) is the module for detecting the center point offset of network C(·), C size (·) represents the target size detection module of network C(·). xyz This represents the label value of the corresponding coordinate in the heatmap, where c represents the target category, R represents the scaling factor, and p and These represent the coordinates of the center point before and after scaling, s k This is the actual size.

[0075] The above are the implementation steps of the object detection enhancement method based on generative adversarial architecture in this embodiment of the invention. Table 1 shows the experimental results of the structure of this embodiment on the MS COCO2017 dataset. The experiments show that after multi-scale enhancement and generative adversarial training, the CenterNet detector based on the Hourglass-104 backbone network has the highest average accuracy of 45.8% (MS), but the slowest inference speed is 1.6 FPS. At this time, the fastest detection network is the detection network based on ResNet-18, with an AP (MS) of 33.8% at 6.9 FPS. The detection speed and detection accuracy of the detectors based on ResNet-101 and DLA-34 are between ResNet-18 and Hourglass-104, but DLA-34 is better than ResNet-101 backbone network.

[0076] Table 1: Detection accuracy and time of the architecture proposed in this embodiment based on different backbone networks.

[0077]

[0078] Table 2 shows the comparative experimental results of the architecture of this embodiment with the original CenterNet network on the MS COCO2017 dataset, comparing the detection accuracy of the architecture of this embodiment with the original network. In the table, Res and C-Net are abbreviations for ResNet and CenterNet, respectively. The experiments include comparative results for four backbone networks, with performance metrics being the average accuracy under different IOUs and three different target sizes. The experimental data shows that for the four backbone networks, the average accuracy improved for most after using generative adversarial training. For ResNet-18 and DLA-34 networks with fewer network parameters, the AP improvement ranged from [0.3, 2.4], with an average improvement of approximately 0.88%. For ResNet-101 and Hourglass networks with more network parameters, the average AP improvement was around 0.86% and 0.30%, respectively. Therefore, for networks with fewer parameters, the additional generative adversarial training has a stronger improvement effect. Furthermore, for large targets, the AP improvement of 0.55% is much smaller than the 1.78% improvement for small targets. While our architecture is more evident for small targets and small network architectures, generative adversarial training can also be effective for larger targets and more complex backbone networks. Most importantly, the network modifications did not negatively impact detection speed. Experiments demonstrate that the architecture of this embodiment can quickly and effectively improve well-trained target detection networks.

[0079] Table 2: Comparison of detection accuracy between the proposed architecture and the original network

[0080]

[0081] It should be noted that this embodiment uses a Ryzen 7 5800X and an NVIDIA GeForce RTX 3090. All models were developed using the PyTorch 1.10.2 and PyTorch Lightning 1.5.10 deep learning frameworks, and implemented using officially pre-trained networks.

[0082] This embodiment conducted comparative experiments on four backbone networks of the CenterNet network: ResNet-18, ResNet-101, DLA-34, and Hourglass-104. The comparative experiments were conducted between the native CenterNet, which had been trained for 150 epochs, and the CenterNet, which underwent adversarial training after training. Most of the experiments in this embodiment were performed on the MSCOCO2017 dataset, which contains 80 categories, 123,000 training and validation images, and 40,000 test images.

[0083] The training input resolution is 512×512, and the output resolution is 128×128. Similar to the original CenterNet setup, this embodiment uses random flipping, random scaling (between 0.6 and 1.3), cropping, and color jitter for data augmentation, and uses Adam to optimize the overall objective. The initial learning rate is 4e-4. This embodiment uses the network's Average Precision (AP) at different IOU values: AP, AP... 50 and AP 75 As the primary performance indicator, AP is used simultaneously S AP L and AP M This test assesses the object detection network's ability to detect objects of different sizes. The PASCAL VOC 2007 dataset has a limited number of categories and images; therefore, it is used as an additional experiment in this embodiment.

[0084] This invention proposes a novel object detection framework that performs adversarial training on object detection networks whose performance is difficult to improve further. This allows well-trained object detection networks to further improve their performance without adding any parameters. It can quickly and effectively improve well-trained object detection networks, with fast training speed, low computational resource consumption, and high efficiency. It does not come at the cost of increased inference time or training difficulty, and is almost a plug-and-play training method.

[0085] Example 2:

[0086] This invention provides a target detection enhancement device based on generative adversarial architecture, the device comprising:

[0087] Separation module: Used to separate the trained CenterNet into a feature extraction network and a classification network;

[0088] Adversarial training module: Used to build a discriminator, and then adversarially train the discriminator with the feature extraction network to form an adversarial structure;

[0089] Network matching module: Used to recombine the feature extraction network and classification network after adversarial training, fix the feature extraction network, and perform classification training on the classification network to match the feature extraction network.

[0090] The target detection enhancement device based on generative adversarial architecture provided in the embodiments of the present invention can execute the target detection enhancement method based on generative adversarial architecture provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0091] Example 3:

[0092] This invention also provides a target detection enhancement device based on generative adversarial architecture, including a processor and a storage medium;

[0093] The storage medium is used to store instructions;

[0094] The processor is configured to operate according to the instructions to execute the steps of the method described in Embodiment 1.

[0095] Example 4:

[0096] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the method described in Embodiment 1.

[0097] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0098] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0099] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0100] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0101] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A target detection enhancement method based on generative adversarial architecture, characterized in that, Includes the following steps: The trained CenterNet is separated into a feature extraction network and a classification network; Construct a discriminator and train it adversarially with the feature extraction network to form an adversarial structure. After the adversarial training is completed, the feature extraction network and the classification network are recombined. The feature extraction network is fixed, and the classification network is trained to match the feature extraction network. The step of constructing a discriminator and training the discriminator adversarially with a feature extraction network to form an adversarial structure also includes: During adversarial training, a portion of images is extracted from the dataset, and image processing techniques are used to augment these images. The image data after augmentation using image processing techniques is the processed data batch; the image data without image processing is the original data batch. The target enhancement includes Gaussian blurring of the background in the image to separate the background, and sharpening of the target details to highlight the details; The processed data batches are fed into a feature extraction network with fixed parameters to obtain high-quality features as the target distribution; the original data batches are used to generate ordinary features through a trainable feature extraction network; the discriminator is used to distinguish the sources of high-quality features and ordinary features, and to provide gradients for the feature extraction network. The step of constructing a discriminator and training the discriminator adversarially with a feature extraction network to form an adversarial structure also includes: During the adversarial training process, network losses are generated; these network losses include generative adversarial losses and detector losses; the generative adversarial losses include adversarial losses from the feature extraction network and the discriminator; constraints are added to the generative adversarial losses, as expressed below: ； In the formula, To generate adversarial losses, The adversarial training objective consists of a feature extraction network and a discriminator. Constraints are formed for the adversarial loss of the feature extraction network and the discriminator. These are the weighting coefficients for the CenterNet network loss.

2. The target detection enhancement method based on generative adversarial architecture according to claim 1, characterized in that, The construction of a discriminator, followed by adversarial training between the discriminator and the feature extraction network to form an adversarial structure, includes: Before adversarial training, CenterNet was fixed and the discriminator was trained for 50 epochs based on the COCO2017 dataset. During adversarial training, high-quality image features are used as the target distribution to further enhance the feature extraction capability of the feature extraction network. After the adversarial training is completed, the detection network that did not participate in the adversarial training is updated.

3. The target detection enhancement method based on generative adversarial architecture according to claim 1, characterized in that, The ratio of alternating adversarial training to classification training is set to 4:

1.

4. A target detection enhancement device based on generative adversarial architecture, characterized in that, include: Separation module: Used to separate the trained CenterNet into a feature extraction network and a classification network; Adversarial training module: Used to build a discriminator, and then adversarially train the discriminator with the feature extraction network to form an adversarial structure; Network matching module: used to recombine the feature extraction network and classification network after adversarial training, fix the feature extraction network, and perform classification training on the classification network to match the feature extraction network; The step of constructing a discriminator and training the discriminator adversarially with a feature extraction network to form an adversarial structure also includes: During adversarial training, a portion of images is extracted from the dataset, and image processing techniques are used to augment these images. The image data after augmentation using image processing techniques is the processed data batch; the image data without image processing is the original data batch. The target enhancement includes Gaussian blurring of the background in the image to separate the background, and sharpening of the target details to highlight the details; The processed data batches are fed into a feature extraction network with fixed parameters to obtain high-quality features as the target distribution; the original data batches are used to generate ordinary features through a trainable feature extraction network; the discriminator is used to distinguish the sources of high-quality features and ordinary features, and to provide gradients for the feature extraction network. The step of constructing a discriminator and training the discriminator adversarially with a feature extraction network to form an adversarial structure also includes: During the adversarial training process, network losses are generated; these network losses include generative adversarial losses and detector losses; the generative adversarial losses include adversarial losses from the feature extraction network and the discriminator; constraints are added to the generative adversarial losses, as expressed below: ； In the formula, To generate adversarial losses, The adversarial training objective consists of a feature extraction network and a discriminator. Constraints are formed for the adversarial loss of the feature extraction network and the discriminator. These are the weighting coefficients for the CenterNet network loss.

5. A target detection enhancement device based on generative adversarial architecture, characterized in that, Including processor and storage media; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method according to any one of claims 1 to 3.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method described in any one of claims 1 to 3.

Images