CN-115984623-B - Target detection enhancement method, device and storage medium based on generation countermeasure architecture

CN115984623BCN 115984623 BCN115984623 BCN 115984623BCN-115984623-B

Abstract

The invention discloses a target detection enhancement method, a target detection enhancement device and a storage medium based on a generated countermeasure architecture, and relates to the technical field of target detection in computer vision. The invention provides a new target detection framework, which performs antagonistic training on a target detection network with difficulty in further improving performance, can enable the target detection network with perfect training to further improve the performance, has no parameter increase, can quickly and effectively improve the trained target detection network, has high training speed, consumes less and efficient computing resources, does not have the cost of increasing reasoning time or training difficulty, and is almost a plug and play training mode.

Inventors

ZHANG YUN
HUANG JINGWEI
GAO GUI
ZHAO YANG
ZHANG YUYAO
HUANG CHENG

Assignees

南京邮电大学

Dates

Publication Date: 20260505
Application Date: 20230130

Claims (6)

1. A method for enhancing target detection based on a generated countermeasure architecture, comprising the steps of: Separating CENTERNET with complete training to obtain a feature extraction network and a classification network; constructing a discriminator, and performing countermeasure training on the discriminator and the feature extraction network to form a countermeasure structure; Combining the feature extraction network and the classification network after the countermeasure training is completed, fixing the feature extraction network, and performing classification training on the classification network to enable the classification network to be matched with the feature extraction network; The construction of a discriminator, the discriminator and the feature extraction network are subjected to countermeasure training to form a countermeasure structure, and the construction of the countermeasure structure further comprises: In the process of countermeasure training, extracting a part of pictures from a data set, and carrying out target enhancement on the part of pictures by using an image processing technology, wherein the picture data subjected to target enhancement by using the image processing technology is a processed data batch; The target enhancement comprises the steps of carrying out Gaussian blur processing on a background part in the picture to separate the background, and carrying out sharpening processing on details of the target to highlight the details; the raw data batch is used for generating common features through a trainable feature extraction network, and the discriminator is used for distinguishing sources of the high-quality features and the common features and providing gradients for the feature extraction network; The construction of a discriminator, the discriminator and the feature extraction network are subjected to countermeasure training to form a countermeasure structure, and the construction of the countermeasure structure further comprises: during the countermeasure training, network losses are generated, wherein the network losses comprise generated countermeasure losses and detector losses, the generated countermeasure losses comprise a feature extraction network and a discriminator countermeasure losses, and constraints are added to the generated countermeasure losses, and the expression is as follows: ; In the formula, In order to generate a countering loss, For an countermeasure training goal consisting of a feature extraction network and a discriminator, Constraints on the formation of countermeasures for the network and the discriminator are extracted for the features, Is a weighting factor for CENTERNET network losses.
2. The method of claim 1, wherein constructing a discriminator, performing challenge training on the discriminator and the feature extraction network to construct a challenge structure, comprises: Prior to challenge training, the discriminator is immobilized CENTERNET, training for 50 cycles based on the COCO2017 dataset; When the countermeasure training is carried out, the high-quality image features are used as target distribution, so that the feature extraction capability of the feature extraction network is further enhanced; after the countermeasure training is completed, the detection network which does not participate in the countermeasure training is updated.
3. A method of target detection enhancement based on generating countermeasure architecture according to claim 1, wherein the ratio of the number of alternations between countermeasure training and classification training is set to 4:1.
4. A target detection enhancement device based on a generation countermeasure architecture, comprising: The separation module is used for separating CENTERNET with complete training to obtain a feature extraction network and a classification network; The countermeasure training module is used for constructing a discriminator, and performing countermeasure training on the discriminator and the characteristic extraction network to form a countermeasure structure; The network matching module is used for recombining the feature extraction network and the classification network after the countermeasure training is finished, fixing the feature extraction network, and carrying out classification training on the classification network so as to enable the classification network to be matched with the feature extraction network; The construction of a discriminator, the discriminator and the feature extraction network are subjected to countermeasure training to form a countermeasure structure, and the construction of the countermeasure structure further comprises: In the process of countermeasure training, extracting a part of pictures from a data set, and carrying out target enhancement on the part of pictures by using an image processing technology, wherein the picture data subjected to target enhancement by using the image processing technology is a processed data batch; The target enhancement comprises the steps of carrying out Gaussian blur processing on a background part in the picture to separate the background, and carrying out sharpening processing on details of the target to highlight the details; the raw data batch is used for generating common features through a trainable feature extraction network, and the discriminator is used for distinguishing sources of the high-quality features and the common features and providing gradients for the feature extraction network; The construction of a discriminator, the discriminator and the feature extraction network are subjected to countermeasure training to form a countermeasure structure, and the construction of the countermeasure structure further comprises: during the countermeasure training, network losses are generated, wherein the network losses comprise generated countermeasure losses and detector losses, the generated countermeasure losses comprise a feature extraction network and a discriminator countermeasure losses, and constraints are added to the generated countermeasure losses, and the expression is as follows: ; In the formula, In order to generate a countering loss, For an countermeasure training goal consisting of a feature extraction network and a discriminator, Constraints on the formation of countermeasures for the network and the discriminator are extracted for the features, Is a weighting factor for CENTERNET network losses.
5. An object detection enhancement device based on a generated countermeasure architecture is characterized by comprising a processor and a storage medium; The storage medium is used for storing 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, on which a computer program is stored, characterized in that the program, when being executed by a processor, implements the steps of the method of any of claims 1-3.

Description

Target detection enhancement method, device and storage medium based on generation countermeasure architecture Technical Field The invention relates to a target detection enhancement method, a target detection enhancement device and a storage medium based on a generated countermeasure architecture, and belongs to the technical field of target detection in computer vision. Background Object detection is one of the fundamental problems of computer vision, and its derivative tasks such as pedestrian detection, face recognition and video monitoring detection are increasingly receiving attention. In recent years, as machine performance increases, application fields and depths increase, a large number of deep and complex convolution layers are beginning to be used to promote the feature extraction capability of the network for dense details of pictures. This directly leads to a dramatic increase in computational resource consumption and difficulty in controlling the feature extraction network during training. In response to these problems, a series of techniques have been proposed to improve network performance while avoiding a significant increase in network parameters. Such as anchor-and anchor-free models, greedy non-maximum suppression, learnable non-maximum suppression, residual connection networks, feature pyramids. In addition, the mapping learning capability of the countermeasure network from the potential space to the real distribution is generated so that the countermeasure network can cope with target detection tasks from a brand new angle, especially the detection of small objects and the detection of low-resolution pictures. For generating the challenge network (GENERATIVE ADVERSARIAL Networks, GAN), once it has obtained the target distribution of a task, it is possible to learn the "near-correct answer" matching approach-i.e. pursuing a non-convex gambling Nash balance between the generator and the discriminator with continuous high-dimensional parameters. Since GAN typically uses a gradient descent method to handle the losses of the generator and discriminator, this can lead to it staying at a locally low value of the artificially designed loss function, rather than at the nash equilibrium point of the non-convex game. In order to solve this problem, studies have been made to propose a feature matching method for generating pictures, seeking a nash balance, and a semi-supervised training method for target detection based on such ideas. It has also been proposed to incorporate a detector as a third party into the two-party game of generator and discriminator to form a "Triple-GAN" with three-party games. Although this approach carefully designs the countermeasure training process (training the discriminator with the generator and detector as the same party, training the generator and detector with the latter and the former rejected, respectively), the negation of the classification result of the real picture by the discriminator for the detector, and the strong coupling of the generator and detector that the approach endeavours to achieve, obviously, can hamper the performance of the detector to some extent. Therefore, expanding the two-party game to three parties is not a method for facilitating stable training. Disclosure of Invention The invention aims to overcome the defects in the prior art, and provides a target detection enhancement method, device and storage medium based on a generated countermeasure architecture, which solve the problem of difficult control of a feature extraction network when a computer performs target detection. In order to achieve the above purpose, the invention is realized by adopting the following technical scheme: in a first aspect, the present invention provides a method for target detection enhancement based on a generated countermeasure architecture, the method comprising: Separating CENTERNET with complete training to obtain a feature extraction network and a classification network; constructing a discriminator, and performing countermeasure training on the discriminator and the feature extraction network to form a countermeasure structure; And recombining the feature extraction network and the classification network after the countermeasure training is finished, fixing the feature extraction network, and carrying out classification training on the classification network so as to enable the classification network to be matched with the feature extraction network. With reference to the first aspect, further, the constructing a discriminator, performing countermeasure training on the discriminator and the feature extraction network to form a countermeasure structure includes: Prior to challenge training, the discriminator is immobilized CENTERNET, training for 50 cycles based on the COCO2017 dataset; When the countermeasure training is carried out, the high-quality image features are used as target distribution, so that the feature extraction capability of the feat