Domain adaptive target detection method and system based on foreground-category perception alignment

Abstract

The invention discloses a domain adaptive target detection method and system based on foreground-category perception alignment. The method comprises the following steps: performing target detection in a scene corresponding to a target domain data set by using a domain adaptive target detection model obtained through the following steps; the method comprises the following steps of: 1, setting an image level domain classifier and an instance level domain classifier, and constructing a reference model; 2, setting a foreground sensing module and a category sensing module, and constructing a domain adaptive target detection model based on foreground-category sensing alignment; and step 3, training and obtaining a domain adaptive target detection model. According to the cross-domain target detection method, the focus of the cross-domain target detection process can be converted from overall feature alignment to foreground feature alignment and then to category feature alignment, and performance improvement on the target domain is achieved under cross-weather, cross-camera and cross-complex scene detection.

Description

technical field [0001] The invention relates to the technical field of target detection based on deep transfer learning, in particular to a domain adaptive target detection method and system based on foreground-category aware alignment. Background technique [0002] Unsupervised domain adaptive object detection improves the detection performance of the object detector on the target domain by transferring the knowledge of the source domain to the target domain when the source domain has labeled data and the target domain has no labeled data. In recent years, unsupervised domain-adaptive object detection has shined in a variety of complex cross-domain detection scenarios by virtue of domain-invariant features learned through adversarial training. These methods use gradient inversion layers to bridge the object detector and domain classifier, and complete the adversarial training by minimizing the object detection loss and maximizing the domain classification loss in the entire...

AI Technical Summary

Problems solved by technology

However, the background area should not be the focus of the migration process for cross-domain detection. If the foreground and background areas are placed in the same position to align the features of the two domains, it is difficult for the cross-domain target detection process to pay attention to the real need to participate in the migration process. the foreground area, thus limiting the improvement of the performance of the cross-domain object detection model

[0004] Although the few existing unsupervised domain adaptive object detection methods try to align the features of the foreground region by decoupling the foreground and background regions, these methods have not yet considered the alignment of features at the category level in the foreground region.

If the foreground area features of the source domain and the target domain are only aligned in a "category unknown" manner, it is very likely to cause misalignment of different class features in the foreground area

Since the category structure of the foreground area has not been fully exploited, it is very likely to cause negative transfer of category features in the foreground area, and the performance of the cross-domain object detection model can only reach a suboptimal state in the end.

Images