Domain adaptive target detection method and system considering category semantic matching

Abstract

The invention discloses a domain adaptive target detection method and system considering category semantic matching. The method comprises the following steps: step 1, acquiring a source domain image with a label and a target domain image without a label; step 2, training by using the source domain image with the label to obtain a pre-trained basic target detector; step 3, adding a domain self-adaptive component on the pre-trained basic target detector, and performing training by using the source domain image with the label and the target domain image without the lable to obtain a trained domain self-adaptive target detection model; and step 4, removing the added domain adaptive component, and performing target detection on the target domain scene by using the trained domain adaptive target detection model. According to the domain adaptive target detection method and system, the problem of semantic matching of specific categories of two domains in cross-domain target detection is considered, and the problem of wrong alignment of the target categories of the source domain and the target domain in a shared category space is avoided, so that the detection performance of a target detection model on the target domain is further improved.

Description

technical field [0001] The invention belongs to the field of computer vision target detection, and in particular relates to a domain adaptive target detection method and system considering category semantic matching. Background technique [0002] Under the joint action of many factors such as the rapid increase in data scale, the advancement of computing power, and algorithm innovation, deep learning has risen rapidly and achieved considerable development, showing strong advantages in the field of computer vision. In recent years, object detection methods based on deep learning have used large-scale, labeled data to train object detection models, and have achieved outstanding results on various public data sets. In practical applications, there are usually differences in the distribution of the training data set (called the source domain) and the test data set (called the target domain). If the target detection model trained from the source domain is applied to the target do...

AI Technical Summary

Problems solved by technology

However, these methods often only align the distribution of the source domain and the target domain from a macro perspective, ignoring the semantic information of the specific categories of the two domains, which can easily cause wrong matching of category semantics, thereby limiting the cross-domain detection performance of the target detection model. improve

Although some self-training methods using pseudo-labels overcome the difficulty of lacking labeled data in the target domain and improve the cross-domain robustness of the target detection model to a certain extent, the strategy of selecting pseudo-labels is not reasonable enough, resulting in errors. The accumulation of the damage to the performance of the cross-domain object detection model

Images