Class regularization method and system for domain adaptive target detection

Abstract

The invention discloses a category regularization method and system for domain adaptive target detection, and the method comprises the steps: 1, obtaining a source domain region-of-interest feature vector, a target domain region-of-interest feature vector, a corresponding prediction category and a corresponding prediction category; 2, calculating an inter-category regularization loss function value; and step 4, calculating an intra-class regularization loss function value, and combining the intra-class regularization loss function value as a regularization item into a loss function of the domain self-adaptive target detection framework so as to jointly optimize the domain self-adaptive target detection loss function and the class regularization loss function. According to the method, reasonable and sufficient alignment of the same category of features of the two domains is realized, the risk of negative migration of the features of the two domains is greatly reduced, and the method can be flexibly expanded to various existing domain adaptive target detection frameworks.

Description

technical field [0001] The invention relates to the technical field of object detection based on deep learning, in particular to a category regularization method and system for domain adaptive object detection. Background technique [0002] Current object detection methods based on deep learning usually assume that the data of the training set (called the source domain) and the test set (called the target domain) obey the same probability distribution, but this assumption is often difficult to hold in many practical application scenarios. The domain drift problem caused by differences in data distribution degrades the detection performance of deep object detection models on object domains. Although the degradation of model performance can be mitigated by collecting and labeling more training data, this process is extremely time-consuming and costly. Unsupervised domain adaptation transfers knowledge from a source domain with labeled data to a target domain without labeled d...

AI Technical Summary

Problems solved by technology

However, whether the features of the source domain and the target domain are aligned at the category level after adaptation has not been really considered.

In cross-domain scenarios, misalignment of features of different categories, or insufficient alignment of features of the same category, may lead to negative transfer of features from two domains, and the trained domain-adaptive object detection model can only achieve suboptimal performance.

Images