Domain adaptive unsupervised target detection method based on feature separation and alignment

Abstract

The invention discloses a domain adaptive unsupervised target detection method based on feature separation and alignment, and the method comprises the following steps: inputting paired RGB samples ofa source domain and a target domain into a two-stage target detection framework, calculating the detection loss through the label of the source domain, training the target detection framework, and obtaining a learned target detection model, wherein in the training process, the multi-level high-dimensional features are aligned; aligning the regional instance features of each group; graying the RGBsamples of the source domain and the target domain which are paired, carrying out feature separation, and separating out high-dimensional features related to detection and disturbance features irrelevant to detection. The invention provides an unsupervised domain self-adaptive target detection method based on feature separation and alignment, which can effectively solve the problems of backgroundinformation noise and candidate region redundancy in domain self-adaptive target detection.

Description

technical field [0001] The invention belongs to the technical field of deep learning and target detection, in particular to a domain adaptive unsupervised target detection method based on feature separation and alignment. Background technique [0002] Object detection is a basic problem in computer vision. Its task is to find out the objects of interest on the image and determine their location and category. Driven by the deep convolutional network, many methods based on deep learning have been proposed, making the target detection model have good performance on some benchmark natural image datasets; but in the real environment, due to the style of the target, background, etc. There is a huge difference in domain distribution between the test set and the training set, which will lead to a significant reduction in the generalization ability of the model in new domains. At the same time, in some new fields, such as artistic images, there are few existing public datasets, whic...

AI Technical Summary

Problems solved by technology

However, there are defects in the existing region alignment methods, which do not take into account the inconsistency of the number of target regions in the image; secondly, for the instance level, it is difficult to extract the target region image that needs to be aligned because there is no label in the target region data. The characteristics of the target area, the candidate area proposed by the network may be redundant, and may be affected by the background, so that the model performance is not ideal

Images