An unsupervised domain-adapted object detection method based on center alignment and relational saliency

Abstract

The invention discloses an unsupervised domain adaptive target detection method based on center alignment and relationship saliency. In the training stage, a detector generates corresponding target area proposals for images in the source domain and target domain; Model the relationship, and update the category center and the target area proposal; use the updated category center to narrow the distance between each category between the target domain and the source domain, so that the different categories of the target domain are separated by the source domain information Distance; after training, the target domain images are directly classified and detected. This method does not need to calculate the category center separately, but puts the category center and the target region proposal into the graph to update together, so the model can be trained end-to-end; when the category center is aligned, it can reduce the distribution of the source and target domains. At the same time, the inter-class differences of the target domain are enlarged to effectively classify the target domain.

Description

technical field [0001] The invention relates to the technical field of target detection, in particular to an unsupervised domain adaptation target detection method based on center alignment and relational saliency. Background technique [0002] As the basic problem of computer vision, target detection has made rapid progress in recent years under the impetus of deep learning. However, target detection faces a serious problem. When the distribution of test data is different from training data, the detection performance will be seriously degraded. This problem is called "domain shift", where the data domain used to train the model is called the source domain, and the test data domain is called the target domain. One way to solve this problem is to collect data from the target domain and label them. Then training is performed based on the target domain data, but manual data labeling consumes a lot of manpower and material resources, especially for tasks such as target detection...

AI Technical Summary

Problems solved by technology

This method has also begun to be applied in the field of target detection. However, the current method basically follows the experience of image recognition and does not fully consider the characteristics of target detection itself. Exploiting the relationship of multiple objects in an image

(2) The unsupervised domain adaptation method used is often aligned from the perspective of the entire domain, and fails to consider the category information of the target domain data in detail, resulting in poor classification effect

Images