An improved transformer-based high-performance object detection method

Abstract

The Transformer architecture is widely used in the field of computer vision due to its ability to relate context and global modeling, and Transformer-based object detection methods have achieved very bright results. However, the existing methods have the problems of underutilizing background images and the model structure is too complex and redundant. In order to solve such problems, an improved high-performance object detection method based on Transformer is proposed. The method first extracts the depth features of the object by means of feature pyramid extraction, then extracts the large, medium and small object regions present in the object image by means of RPN and ROI Pooling, and finally seeks the extrinsic connection between different regions by means of the attention mechanism in the Transformer method, and predicts the object location and category using the proposed loss function. Compared with state-of-the-art object detection methods on MS-COCO datasets, The effectiveness and superiority of the proposed method were demonstrated.