Small Object Detection Based on Context Information and Attention Mechanism

Abstract

In order to solve the problem of missed detection and false detection of small targets in object detection, and to improve the detection precision and recall of small object, this paper proposes a small object detection algorithm that introduces context information and attention mechanism. The algorithm is improved on the Faster RCNN network architecture, and a multilevel feature fusion module is proposed to solve the problem of incomplete extraction of detailed information. The proposed regional attention module solves the interference of background noise and focuses on the target to be detected. At the same time, in order to more effectively meet the characteristics of small target detection, we have improved the anchor box. The method proposed in this paper is verified on DIOR, PASCAL VOC2007 and MS COCO datasets. Experiments show that the algorithm proposed in this paper and the current advanced algorithm have better accuracy and precision in detecting small targets.