Object Detection Algorithm Based on Alternate-Attention

Abstract

The one-stage object detection algorithm, YOLOv3, has a fast detection speed and can meet real-time requirements. But giving the same attention weight to all grids during detection will result in the inability to highlight the detection subject, so the positioning accuracy of the bounding box is still room for improvement. In order to improve the detection accuracy, this paper proposes an Alternate-Attention mechanism, using the global pooled attention mechanism to highlight the overall characteristics, and the self-attention mechanism to reflect the self-weight relationship between features. The two attention mechanisms are alternated and applied to the two dimensions of channel and space, and finally enhance the features extracted by Darknet-53. Experiments on the PASCAL VOC2007 dataset shows that this algorithm can effectively improve the detection accuracy. Compared with Faster RCNN, YOLO series and SSD series algorithms, the mAPlouo.5 value of this algorithm is higher, up to 80.24.