Residual attention mechanism and weighted feature fusion for multi-scale object detection

Object detection is one of the critical problems in computer vision research, which is also an essential basis for understanding high-level semantic information of images. To improve object detection performance, an improved YOLOv3 multi-scale object detection method is proposed in this article. Firstly, a residual attention module is introduced into the neck of YOLOv3, which includes the channel attention module, spatial attention module, and skip connection. The residual attention module is applied to the three layers of features obtained from the backbone, making the output feature focus on the channels and regions related to the object. Secondly, an additional weight is proposed to add to each input feature in the top-down feature fusion stage of YOLOv3, the size of which is determined by the degree of contribution of each input feature to the output features. The experimental results on KITTI, PASCAL VOC, and bird’s nest datasets fully verify the effectiveness of the proposed method in object detection. The proposed method has significant value in electric power inspection and self-driving automobiles.