An Anchor-Free Target Detection Algorithm Combining Attention and Dilation Convolution

Abstract

Aiming at the problem of insufficient target detection capability in CenterNet, an improved target detection model combining attention and cavity convolution is proposed. Firstly, in order to improve the ability of the network to obtain the semantic and location features of the target, an improved nonlocal attention mechanism module (CANL) is designed to capture the remote dependence of the target in the image along the channel domain and the spatial domain, respectively. Secondly, a multi-scale feature extraction network based on dilation convolution (MSNet) is designed to improve the expression ability of the network to different scale targets, the residual structure is used to fuse the receptive field features of multiple scales in parallel, and the feature information obtained by the target in the image at multiple scales is retained. Finally, the proposed algorithm is verified on PASCAL VOC dataset. The detection accuracy of the proposed algorithm is 2.65 % higher than that of the baseline algorithm CenterNet, which effectively improves the performance of the anchorless object detection algorithm.