Abstract: Target detection algorithms cannot effectively use the edge texture and semantic information of small targets in the feature map due to low data resolution and noise interference, resulting in poor detection results. To solve this problem, this paper proposes an improved algorithm for small target detection based on SSD. Firstly, common convolution and deep separable convolution are used for synchronous feature learning and fusion, and the information-rich shallow features are obtained. Then  the channel and space adaptive weight distribution network is added after the inherent 5 scale feature layer, so that the model pays more attention to the important feature information of the channel and space. Finally, the candidate target frame is subjected to non-maximum suppression screening to obtain the detection result. By comparing the improved method with Faster RCNN, SSD and other methods on the VOC2007 data set, the method reduces the false detection rate of small targets and improves the accuracy of the overall target. The proposed model mAP reaches 78.94%. It is 3.13% higher than the SSD model.

Key words: small target detection, depth separable convolution, multi-scale, weight distribution network, SSD