An improved small target detection algorithm for SSD

Abstract

In order to improve the accuracy and robustness of the small target detection for SSD (Single Shot Multi-Box Detector) algorithm, this paper proposes an improved SSD algorithm. Firstly, the Mish activation function is introduced in the backbone network to provide more linear relationships. Secondly, a copy-reduce-paste data enhancement method is proposed to ensure a more balanced train. In the multi-scale detection stage, feature enhancement module and feature fusion strategy guided by channel attention mechanism are used to improve the feature information. At the same time, modify the loss function to improve the model training effect. Experiments on the VOC2007+2012 dataset show that the detection performance of the proposed algorithm is better than that of the SSD algorithm.