SSD with Multi-scale feature Fusion and Attention mechanism

Abstract

Abstract In the Internet of Things, image acquisition equipment is very important, which will generate lots of invalid data when monitoring in real time. Analyzing the data collected from the terminal directly by edge calculation can remove invalid frames and improve the accuracy of system detection. SSD algorithm model is relatively light and fast detection speed. However, SSD algorithms don't take full advantage of both shallow and deep information, a multiscale feature fusion attention mechanism structure based on SSD algorithm was proposed in this paper, which combines the idea of multiscale feature fusion and attention mechanism. Improve the feature information expression ability by fusing adjacent feature layers for each detection layer. Then, add the attention mechanism to to improve the algorithm's attention to the feature map channels. The results of the experiment show that the optimized model detection accuracy is improved, which will greatly improve the reliability of edge calculation.