Improved Target Detection Algorithm Based on YOLO

Abstract

In the field of target recognition, YOLO algorithm has performed well. In this paper, we improve the latest YOLO network model by changing the residual units to dense connection in the CSP module and adding channel attention mechanism. The improved network model alleviates the vanishing-gradient problem, enhances feature propagation, encourages feature reuse, and reduces the number of parameters. What's more, it can adaptively recalibrate the channel information of the feature maps and improve the performance of target detection. Experimental results show that the improved YOLO network model greatly improves the detection accuracy. In addition, it optimizes the problem of missing and mis-detecting targets.