YOLO lightweight contraband detection network using attention mechanism

Abstract

In stations, airports and other places, contraband detection faces many problems such as false positives, omissions and slow detection speed caused by object background interference and human factors. This paper proposes an improved network based on YOLO-lightweight. The attention mechanism module is embedded in the backbone network, focusing on the important features from different channels. CBAM-FPN (Convolution Block Attention Module and Feature Pyramid Networks) structure is adopted in the network neck to reduce the loss of network features. Attention mechanism module is added in the bottom-up feature fusion process. Finally, CIOU is used as the edge optimization loss function to accelerate the network convergence and optimize the network model. Compared with YOLOv4-tiny, the precision is improved by 3.8%, reaching 87.5%. The detection speed reaches 60.3fps. The improved network only occupies 23.4M memory, which is convenient for embedding mobile devices. The improved network meets the real-time detection requirements.