Research on x-ray contraband detection and overlapping target detection based on convolutional network

The vast majority of target detection methods are based on convolutional networks to extract features. Yolo series is the mainstream algorithm for target detection today. Given the characteristics of a large number of occlusion and overlapping targets detected by X-ray contraband, it is easy to miss detection and erroneous detection. The yolo algorithm is improved to improve the network performance of the algorithm in detecting overlapping targets. To better reflect the improvement effect, the yolov4 algorithm is used as the baseline for improvement. Firstly, the SPPFCSPC module is proposed to replace multi-space pyramid pooling. Secondly, the CBAM mechanism is introduced to resist invalid features and improve accuracy. Finally, the path aggregation network module is improved, and the convolution module is optimized to improve the prediction accuracy of the network for overlapping targets. The improved network increases average accuracy by 4.86 points in the Sixray dataset, and has certain advantages over advanced algorithms such as yolov5 and yolox. Meet real-time detection standards and effectively improve detection accuracy.