Dangerous Goods Detection in X-ray Security Inspection Images Based on Improved YOLOv7

According to multi-scale targets and overlapping targets in X-ray security inspection images, this paper proposes a dangerous goods detection algorithm called as YOLOv7-MPCN(MaxPooling with CA and BiFPN). As the convolutional network layers overlap, the resolution of the feature map decreases, resulting in the loss of some position information. Firstly, the coordinate attention mechanism is combined with the MPConv structure in the downsampling stage to enrich semantic information while preserving position perception information, improving the ability to extract dangerous goods feature information in complex overlapping backgrounds. Secondly, using the BiFPN(Bidirectional Feature Pyramid Network) feature fusion module with smaller parameter values as the feature pyramid structure of the model effectively reduces the computational complexity of the model; Finally, referring to the residual network structure in RepVGG, adding skip connections in the lower branch of MPConv enables the network to reach deeper depths without the gradient vanishing. The average detection accuracy of YOLOv7-MPCN on the public dataset SIXray reached 92.5%, which is 0.9% higher than YOLOV7. The experimental results show that the proposed algorithm in this paper has an improvement in detection accuracy compared to the original YOLOv7, proving the effectiveness of YOLOv7-MPCN.