Recursive CNN Model to Detect Anomaly Detection in X-Ray Security Image

Abstract

To address the issue of contraband scale difference in the identification of X-ray pictures during security inspection, we upgrade the Faster RCNN network and propose a multi-channel region proposal network (MCRPN). Multi-layer feature extraction is achieved using the complementarily of distinct levels of convolution features in visual semantics, and the richer semantic components of VGG16 high-level layers and the shallower edge features of low-level layers are fused; To construct a multi-scale contraband detection network, the multi-scale candidate target regions are mapped to the corresponding feature maps; dilated convolutions are introduced into the multi-channel, and a multi-branch dilated convolutions module (DCM) is designed to increase the Receptive field and thus enhance features at different scales. On the self-created data set SIXray OD, the enhanced algorithm achieves an average detection accuracy of 84.69 percent and a test performance improvement of 6.28 percent over the original network. Additionally, the testing findings indicate that the enhanced algorithm's recognition accuracy has been increased to a considerable level.