Multi-Scale Dense Detector for Prohibited Items Detection in X-Ray Images
Mingyuan Li, Tong Jia, Hui Min Lu, Yingbo Li, Bowen Ma, Hao Wang, Dongyue Chen · 2024
Given the overlapping phenomena observed in X-ray images within security inspection scenarios, as well as the significant variance in the size of prohibited items, we introduce a Multi-Scale Dense Detector (MSDDet). Specifically, to address the issue of overlap in X-ray images, we propose a Feedback Channel Attention Mechanism (FCAM), which refines class semantic information based on the inter-channel correlations in the high-level features and feeds it back to lower-level features. This enhances the capacity of the network to extract target category information from overlapping foreground and background regions. Furthermore, to address the notable size variation of objects in images, we propose a Multi-Scale Feature Aware Module (MFAM). This module mainly utilizes several atrous convolutions with different dilation rates to capture context information across various scales, thereby enhancing the compatibility of the model in detecting prohibited items of multiple scales. Extensive experiments conducted on the PIXray and SIXray dataset demonstrate that the proposed method outperforms state-of-the-art object detectors, indicating its potential application in the field of prohibited item detection.