Material-aware Cross-channel Interaction Attention (MCIA) for occluded prohibited item detection

For security inspection, detecting prohibited items in X-ray images is challenging since they are usually occluded by non-prohibited items. In X-ray images, different materials present different colors and textures. On this basis, we exploit the material characteristics to detect occluded prohibited items. Moreover, the occlusion mainly exists between prohibited items and non-prohibited ones, belonging to inter-class occlusion. We propose a Material-aware Cross-channel Interaction Attention (MCIA) module which can use the material information of X-ray images to deal with the inter-class occlusion. Specifically, MCIA is composed of Material Perception (MP) and Cross-channel Interaction (CI). MP captures distinctive material information of X-ray images and CI gets the local cross-channel interaction to convert material information into channel-wise weights. By combining MP and CI, MCIA effectively helps the network to highlight the core features of prohibited items while suppressing non-prohibited items. Meanwhile, we design the MCIA-Net and MCIA-FPN by placing our MCIA module behind each stage in ResNet. Our MCIA-Net and MCIA-FPN can be used as backbones to detect occluded prohibited items. Note that MCIA-FPN also takes into account the prohibited items of various sizes. Our MCIA-Net and MCIA-FPN have been comprehensively validated on the SIXray dataset and OPIXray dataset. The experimental results prove the superiority of our method. Furthermore, our proposed MCIA module outperforms several widely used attention mechanisms and effectively improves the performance of Faster R-CNN and Cascade R-CNN in detecting occluded prohibited items.