MAM Faster R-CNN: Improved Faster R-CNN based on Malformed Attention Module for object detection on X-ray security inspection

X-ray security inspection machine is currently the most widely used security inspection technology in our country. It is widely used in urban rail transit, railways, airports, key venues, logistics delivery and other scenarios. Using artificial intelligence technology to assist front-line security inspectors in X-ray security inspection can effectively reduce problems such as missed reports due to personnel fatigue or inattention. However, in actual scenes, the detection accuracy is not high due to the diversity of item types, the particularity of imaging angles, and the limitations of detection algorithms. In order to solve related problems, we propose MAM Faster R-CNN. First, in order to expand the receptive field of the feature map and effectively extract the regional features of the target object with shape distortion in the feature map, we propose the Malformed Attention Module (MAM). Secondly, the Large Kernel Attention (LKA) block is used to connect the corresponding backbone output feature layer to use the adaptive selection feature of self-attention module to better focus on the effective feature information in the feature map. Finally, for the neck part, we replace feature pyramid network (FPN) with path aggregation network (PAN), and add the Conv-MLP block to the bottom-up feature fusion part on the right side of the PAN network to reduce the loss of some low-level details. In this paper, the Faster R-CNN benchmark model with ResNet50 as the backbone network is used to evaluate our proposed MAM Faster R-CNN model on HiXray and OPIXray datasets, and the results show that it outperforms the SOTA detection method. We do ablation experiments on the HiXray dataset, and the results show that our proposed MAM Faster R-CNN improves about 2.0%, compared to baseline model (Faster R-CNN).