Enhancing Prohibited Item Detection in X-Ray Images with Improved YOLOv8

Using computer vision technology to detect contraband in X-ray security inspection images is an effective means to promote the implementation of smart security inspection technology. The background of X-ray security inspection images is cluttered, and the targets occlude each other, which poses a challenge to the performance of object detection models. Based on the YOLOv8 framework, this paper proposes a high-performance contraband detection model for X-ray images, where two modifications are carried out. Firstly, the dynamic detection head, integrating the multi-scale attention mechanism is applied at the object detection stage. By coherently combining multiple self-attention mechanisms between feature layers (for scale perception), between spatial positions (for spatial perception), and within output channels (for task perception), the proposed model can more effectively handle the occlusion and multi-scale problems of prohibited items. Secondly, we increase the number of loop iterations and enhance the receptive field of the feature pyramid layer to make the extracted features more suitable for the prohibited item detection. Simulation experiments are conducted on two public X-ray image data sets, which shows the proposed model increases the Average Precision (mAP) metric on the SIXray dataset from 73.7% to 76.2%, and on the CLCXray dataset from 77.3% to 78.7%.