Abstract: Due to the complex density overlap and texture interference between prohibited items and background materials in subway security X-ray images, the feature representation ability is insufficient, making it difficult to effectively distinguish prohibited items from normal items. The traditional methods are prone to losing key spatial information of small prohibited objects during feature extraction, ultimately leading to serious missed and false detections in detection systems. To address this issue, a subway security X-ray prohibited object intelligent recognition algorithm based on feature extraction and enhancement is proposed. A multi-scale feature extraction framework is constructed based on improved SSD-VGG16(Single Shot MultiBox Detector-Visual Geometry Group 16). The ability to extract microscopic features of prohibited objects is enhanced by adding Conv3 _3 detail capture layer and Conv5_3 small object sensitivity layer, and integrating semantic information from Conv4_3 and other basic network layers using feature fusion technology, significantly improving the completeness of feature representation; On this basis, a spatial attention mechanism is introduced to obtain X-Y bidirectional attention vectors by decomposing and aggregating features, effectively focusing on key areas of prohibited items. At the same time, an ECA( Efficient Channel Attention) channel attention module is embedded to implement cross channel interactive learning, achieving dynamic enhancement of discriminative features of prohibited items; By using the DIoU-NMS (Distance-Intersection over Union Non-Maximum Suppression)algorithm to comprehensively consider the target box overlap rate and center distance for optimization screening, the missed detection rate in dense scenes is significantly reduced; By using adaptive threshold segmentation method and combining Wiener filtering and median filtering preprocessing techniques to eliminate image noise interference, accurate area segmentation of prohibited objects is achieved based on grayscale or pseudo color distribution characteristics, thereby realizing X-ray prohibited object recognition. According to the experimental results, the pixel brightness corresponding to the metal knife, fire machine, and glass bottle recognized by the algorithm is 255, 153, and 51, respectively, which is consistent with the experimental indicators and can accurately identify various prohibited items.

Key words: feature extraction, feature enhancement, subway security X-ray, contraband, intelligent identification