Design of security image contraband detection system based on PP-YOLOE+_DCS
Chen Fan, Hua Jin, Qinghan Li · 2023
Traditional security screening methods mainly use manual identification of security images, there is a low identification of inefficiency, high error of judgement rate, which has become a bottleneck limiting public safety and security. Therefore, to deal with this problem, this paper proposes the security image contraband detection model PP-YOLOE+_DCS, which makes three main improvements on the basis of the PP-YOLOE+ model. To begin with, we introduced deformable convolution within the backbone network that strengthen the models' feature extraction capability. Secondly, we introduced a coordinated attention mechanism among the backbone network and the detection neck for better focusing the model on the object region. Finally, we replaced the original GIOU loss function with the SIOU loss function to improve the detection accuracy and training speed. The improved PP_YOLOE+_DCS model obtained achieves 91.4% detection accuracy, 2.8% improvement compared with the baseline model mAP, only 0.24 M additional parameters, and 420.2 ms inference delay on embedded devices, which provides a new solution for the intelligence of contraband detection.