X-ray Security Contraband Detection Based on Improved YOLOX

Aiming at the problems of complex background, sample imbalance and easy to squeeze and deform objects in X-ray security inspection contraband images, a YOLOX-based algorithm for detecting contraband objects in X-ray security inspection images is proposed. First, Deformable convolution is added to the backbone network of YOLOX so that the model can learn to adapt to the sampling offsets of different object deformations, thus enhancing the extraction of spatial feature information. Second, Triplet attention mechanism is used to establish inter-dimensional dependencies, which is used to enhance the model's perception of contraband while effectively suppressing background interference. Finally, focus loss is used in the model to improve the confidence loss function of YOLOX, which improves the model's recognition ability in the case of imbalance between difficult and easy samples. The proposed algorithm is tested on the SIXray dataset and the mAP is improved from the original 88.3% to 92.7%. Many existing state-of-the-art algorithms are compared with better detection accuracy and real-time performance.