An Improved Cigarette Defect Detection Algorithm Based on YOLOX

Abstract Surface defect detection has always been a difficult challenge. Aiming at the detection of surface defects in cigarettes, an improved YOLOX-S model is proposed. Firstly, an improved attention mechanism named MS-GCT (Multi-Spectral Gaussian Context Transformer) is introduced in the backbone network to enhance the model’s ability to capture the global context information in the image and improve its understanding of semantic feature information; secondly, a DMG (Dynamic convolution and MS-GCT) module is proposed and combined with the C2f (CSPLayer with 2 convolutions) module to construct the C2f-DMG module,which is introduced into the model to enhance feature interaction, improve specific feature extraction ability, and strengthen long-distance dependency ability of global features; finally, the loss function is replaced with SIoU to improve model performance and accelerate model convergence. The effectiveness of the model is verified on both the self-made cigarette dataset and the public dataset. The experimental results show that the improved model not only ensures the lightweight of the model, but also improves the model mAP by 2.02, while achieving a detection speed of 73.17 frames− 1. The proposed algorithm also meets the real-time detection requirements for cigarette appearance defects.