Towards Object Detection Algorithm Based on LCAS-YOLOX

This paper proposes a composite floor defect detection method based on improved YOLOX network, introduces coordinate attention mechanism (CA) and adaptive spatial feature fusion (ASFF) modules to improve the enhanced feature extraction network of YOLOX network, and uses SIoU loss function to replace the original IoU loss function. The improved network is named CAS-YOLOX. Through ablation experiments, it has been proven that all three improvement methods have good improvement effects. Compared with other algorithms, the CAS-YOLOX network model has higher detection accuracy, with an mAP index of 97.19% and a detection speed of 46 FPS. Based on the design concept of deep separable convolution, the BottleNeck module in CAS-YOLOX was improved to obtain a lightweight network model named LCAS-YOLOX. Experiments have shown that compared with other algorithms, the LCAS-YOLOX network can better balance the relationship between detection accuracy, detection speed, and model lightness. The mAP (mean Average Precision) index is 96.84%, the detection speed reaches 53 FPS (Frame Per Second), and the parameter quantity is only 6.72M.