ASC-YOLOv8n: Enhanced multi-scale feature fusion for accurate detection of four cigar appearance defects

Abstract

Abstract The appearance defects of cigars can significantly compromise their overall quality, with detection currently relying mainly on manual inspection, a process that is time-consuming and inefficient. The ASC-YOLOv8n model has been proposed for high-precision automated defect detection in full-leaf handmade cigars, targeting defects such as green spots, holes, breaks, and tail breaks during production. This model incorporates several key improvements: it integrates an ASC (Adaptive Spatial Context) module to provide a more hierarchical receptive field, enhancing its ability to detect intricate defect patterns; replaces the conventional C2f module with a more advanced Fusion module, which combines multiple feature maps to improve feature extraction capabilities; and employs a novel WIoU (Weighted Intersection over Union) localization loss function, significantly refining defect localization precision. Experimental results show that the ASC-YOLOv8n model achieves a performance boost, with the mean average precision at an IoU threshold of 0.5 (mAP@0.5) increasing by 1.7% points to 94.00%, outperforming the baseline YOLOv8n model. This demonstrates the model’s effectiveness in accurately identifying critical cigar defects, making it a reliable and robust solution for intelligent cigar inspection, and contributing to enhanced quality control in cigar production.