Abstract:To overcome the challenges of low detection accuracy in identifying surface character defects (including overlapping, misalignment, and displacement) on cylindrical products such as cigarettes, a lightweight defect detection network based on temporal and frequency domain feature fusion is proposed. High-definition character edges and detailed information are obtained by integrating the Scharr operator with temporal-frequency features, which effectively captures both local image variations and structural periodicity characteristics. A shared convolution mechanism is incorporated to develop a lightweight detection head, achieving enhanced inference speed while maintaining detection accuracy. Experimental validation was performed using 2,372 cigarette character defect images. The optimized network demonstrates significant improvements with a mean average precision (mAP) of 96.37%, recall rate of 91.13%, and precision of 95.47% - representing respective enhancements of 6.97%, 4.62%, and 6.3% over the baseline model. Furthermore, the detection speed is boosted to 87.9 frames per second (FPS) , achieving a 6.6 FPS increase compared to the original implementation. This methodology provides an effective solution for industrial defect detection in cylindrical products, particularly demonstrating strong applicability in cigarette quality control systems.