Feature-enhanced small object detection algorithm for surface defects of cigarette pack

The surface defects of cigarette pack are characterized by varying scales and complex backgrounds. Small target detection algorithms, as a potential solution, have been proven to have the ability to identify such surface defects. However, the existing small target detection algorithms still have many limitations, especially in terms of detection accuracy, model complexity, and computing speed. To address these challenges, we propose a lightweight cigarette pack defect detection algorithm, YOLOv5s-FDE. In the backbone network, we introduce the GSConv structure, which incorporates channel shuffle and depth-wise separable convolution to construct the feature enhancement module (FEM) with a cross-stage architecture. It aims to reduce information loss and enhance feature utilization. In the feature fusion stage, we introduce the DySample upsampling module, which effectively preserves target detail features through a dynamic sampling mechanism, and reduces the memory usage. Finally, the E-Triplet cross-dimensional attention module is proposed, which leverages a multi-branch structure to enhance cross-dimensional information interaction and suppress irrelevant background information. Experimental results demonstrate that, compared to the YOLOv5s baseline algorithm, YOLOv5s-FDE achieves an average precision of 95.6%. The number of model parameters and computational cost are reduced by 0.88M and 3.5G, respectively. The proposed method outperforms recent state-of-the-art detection algorithms, whiles providing theoretical support for intelligent defect detection of cigarette pack.