Fine-Grained Tobacco Impurity Detection by Encoding Global Contextual and Local Information

Reconstituted stem silk (ReS) is a transformative material in the tobacco industry, where impurity detection is critical for ensuring product quality. However, automating this task via deep learning faces three major challenges: complex image backgrounds, extremely small impurities, and the absence of annotated datasets. To overcome these issues, this paper introduces three novel modules: a Coordinate Attention Encoding module for capturing global context, a Multi-Scale Downsampling module for preserving fine details, and a Coordinate-Focus Spatial Pyramid Pooling module for position-aware, multiscale feature fusion. Furthermore, a dedicated annotated ReS dataset is constructed and released to support this research. Comprehensive experiments validate that the proposed method significantly outperforms state-of-the-art models, achieving an 11.8% improvement in mAP and demonstrating a strong potential for industrial inspection.