YOLOv8n-COSD: A lightweight multi-mechanism network for straw plastic impurity detection

In straw recycling scenarios, challenges such as detail loss, scale variation and target deformation significantly hinder accurate detection. To address these issues, this study proposes YOLOv8n-COSD, a lightweight detection model with multi-mechanism collaborative optimisation, achieving synergistic improvements in detection accuracy and efficiency through hierarchical architectural innovations. First, the Coordinate Attention (CA) mechanism is embedded in shallow layers of the Backbone and Neck networks to enhance spatial localisation accuracy for small targets. Second, the C2f_OD module is constructed by integrating Omni-Dimensional Dynamic Convolution (ODConv) to strengthen morphological diversity representation in shallow-layer features. Furthermore, Switchable Atrous Convolution (SAConv) is deployed in deep Backbone and shallow Neck layers to balance local and global multi-scale feature fusion. Finally, Deformable Convolutional Networks v3 (DCNv3) is integrated into deep Neck layers to improve geometric adaptability for complex deformable targets. Experiment results show that the improved model achieves an F1-score of 91%, mAP50 of 94.1%, and mAP50-95 of 59.8%, representing improvements of 1.1, 0.4, and 2.4 percentage points over the baseline model, respectively, while reducing computational costs by 14.5% to 5.8G. This work establishes a cross-layer collaborative mechanism for dynamic feature optimisation, multi-scale balance, and deformation modelling, providing a reliable algorithmic foundation for developing straw impurity removal equipment and offering a high-precision, low-power algorithmic prototype for lightweight agricultural detection systems.