A lightweight context-aware framework for toxic mushroom detection in complex ecological environments

The accidental proliferation of toxic mushrooms in natural ecosystems poses risks to both biodiversity and human activities in forested regions. Existing detection methods struggle with three key challenges in environmental monitoring: (1) poor discrimination of morphologically similar species in wild habitats, (2) high computational costs limiting deployment in resource-constrained field settings, and (3) performance degradation under ecological variations such as weather changes and terrain complexity. To address these challenges, we propose PM-YOLO which integrates the Contextual and Spatial Feature Calibration Network (CSFCN) and Contextual Anchor Attention (CAA) mechanisms, and is specifically designed for poisonous mushroom recognition. With the help of knowledge distillation technology, our model achieves an mAP@0.5 with 92.64 %, which is 2.06 % higher than that of YOLOv8s. Meanwhile, the number of parameters is only 31.25 % of that of YOLOv8s (3.5 M vs. 11.2 M). Rigorous 10-fold cross-validation demonstrates its excellent robustness, with performance differences of less than 2 % across various test scenarios. PM-YOLO achieves multi-scale feature alignment through hierarchical context fusion, performs adaptive attention weighting for morphological variations, and maintains a low computational cost while significantly improving accuracy. This breakthrough enables the practical application of AI-assisted mushroom identification, effectively bridging the critical gap between academic research and field applications in the field.