A real-time detection method for classification and grading of multi-variety shiitake mushrooms in natural growth environments.

The intelligent production of edible fungi requires accurate and real-time monitoring of mushroom growth status throughout cultivation. However, real-time classification and grading of multiple shiitake mushroom varieties in natural environments remain challenging due to illumination variation, pose diversity, occlusion, dense small targets, and subtle inter-variety morphological differences. To address these task-specific challenges, this study proposes a lightweight real-time detection framework for fine-grained shiitake mushroom recognition. The model is designed from a problem-driven perspective by jointly considering deployment efficiency, small-object perception, and subtle feature discrimination. First, a Spatially-weighted Token Gated Convolution (STGConv) module is proposed to achieve lightweight feature extraction by reducing computational redundancy while maintaining robust feature learning. Second, an Adaptive Residual Fusion Head (ARFHead) is proposed to enhance multi-scale feature interactions and improve the perception of densely distributed, partially occluded mushrooms. Finally, a Dynamic Sampling operator is incorporated into the upsampling process to adaptively restore critical detail features and improve discrimination of subtle differences among varieties and growth grades. This modular combination provides a balance between lightweight computation, small-object perception, and fine-detail reconstruction, making it suitable for agricultural vision tasks in natural growth environments. Experimental results show that the proposed model achieves an mAP@0.5 of 97.9%, precision of 96.0%, recall of 96.3%, and 110.8 FPS, with a model size of only 8.27 MB, indicating its suitability for real-time monitoring and edge deployment in practical cultivation environments.