A Lightweight Intelligent Vision Architecture With Receptive Field Enhancement and Channel Similarity Attention for Plant Disease Diagnosis

Accurate and interpretable intelligent expert systems are essential for early plant disease diagnosis and precision agricultural management. However, existing diagnostic systems struggle to identify subtle lesions and lack sufficient interpretability, often functioning as “black‐box” classifiers rather than integrated computational cores. This paper proposes PCSAM‐YOLO, a lightweight visual inference architecture based on YOLO11, designed as the computational engine for intelligent plant disease diagnostic systems. Unlike conventional attention‐based models that treat spatial and channel features in isolation, PCSAM‐YOLO integrates Pinwheel Convolution (PConv) and a Channel Similarity Attention Mechanism (CSAM) to achieve synergistic dual‐domain feature refinement. Specifically, PConv effectively expands the receptive field to aggregate directional spatial context, providing a robust structural foundation for feature extraction. Building upon this, CSAM implements second‐order channel interaction modelling to capture fine‐grained inter‐channel correlations, significantly enhancing the model's sensitivity to subtle pathological features. Evaluated on the WCG and Paddy Doctor datasets, the proposed architecture achieves state‐of‐the‐art performance (97.29% Accuracy, 96.84% Precision, 95.91% Recall, and 96.35% F1‐Score). Visual analysis confirms that the integration of PConv and CSAM enables the model to accurately focus on lesion regions, aligning with the diagnostic logic of agricultural experts. Rather than serving only as a standalone classifier, PCSAM‐YOLO functions as a high‐performance inference core capable of supporting downstream decision‐making and automated management recommendations in complex agricultural IoT environments.