GRI-Net: A Gated Regional Interaction Network With Dual Cross-Axis Attention for Plant Disease Diagnosis in Real Field Conditions

초록

Plant diseases represent a serious risk to global food security, causing significant economic instability and substantial crop yield losses each year. Early and precise diagnosis is therefore needed for sustainable farming and stable food production. Plant disease classification under real-field agricultural conditions is challenging due to complex backgrounds, inconsistent illumination, unclear infected region boundaries, and highly similar disease patterns. Existing methods often rely on small fixed receptive fields that primarily capture local textures while failing to model broader infection structures. Conventional attention mechanisms also fail to capture multi-scale disease structures and suppress irrelevant background activations. Therefore, we propose a gated regional interaction network (GRI-Net) utilizing a multi-scale spatial energy response normalization (MS-SERN) block, dual attention cross-axis modulation (DACAM) block, and gated regional interaction block (GRIB). MS-SERN extracts contextual representations of plant infections at fine-grained, medium-scale, and global levels. DACAM uses spatial axis interaction to suppress background-correlated channels and amplify disease-discriminative channels, while GRIB enables contextual relationship modeling among neighboring disease regions and feature sharing among deep semantic lesion areas. Experiments were performed on three open-source datasets, TLD-PK, SLIF-Brinjal, and JWS. GRI-Net achieved average accuracies of 83.24%, 76.00%, and 89.02%, with corresponding harmonic means of precision and recall (F1-scores) of 82.54%, 76.61%, and 89.65%, respectively. Experimental results highlight that GRI-Net outperforms state-of-the-art approaches and effectively classifies plant diseases under real-field conditions.

키워드