DAF-EfficientNetV2: distribution-aware feature enhancement for plant leaf disease recognition in complex environments

Plant leaf diseases in complex natural growing environments are characterized by significant scale variation and morphological heterogeneity, which pose severe challenges to the feature representation capability and inference efficiency of recognition models. Although existing convolutional neural networks have achieved certain progress in disease recognition tasks, they are still constrained by inherent local inductive bias, resulting in limited capability in modeling long-range dependencies and bridging the semantic gap across hierarchical features. To address these issues, this paper proposes a plant leaf disease recognition model termed DAF-EfficientNetV2 (Distribution-Aware Feature-enhanced EfficientNetV2) based on the EfficientNetV2-S backbone. Under the premise of maintaining parameter efficiency and fast inference of the backbone network, the proposed model introduces an integrated Global–Local Semantic Fusion Module (GMSF). Through a spatially orthogonal global modeling strategy, the module efficiently captures cross-region lesion correlations within a single computational unit, and, together with an adaptive semantic calibration mechanism, effectively alleviates the semantic inconsistency among features from different levels. As a result, the model’s panoramic perception capability for scale distortion and complex lesion morphologies is significantly enhanced. In addition, a Shuffle Attention feature re-calibration mechanism is incorporated, which jointly optimizes channel-wise and spatial responses through a “group–parallel–reassemble” strategy, guiding the model to focus on diagnostically valuable lesion regions while suppressing redundant noise from complex field backgrounds in a computationally efficient manner. Furthermore, a lightweight cross-scale aggregation strategy is adopted to achieve efficient integration and alignment of features at different resolutions. Experimental results on the Plant Pathology 2020-FGVC7 and Plant Pathology 2021-FGVC8 datasets demonstrate that the proposed method achieves classification accuracies of 95.43 ± 0.51% (best 96.14%) and 88.39%, respectively, outperforming existing mainstream approaches and validating the effectiveness and robustness of DAF-EfficientNetV2 in complex plant disease recognition tasks. The code is publicly available at: https://github.com/mayue0606/DAF-EfficientNetV2 .