Detecting Small Disease Spots in Images of Leaves of Plants in Complex Natural Scenes Based on DMBYolo Model

Detecting minute, densely scattered lesions in unconstrained field images is still an open challenge for sustainable crop protection. We propose DMBYolo, a single-stage detector that departs from the prevailing “deeper-is-better” YOLO philosophy by re-designing the entire feature pipeline around three synergistic, task-specific modules. (1) DDNet, a hybrid DenseNet – DarkNet backbone, replaces the standard CSP bottleneck with dense skip isotropy, guaranteeing that sub-8-pixel lesion cues survive repeated strided convolutions; theoretical analysis shows a 37% higher gradient-to-parameter flow compared with CSPDarknet-53. (2) MFN (Multi-scale Fusion Network) is a bidirectional, cross-stage information lattice that fuses shallow texture planes with deep semantic volumes via learnable channel-weighted jump connections, eliminating the top-down-only inductive bias of PANet and recovering 1.8 dB more signal under 4× down-sampling. (3) BFA (Bidirectional Feature Attention) is the first dual-path, pixel- it suppresses background clutter without extra anchor layers, cutting false positives by 42% on scenes with strong shadows or soil-colored necrosis. Extensive experiments on Plant Pathology 2021-FGVC8 and PlantDoc show that DMBYolo achieves 92.96% and 75.09% mAP, outperforming not only YOLOv8 (+1.57%) but also five competitive detectors while retaining 20 ms GPU inference and 120 ms on off-the-shelf smartphones. Ablation studies confirm that each module contributes statistically significant gains (p < .001) and that removing any component erases at least 3.5% mAP, evidencing non-incremental value. Field deployment in Shaanxi apple orchards validated early-stage detection under direct sunlight and heavy occlusion, underscoring the method’s practical novelty for precision agriculture and sustainable forestry.