DSLNDD-Net: A multi-scale Edge-aware attention network for nutrient deficiency detection in dense strawberry leaves

Real-time precise detection of nutrient deficiency symptoms in strawberry leaves is crucial for enabling precision fertilization and growth regulation. To overcome the limitations of existing methods in handling dense leaf distribution, subtle symptoms, and variable target scales lead us to propose a real-time DSLNDD-Net (Dense Strawberry Leaf Nutrient Deficiency Detection Network). Specifically, the optimized backbone integrates multi-path complementary downsampling with high-frequency compensation, preserving deficient-region edges while suppressing noise to strengthen low-level representation, thereby facilitating subtle lesion differentiation and establishing a discriminative foundation for subsequent. During feature fusion, cross-scale cascaded calibration module is coupled with small-target detection head to mitigate the dilution of shallow, fine-grained features by high-level semantics. By aligning low-level details with semantic information through progressive cross-scale calibration, this module ensures the robust propagation of detailed symptomatic traits throughout the fusion process. Furthermore, the dynamic shape-aware bounding box loss is introduced to optimize the regression process by synergizing dynamic weight allocation with anisotropic shape constraints, which alleviates regression biases from imbalanced gradient contributions and geometric heterogeneity, thereby reshaping the gradient distribution of predicted boxes to enhance localization precision and stability in complex scenarios. Experiments demonstrate that DSLNDD-Net achieves 90.8% mAP (mean average precision), 85.3% recall, 85.5% precision, and 85.4% F1 score for detecting calcium-, iron-, and phosphorus-deficient leaves, outperforming YOLOv10s by 3.7% in mAP, 4.8% in precision, and 3.3% in recall while achieving 125.0 FPS. DSLNDD-Net exhibits strong robustness and high real-time performance, enabling rapid and accurate identification of strawberry leaf deficiency symptoms, facilitating dynamic fertilization adjustments to reduce yield loss.