LMF-YOLO: a multi-scale inception model for edge-based strawberry disease detection

Abstract

Rapid and precise identification of strawberry diseases is paramount for securing crop yield and quality. To address challenges arising from significant scale variations, diminutive early-stage lesions, and high inter-class visual similarity, this study proposes a parameter-efficient multi-scale feature fusion network derived from YOLO26, termed LMF-YOLO. The architecture incorporates three strategic enhancements: First, leveraging the end-to-end NMS-free framework of YOLO26, a Triple Feature Encoder (TFE) is developed to aggregate multi-scale features in a one-shot manner, preserving critical high resolution spatial details. Second, a lightweight Multi-scale Initialization Module (MSInit), utilizing separable dilated convolutions, captures contextual information across diverse receptive fields, enhancing the model’s discriminative power against visually similar diseases. Third, a P2 micro-detection head is added to provide explicit supervision for extremely early tiny lesions. Evaluations on a custom strawberry disease dataset (7 categories, 6,246 annotations) demonstrate that the proposed model LMF-YOLO (YOLO26n + LMF neck + P2 head) achieves an mAP50 of 82.0%, marking a 1.9 percentage point increase over the baseline YOLO26n. Notably, the model yields an 11.5% improvement in AP_small for early lesions. Following TensorRT acceleration on the Jetson Nano, the model achieves an inference latency of 33.6 ms (~30 FPS), meeting the real-time constraints for mobile-based recognition. Ablation and comparative studies confirm that LMF-YOLO outperforms mainstream YOLO variants in balancing computational efficiency with detection sensitivity, providing a robust technical solution for mobile-intelligent strawberry disease diagnostics.