MFD-YOLO: A fast and lightweight model for strawberry growth state detection

Abstract

Strawberry farming requires efficient and adaptable solutions for real-time monitoring to tackle challenges like rapid ripening, perishability, and bad fruit recognition in field applications. However, existing methods often lack the robustness and lightweight design necessary for resource-constrained environments. To address these limitations, we propose MFD-YOLO, a feature-enhanced, distilled neural architecture based on YOLOv7-tiny, for accurate detection of strawberry growth states. First, we developed the MobileNet-MCA (M-MCA) backbone, which enhances feature extraction while significantly reducing redundant computations. Additionally, Partial Convolution (PConv) is incorporated into the E-ELAN module in the neck, improving feature fusion efficiency while reducing parameters. We also proposed the FocusDownNet (FDN) adaptive downsampling method to better capture and fuse multi-scale features. The DepthLiteBlock is designed to replace the CBL module in the prediction layer, further reducing computational complexity. Finally, an adaptive weighted knowledge distillation (AWKD) strategy is employed to balance performance and efficiency. Experimental results demonstrate that MFD-YOLO achieves a [email protected] of 97.5%, precision of 96.5%, recall of 93.8%, and an F1 score of 95.0%, operating at 128 FPS with a model size of only 3.58 MB. The proposed model outperforms state-of-the-art models and is successfully deployed on both desktop and Android devices, enabling real-time, efficient detection in resource-constrained environments.