YOLOv8-MDN-Tiny: A lightweight model for multi-scale disease detection of postharvest golden passion fruit

Abstract

Passion fruit, a commercially significant fruit crop, is easily infected by anthracnose and scab, which declines it economic value. However, at the present time, passion fruit quality grading is mainly judged by manual assessment, with strong subjectivity, poor efficiency and low accuracy. Intelligent classification of postharvest passion fruit is essential, with skin disease being a critical factor in grading fruit quality. In view of the shortcomings in traditional deep learning model, such as weak multi-scale detection ability and low accuracy, we propose a YOLOv8-MDN-Tiny model to improve the ability of passion fruit small-scale disease detection. The backbone layer is replaced by the self-made MFSO structure to expand the feature pixels of small target information and enrich their feature expression. An improved DyRep module is proposed to realize the interactive fusion of disease features at different scales and depths. NWD loss function is introduced to accurately measure the overlap of two bounding boxes. Finally, Slimming pruning and CWD are used to compress the model. Compared with YOLOv8s, our improved lightweight model achieves more accurate localization of small passion fruit targets. Specifically, the mAP₅₀ is increased by 2.2–94.8 %, the precision and recall are improved by 1.5 % and 6.0 %. Meanwhile, the number of model parameters and memory usage are decreased by 90.1 % and 88.9 %. The results technically support the disease detection in postharvest passion fruit and real-time grading of their quality.