Feature enhancement and fusion-optimized defect detection model for Sanhua plums

TL;DR

This paper introduces a new model for detecting defects in Sanhua plums using enhanced image data and optimized algorithms.

Contribution

The novel YOLO-CMA model improves detection of small defects in Sanhua plums with enhanced feature extraction and fusion.

Findings

The YOLO-CMA model achieved 0.639 mAP50 and 0.91 precision for insect-damaged fruit detection.

YOLO-CMA has a computational cost of 5.9 GFLOPs and 2.43 M parameters, suitable for edge deployment.

Ablation experiments confirmed the effectiveness of the C2fCIB and C3k2_Mambaout modules.

Abstract

To achieve precise and efficient detection of abnormal Sanhua plums, this study first constructed a specialized image dataset encompassing five categories: diseased fruit, insect-damaged fruit, bird-pecked fruit, cracked fruit, and normal fruit. To mitigate the initial class imbalance, a multi-weather simulation data augmentation strategy was employed, which expanded the dataset to 10,000 images and achieved a balanced distribution. After systematically evaluating multiple state-of-the-art detection models, YOLOv12 was selected as the baseline model due to its high recall and extreme lightweight nature. To overcome the core challenge of detecting minute defects like insect-damaged fruit, this study innovatively proposed the YOLO-CMA model. This model integrates the C2fCIB module to enhance small-object feature extraction capabilities and incorporates the C3k2_Mambaout module to optimize the discriminative fusion of multi-scale features. Ablation experiments demonstrated that the synergistic operation of C2fCIB and C3k2_Mambaout improved detection performance. For insect-damaged fruit detection, mAP50 and precision were boosted by 2.7% and 5.8%, respectively, compared to the baseline model, reaching 0.639 and 0.91. When compared to other YOLO variants, YOLO-CMA achieved the lowest computational cost (5.9 GFLOPs) and parameter count (2.43 M) while maintaining competitive detection precision, demonstrating significant edge deployment advantages. This study provided a comprehensive technical solution—from data construction to model optimization—for addressing the challenge of detecting minute defects in agricultural products, offering substantial practical value and application potential.

Figures