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

Area of Science:

• Agricultural Technology
• Computer Vision
• Machine Learning

Background:

• Accurate detection of fruit defects is crucial for quality control in agriculture.
• Existing models struggle with identifying subtle defects like insect damage on Sanhua plums.

Purpose of the Study:

• To develop a precise and efficient model for detecting abnormal Sanhua plums, focusing on small defects.
• To create a specialized dataset and optimize a deep learning model for fruit quality assessment.

Main Methods:

• Constructed a specialized dataset of 10,000 images across five categories (diseased, insect-damaged, bird-pecked, cracked, normal).
• Employed multi-weather simulation data augmentation to address class imbalance.
• Developed the YOLO-CMA model, integrating C2fCIB and C3k2_Mambaout modules for enhanced small-object detection and feature fusion, based on the YOLOv12 architecture.

Main Results:

• The YOLO-CMA model significantly improved detection for insect-damaged fruit, boosting mAP50 by 2.7% and precision by 5.8% over the baseline YOLOv12.
• Achieved state-of-the-art performance with the lowest computational cost (5.9 GFLOPs) and parameter count (2.43 M) among YOLO variants.
• Demonstrated superior detection precision and efficiency for minute defects in agricultural products.

Conclusions:

• The YOLO-CMA model provides a robust and efficient solution for detecting subtle defects in agricultural products.
• The proposed model has significant advantages for edge deployment in real-world agricultural quality control scenarios.
• This research offers practical value and potential for widespread application in automated fruit inspection.