Automatic Detection of Small Sample Apple Surface Defects Using ASDINet.

Area of Science:

• Computer Vision
• Agricultural Technology
• Machine Learning

Background:

• Apple appearance quality significantly impacts market price.
• Automated apple grading requires accurate surface defect detection.
• Existing methods struggle with low recognition rates in small sample scenarios.

Purpose of the Study:

• To develop an effective apple surface defect detection network (ASDINet) for small sample learning.
• To improve the accuracy and speed of automated apple grading systems.
• To address the challenge of low recognition rates in detecting various apple defects.

Main Methods:

• Designed an apple surface defect detection network (ASDINet) incorporating an AU-Net segmentation network.
• Integrated a Dep-conv module to enhance receptive field feature capacity and a global spatial domain attention mechanism (GSAM) within a global decision module (GDM).
• Optimized for real-time segmentation by modifying feature map processing and utilizing mask outputs for fast prediction.

Main Results:

• ASDINet achieved a 98.8% AP and 97.75% F1-score, surpassing most state-of-the-art models.
• The network demonstrated a detection speed of 39ms per frame, suitable for practical deployment.
• Effective performance was maintained even with limited training data (42 defective images).

Conclusions:

• ASDINet offers a robust solution for small sample apple surface defect detection.
• The network's high accuracy, speed, and data efficiency align with actual production requirements for automated grading.
• The study validates the practical applicability and excellent performance of ASDINet in real-world scenarios.