Green Apple Detection Method Based on Multidimensional Feature Extraction Network Model and Transformer Module.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Food Science

Background:

• Accurate detection of pollution-free green apples is crucial for food safety.
• Existing methods may struggle with variations in scale, angle, and near-color detection.
• The DETR (Detection Transformer) network offers a promising framework but requires optimization.

Purpose of the Study:

• To develop a novel, enhanced DETR-based method for fast and accurate detection of pollution-free green apples.
• To improve the model's adaptability to variations in fruit appearance and environmental conditions.
• To enable efficient, AI-driven robotic harvesting for food safety.

Main Methods:

• An improved DETR network incorporating ResNet18 with deformable convolutions (DCNv2) for feature extraction.
• Integration of Extended Spatial Pyramid Pooling (DSPP) and Feature Refinement Aggregation Module (FRAM) to reduce noise and preserve features.
• Transformer attention mechanism optimization to accelerate convergence and reduce computational load.

Main Results:

• The proposed method significantly improved detection accuracy, with AP50 reaching 97.12% compared to the original DETR.
• When deployed on a picking robot, the model achieved 96.58% average detection accuracy and a 51% increase in detection speed.
• The detection rate for non-polluted fruits exceeded 0.95, demonstrating effective application in robotic harvesting.

Conclusions:

• The enhanced DETR network provides a robust solution for detecting pollution-free, near-color fruits.
• This AI-driven approach shows great potential for improving food safety and automating fruit picking processes.
• The method ensures efficient and accurate identification of pollution-free fruits, supporting AI applications in food safety.