Enhanced YOLO-based framework for accurate detection and identification of common wheat impurities with distinct...

Area of Science:

• Agricultural Engineering
• Computer Vision
• Food Science

Background:

• Accurate detection of wheat grain impurities is vital for storage, milling, and harvesting.
• Real-time detection requires balancing accuracy with processing speed.

Purpose of the Study:

• To evaluate the performance of YOLOv5n, YOLOv5x, YOLOv8n, and YOLOv8x models for detecting wheat grain impurities.
• To determine optimal algorithms and resolutions for both laboratory and real-time applications.

Main Methods:

• Trained four YOLO models (YOLOv5n, YOLOv5x, YOLOv8n, YOLOv8x) on 700 labeled images across three resolutions.
• Evaluated models based on detection accuracy (mAP@50) and processing speed.

Main Results:

• Larger YOLO models (YOLOv5x, YOLOv8x) showed similar performance for lab applications but reduced speed.
• YOLOv5n at 320x320 resolution improved detection speed by 4% while maintaining accuracy, ideal for real-time use.
• High mAP@50 achieved for visually similar impurities (85-88%) and others (>95%).

Conclusions:

• YOLO models are effective for non-destructive wheat impurity detection.
• Model selection depends on application needs: YOLOv5n for real-time, YOLOv5x/YOLOv8x for lab analysis.
• The approach can be adapted for impurity detection in other grains.