Online Detection System for Wheat Machine Harvesting Impurity Rate Based on DeepLabV3

Area of Science:

• Agricultural Engineering
• Computer Vision
• Machine Learning

Background:

• Mechanized wheat harvesting relies on combine harvesters, where impurity rate is a key quality indicator.
• Accurate, real-time assessment of wheat impurity rates is crucial for optimizing harvest quality and efficiency.

Purpose of the Study:

• To design and evaluate a vision system for online detection of wheat impurity rates during mechanized harvesting.
• To identify the optimal deep learning model for segmenting wheat grains and impurities.

Main Methods:

• Developed a vision system utilizing the DeepLabV3+ deep learning model.
• Trained and compared DeepLabV3+ with four backbones: MobileNetV2, Xception-65, ResNet-50, and ResNet-101.
• Evaluated models based on accuracy, comprehensive evaluation index, and average intersection ratio for grain and impurity identification.

Main Results:

• The ResNet-50 backbone demonstrated superior performance in recognition and segmentation.
• ResNet-50 achieved high accuracy rates for grain (86.86%) and impurity (89.91%) identification.
• The system exhibited low detection errors in both bench (max absolute error 0.2%) and field tests (max absolute error 0.06%).

Conclusions:

• The DeepLabV3+ model with the ResNet-50 backbone provides an effective and real-time solution for measuring wheat impurity rates during mechanized harvesting.
• This technology facilitates immediate quality assessment, potentially improving wheat yield and quality management.