Advancing Rice Grain Impurity Segmentation with an Enhanced SegFormer and Multi-Scale Feature Integration.

Area of Science:

• Agricultural Engineering
• Computer Vision
• Machine Learning

Background:

• Accurate identification of rice grains versus impurities (straw, leaves) is challenging during harvesting due to occlusion and adhesion.
• Existing methods struggle with precise segmentation in complex agricultural environments.

Purpose of the Study:

• To develop a lightweight semantic segmentation algorithm for improved impurity detection in rice harvesting.
• To enhance the accuracy and efficiency of distinguishing rice grains from contaminants.

Main Methods:

• An improved SegFormer network was designed for lightweight semantic segmentation.
• The decoder was redesigned incorporating Feature Pyramid Network (FPN) for feature fusion.
• A Part Large Kernel Attention (Part-LKA) module and Bottleneck Recursive Gated Convolution (B-gnConv) were introduced for enhanced feature focus and spatial interaction.

Main Results:

• The improved model demonstrated increased accuracy in segmenting rice grains and impurities.
• Pixel accuracy (PA) increased by 1.6%, and the F1 score improved by 3.1% compared to the original SegFormer model.
• The algorithm achieved effective segmentation, simplifying the model and accelerating computation.

Conclusions:

• The proposed lightweight semantic segmentation algorithm offers a valuable solution for real-time impurity monitoring in rice combine harvesters.
• The enhanced SegFormer network provides a robust algorithmic reference for agricultural automation and quality control.