Segmentation of Unsound Wheat Kernels Based on Improved Mask RCNN

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Wheat quality grading relies on identifying unsound kernels, a process currently inefficient due to manual sorting.
• Traditional image processing methods struggle with separating clumped or adherent wheat kernels.
• Accurate and rapid detection of unsound kernels is crucial for effective wheat quality evaluation.

Purpose of the Study:

• To develop a novel algorithm for rapid and accurate recognition of unsound wheat kernels.
• To address the challenges of manual inspection and limitations of traditional image processing in kernel separation.
• To improve the efficiency and precision of wheat grading systems.

Main Methods:

• An improved Mask R-CNN model was proposed, incorporating a bottom-up pyramid network to enhance low-level feature information.
• An attention mechanism (AM) module was integrated to boost detection accuracy for small or occluded targets.
• The regional proposal network (RPN) was optimized to refine prediction performance for unsound kernel identification.

Main Results:

• The improved Mask R-CNN algorithm demonstrated faster and more accurate identification of unsound wheat kernels.
• The model effectively handled issues with adherent kernels, improving segmentation capabilities.
• Achieved a precision of 86% and a recall of 91%, with an inference time of 7.83 seconds for images with approximately 200 targets.

Conclusions:

• The developed algorithm significantly improves the speed and accuracy of unsound wheat kernel detection compared to existing methods.
• This approach offers a robust solution for automated wheat grading, addressing key limitations in current practices.
• The findings lay a strong foundation for advancing automated quality assessment in the grain industry.