SY-Net: A Rice Seed Instance Segmentation Method Based on a Six-Layer Feature Fusion Network and a Parallel...

Area of Science:

• Agricultural Engineering
• Computer Vision
• Image Processing

Background:

• Accurate rice seed segmentation is crucial for quality testing.
• Challenges include similar seed characteristics, small size, and dense distribution.
• Existing methods struggle with precise target segmentation in complex rice seed samples.

Purpose of the Study:

• To propose SY-net, a novel network for precise rice seed instance segmentation.
• To enhance feature extraction, fusion, and mask generation for improved accuracy.
• To evaluate SY-net's performance on public and private rice seed datasets.

Main Methods:

• Developed SY-net with four modules: feature extractor, feature pyramid fusion, prediction head, and prototype mask generation.
• Utilized a transformer backbone for enhanced feature learning.
• Implemented a six-layer feature fusion network and parallel prediction heads.
• Employed a large feature map for high-quality mask generation.

Main Results:

• SY-net achieved 90.71% mean average precision (mAP) on a private rice seed dataset.
• Demonstrated 16.5% average precision (AP) for small targets in the COCO2017 dataset.
• Significantly improved the efficiency of rice seed segmentation.

Conclusions:

• SY-net offers a robust solution for rice seed instance segmentation.
• The network shows strong potential for applications in automated rice quality testing.
• SY-net effectively addresses challenges posed by small and densely packed rice seeds.