Adhered Buckwheat Seed Segmentation Method Based on Improved UNet

To address the issue of adhesion segmentation caused by the small volume, diverse morphology, large quantity, and fuzzy adherence boundaries of seeds in the output image of the buckwheat hulling machine, this paper proposes a semantic segmentation model, ResCo-UNet. The model integrates the ResNet18 network structure in the encoder of UNet, enhancing feature extraction capabilities and accelerating network training speed. To improve the recognition of adhered target boundaries, a novel parallel attention mechanism, CA², is designed and incorporated into ResNet18, thereby enhancing the extraction of high-level semantic information. In the decoder stage, ConvNeXt modules are introduced to expand the receptive field, enhancing the model's ability to reconstruct complex boundary features. Results demonstrate that ResCo-UNet exhibits stronger generalization capabilities compared to other models, showing significant enhancements across multiple metrics: 87.81% mIoU, 92.71% recall, and 93.33% F1-score. Compared to the baseline model, these metrics increased by 6.71%, 5.13%, and 4.32%, respectively. Analysis of detection results across images with different distribution densities revealed an average counting accuracy of 98.89% for adherent seeds. The model effectively segments adhered seed images with varying density distributions, providing reliable parameter feedback for adaptive control of intelligent hulling equipment.