FSUNet: lightweight full-scale information fusion UNet for seed coat thickness measurement

With the rapid development of seed research, understanding the external structure and characteristics of plants has become essential. Seed coat thickness plays a key role in assessing seed quality in seed reproduction, storage, processing, and research. However, traditional manual measurement methods are labor-intensive, subjective bias, and measurement errors. To solve this problem , we propose an efficient and universal deep learning method, FSUNet. Base on the UNet model, FSUNet introduces a new Full-Scale Information Fusion module (FSIF), adopting 7 × 7 depth-wise separable convolutions and point convolutions to capture remote information, and designs the MConv module. The FSUNet model has 5.109 M parameters, only 14.79% of the UNet model. This research validated the method with a self-constructed corn seed coat segmentation dataset. FSUNet achieved segmentation results of 69.37%, 67.80%, 83.85%, and 85.40% on the self-constructed dataset and the three public datasets, respectively, which are 0.24%, 1.95%, 1.51%, and 0.73% higher than those of the UNet model. Compared to other recent lightweight models like CMUNext, FSUNet shows significant advantages. In addition, we also provide a seed coat thickness measurement algorithm that can obtain stable and accurate measurement results. Through our method, the measurement efficiency and accuracy of seed coat thickness can be significantly improved, providing strong tools and technical support for plant seed research.