| 摘 要: 针对蜂窝状植物细胞图像包含大量噪声、细胞结构复杂以及细胞边界不清晰等问题,提出了一种基于WT-SR和改进 U-Net的分割模型。利用小波变换和稀疏编码,将沿Z轴不同层的图像序列融合生成信息更丰富的单张图像,使用改进 U-Net对图像进行分割。在 U-Net编码过程中,将小波变换提取的低频信号与网络层特征结合,增强对细胞边界的辨识能力。引入混合注意力模块,融合通道信息和空间信息,提高有效特征的表达能力。在整个网络中使用残差连接,加强浅层和深层信息的融合,改善分割效果。通过两组噪声含量不同的数据进行实验,结果显示所提模型在 MIoU,Dice和 Acc指标上平均达到了85.85%、90.99%、97.98%,相较于原始 U-Net模型,提高了8.49%、7.91%、8.07%,由此验证了该模型的有效性和可行性。 |
| 关键词: 小波变换 图像分割 U-Net 注意力机制 |
|
中图分类号: TP391.41
文献标识码: A
|
| 基金项目: 湖南省自然科学基金部门联合基金资助(2023JJ60194) |
|
| Plant Cell Image Segmentation Based on WT-SR and Improved U-Net |
|
LI Jieqin1, XIE Dingfeng1, WANG Xueping2
|
(1.College of Information Engineering, Hu’nan Industry Polytechnic, Changsha 410000, China;
2.College of Information Science and Engineering, Hu’nan Normal University, Changsha 410081, China)
ijieqin4568@163.com; coolboyxie@163.com; 1139590669@qq.com
|
| Abstract: Aiming at the problems of honeycomb plant cell images containing heavy noise, complex cell structures, and unclear cell boundaries, a segmentation model based on WT-SR and an improved U-Net was proposed.By leveraging wavelet transform and sparse coding, image sequences along the Z-axis at different layers were fused to generate a single image with richer information. The improved U-Net was used for image segmentation. During the U-Net encoding process, low-frequency signals extracted via wavelet transform were combined with network layer features to enhance the identification capability of cell boundaries. A hybrid attention module was introduced to integrate channel information and spatial information, improving the representation of effective features. Residual connections were employed throughout the network to strengthen the fusion of shallow and deep information, thereby enhancing segmentation performance. Experiments were conducted on two datasets with varying noise levels, and the results showed that the proposed model achieved average MIoU, Dice, and Acc metrics of 85.85% , 90.99% , and 97.98% , respectively. Compared with the original U-Net model, these metrics improved by 8.49% , 7.91% , and 8.07% ,demonstrating the effectiveness and feasibility of the proposed model. |
| Keywords: wavelet transform image segmentation U-Net attention mechanism |
