融合稀疏表示和字典学习的脑部 MR 图像分割

摘要

For segmentation accuracy problem of brain MR image in regard to white matter, gray matter and cerebrospinal fluid, we proposed an image segmentation method which fuses the sparse representation and dictionary learning.First, it trains the over-completed dictionary using block-based input data.Then, it obtains the high-dimensional feature represented by optimal sparse according to the dictionary learnt.Finally, it implements the segmentation by combining the local and nonlocal reconstruction errors of each pixel.Results of experiment on simulated and real image database show that the proposed method can use the formula with distance factor and sparse factor to accurately segment MR images, and is superior to other MR segmentation methods in terms of stability.%针对脑部磁共振图像中白质、灰质和脑脊液的分割精度问题，提出一种融合稀疏表示和字典学习的图像分割方法。首先，利用基于块的输入数据来训练过完备字典；然后，根据学习到的字典获得最优稀疏表示的高维特征；最后，结合每个像素局部和非局部重构误差实现分割。在模拟和真实图像数据库上的实验结果表明，该方法能利用带有距离因子和稀疏因子的公式准确分割MR图像，在稳定性方面优于其他MR分割方法。