基于曲率信息的乘性噪声图像自适应恢复与分割算法研究与应用

The restoration of multiplicative noisy image has very important applications in the fields of medical imaging. The existing methods are usually only focus on the question of multiplicative noise, but in practice, blurred degradation is universal. Those algorithms often cannot provide satisfactory results for the task of restoring the image which is contaminated seriously by multiplicative noise and blurring kernel, particularly the unknown blurring kernels. On the other hand, the existing multiplicative noise models and the algorithms are mostly based on the total variation and its improved model, which cannot provide satisfactory results for the images with piecewise constant character. To solve these problems, the project intends to further study the solution of multiplicative noisy image restoration and segmentation based on Gaussian curvature and mean curvature information of the surface of image with regularization method. The main contents include: multiplicative noisy image denoising with adaptive regularization parameters, multiplicative noisy blurred image restoration for the problems of non-blind deblurring and blind deblurring, and the robust segmentation of the multiplicative noisy blurred image. At the same time, we will also focus on the application of the proposed algorithms in medical ultrasound image restoration. The project will improve the framework of adaptive denoising and blur kernel estimation, blind deblurring and robust image segmentation under multiplicative noise. It will give full play to the advantages of curvature regularization to achieve efficient and accurate restoration and segmentation of the degraded image. And it will also effectively improve the visual effect of the image, the segmentation accuracy and objective evaluation index, beyond the existing the model and algorithm in the aspects of subjective and objective indicators. Therefore the research of the project has the important theory significance and the practical value.

乘性噪声图像恢复在医学成像领域有着重要的应用。现有方法较多关注乘性噪声去除问题，对于实际中普遍存在的同时受乘性噪声和模糊影响，尤其是未知模糊影响而严重退化的图像，不能得到满意的恢复结果。另外，现有方法多是基于全变分及其改进模型，对非分片常数图像的恢复结果较差。针对以上不足，本项目拟在前期工作基础上，进一步深入研究应用图像曲面的平均、高斯曲率信息，基于正则项方法解决乘性噪声图像恢复和分割问题，具体包括乘性噪声自适应去除、乘性噪声下(盲)去模糊和乘性噪声模糊图像的鲁棒分割，并将所提算法应用于临床超声影像的自适应恢复与分割问题中。本项目将发展并完善乘性噪声下的自适应去噪、模糊核估计、盲去模糊和鲁棒图像分割理论框架；充分发挥曲率正则项的优势，对退化图像实现高效而准确的恢复与分割，有效提高图像的视觉效果、分割精度及客观指标，在主观评价和客观指标两方面超越现有模型和算法，具有重要的理论意义和实用价值。

图像恢复问题在医学、国防等领域都有着非常重要的应用前景。本项目我们关注数学工具与机器学习方法在图像处理、分割、识别等问题中的应用。首先提出了一种多映射残差卷积神经网络(MMRCNN)实现对噪声、模糊、分辨率不足退化图像进行有效的恢复，不需要预处理和人工干预，就能够直接完成清晰图像的自动恢复重建。同时重点研究了X-射线、CT等模态医疗影像的分割、处理、识别等问题：提出了基于数学工具与机器学习方法的X射线肺炎（二分类）以及新冠肺炎医疗影像诊断分析算法，通过 X 射线图像挖掘COVID-19 和常规肺炎的深度学习结构，实现精准检测；对胸部CT图像肺结节检测算法进行了深入研究，提出了基于尺寸自适应深度神经网络的肺结节检测算法，可以快速有效地检测出潜在的危险结节。通过临床数据的对比实验结果表明，在检测小结节方面，该系统的性能可与经验丰富的放射科医生相媲美；为了实现对血肿区域的准确、快速分割，项目负责人和参与人提出了改进注意力 U-net 的微小脑出血计算机断层扫描（CT）影像自动分割算法研究。此外，我们对数学理论、机器学习方法在生物信息科学中的相关问题也进行了研究，提出了基于矩阵补充和岭回归的集成方法实现抗癌药物反应的预测；提出了基于高斯混合和BP神经网络的卵巢癌质谱数据三分类模型。