Wang Wensheng, Su Chang
School of Mechanical and Electrical Engineering, Beijing Information Science and Technology University, Beijing, China; Key Laboratory of Modern Measurement and Control Technology, Ministry of Education, Beijing Information Science and Technology University, Beijing, China.
Changchun Institute of Optics, Fine Mechanics and Physics, Chinese Academy of Sciences, Changchun, China.
ISA Trans. 2022 Dec;131:650-661. doi: 10.1016/j.isatra.2022.05.005. Epub 2022 May 13.
Since the image sensor will produce blur problems in the process of collecting data of moving objects, the image needs to be restored. Ringing is one of the most common artifacts in deblurred images. This paper proposes a non-blind image deconvolution method based on texture mapping segmentation, named texture-Richardson-Lucy (TRL) algorithm, which suppresses ringing while deblurring the image. TRL is based on a novel ringing removal deconvolution algorithm, which adds a ringing detection term as regularization in the iterative process of the Richardson-Lucy algorithm. Taking into account the structural difference between the texture and the flat area, the image is segmented into several blocks and restored through adaptive iterative texture maps based on the pixel intensity and texture features of the image. In order to obtain a reasonable texture map, a Gaussian mixture model is used to fit the pixel intensity distribution, and use the expectation maximization algorithm and local binary mode to estimate. Experimental results and quantitative evaluations show that TRL can effectively reduce ringing artifacts while retaining details and achieving robustness to suppress ringing of different blur kernels. The processing time of a single 1 million pixel image in an 8-core CPU environment is about 3.5 s. And the PSNR and SSIM parameters are above 30 dB and 0.92, respectively. In conclusion, TRL is superior to the current popular algorithms.
由于图像传感器在采集运动物体数据的过程中会产生模糊问题，因此需要对图像进行恢复。振铃是去模糊图像中最常见的伪影之一。本文提出了一种基于纹理映射分割的非盲图像去卷积方法，即纹理-理查森-露西（TRL）算法，该算法在对图像去模糊的同时抑制振铃。TRL基于一种新颖的去除振铃的去卷积算法，该算法在理查森-露西算法的迭代过程中添加了一个振铃检测项作为正则化。考虑到纹理和平坦区域之间的结构差异，将图像分割成几个块，并根据图像的像素强度和纹理特征通过自适应迭代纹理映射进行恢复。为了获得合理的纹理映射，使用高斯混合模型拟合像素强度分布，并使用期望最大化算法和局部二值模式进行估计。实验结果和定量评估表明，TRL可以有效地减少振铃伪影，同时保留细节，并对不同的模糊核实现抑制振铃的鲁棒性。在8核CPU环境下处理单个100万像素图像的时间约为3.5秒。并且PSNR和SSIM参数分别高于30dB和0.92。总之，TRL优于当前流行的算法。
