Accurate and efficient extraction of fringe orientation from the poor-quality ESPI fringe pattern with a convolutional neural network

Digital phase-shift method based on distance mapping for phase recovery of an ESPI fringe pattern

In view of the limitation of the traditional method to recover the phase of the single fringe pattern, we propose a digital phase-shift method based on distance mapping for phase recovery of an electronic speckle pattern interferometry fringe pattern. First, the direction of each pixel point and the centerline of the dark fringe are extracted. Secondly, the normal curve of the fringe is calculated according to the fringe orientation to obtain the fringe moving direction. Thirdly, the distance between each pixel point and the next pixel point in the same phase is calculated by a distance mapping method according to the adjacent centerlines; then the moving distance of the fringes is obtained. Next, combining the moving direction and moving distance, the fringe pattern after the digital phase shift is obtained by full-field interpolation. Finally, the full-field phase corresponding to the original fringe pattern is recovered by four-step phase shifting. The method can extract the fringe phase from a single fringe pattern through digital image processing technology. The experiments show that the proposed method can effectively improve the phase recovery accuracy of a single fringe pattern.

DeepOrientation: convolutional neural network for fringe pattern orientation map estimation

Fringe pattern based measurement techniques are the state-of-the-art in full-field optical metrology. They are crucial both in macroscale, e.g., fringe projection profilometry, and microscale, e.g., label-free quantitative phase microscopy. Accurate estimation of the local fringe orientation map can significantly facilitate the measurement process in various ways, e.g., fringe filtering (denoising), fringe pattern boundary padding, fringe skeletoning (contouring/following/tracking), local fringe spatial frequency (fringe period) estimation, and fringe pattern phase demodulation. Considering all of that, the accurate, robust, and preferably automatic estimation of local fringe orientation map is of high importance. In this paper we propose a novel numerical solution for local fringe orientation map estimation based on convolutional neural network and deep learning called DeepOrientation. Numerical simulations and experimental results corroborate the effectiveness of the proposed DeepOrientation comparing it with a representative of the classical approach to orientation estimation called combined plane fitting/gradient method. The example proving the effectiveness of DeepOrientation in fringe pattern analysis, which we present in this paper, is the application of DeepOrientation for guiding the phase demodulation process in Hilbert spiral transform. In particular, living HeLa cells quantitative phase imaging outcomes verify the method as an important asset in label-free microscopy.

BM3D adaptive TV filtering-based convolutional neural network for ESPI image denoising

Image denoising is a fundamental part of image processing. The real electronic speckle pattern interferometry (ESPI) contains a large amount of speckle noise, which affects the image quality and adversely affects subsequent studies. In this paper, a method based on an improved denoising convolutional neural network (CNN) has been proposed, with the goal of reducing noise while maintaining accurate information. The block matching 3D-based adaptive TV denoising CNN can protect the valid information while preventing the information of the original image itself from being corrupted. A two-channel model is used to improve the noise reduction effect of real images. The proposed method is compared with the conventional denoising algorithms and the deep-learning denoising algorithms. Experimental results show that the proposed method can maintain accuracy, integrity, and stability while preserving the details, texture, and edge information of the stripe pattern.

Binarization of ESPI fringe patterns based on an M-net convolutional neural network

The fringe skeleton method is the most straightforward method to estimate phase terms in electronic speckle pattern interferometry (ESPI). It usually needs to binarize the fringe patterns. However, the massive inherent speckle noise and intensity inhomogeneity in ESPI fringe patterns make it difficult to binarize the ESPI fringe patterns. In this paper, we propose a binarization method for ESPI fringe patterns based on a modified M-net convolutional neural network. Our method regards the binarization of fringe patterns as a segmentation problem. The M-net is an excellent network for segmentation and has proven to be a useful tool for skeleton extraction in our previous work. Here we further modify the structure of the previous network a bit to suit our task. We train the network by pairs of ESPI fringe patterns and corresponding binary images. After training, we test our method on 20 computer-simulated and three groups of experimentally obtained ESPI fringe patterns. The results show that even for fringe patterns with high noise and intensity inhomogeneity, our method can obtain good binarization results without image preprocessing. We also compare the modified M-net with a classic segmentation network, the U-net, and a residual encoder-decoder network (RED-net). The RED-net was used for binarization of document images. The experimental results prove the effectiveness of our method.

Skeleton extraction and inpainting from poor, broken ESPI fringe with an M-net convolutional neural network

Extracting skeletons from fringe patterns is the key to the fringe skeleton method, which is used to extract phase terms in electronic speckle pattern interferometry (ESPI). Because of massive inherent speckle noise, extracting skeletons from poor, broken ESPI fringe patterns is challenging. In this paper, we propose a method based on a modified M-net convolutional neural network for skeleton extraction from poor, broken ESPI fringe patterns. In our method, we pose the problem as a segmentation task. The M-net performs excellent segmentation, and we modify its loss function to suit our task. The broken ESPI fringe patterns and corresponding complete skeleton images are used to train the modified M-net. The trained network can extract and inpaint the skeletons simultaneously. We evaluate the performance of the network on two groups of computer-simulated ESPI fringe patterns and two groups of experimentally obtained ESPI fringe patterns. Two related recent methods, the gradient vector fields based on variational image decomposition and the U-net based method, are compared with our method. The results demonstrate that our method can obtain accurate, complete, and smooth skeletons in all cases, even where fringes are broken. It outperforms the two compared methods quantitatively and qualitatively.