Invalid point removal method based on error energy function in fringe projection profilometry

A Markov random field based method for removing invalid unwrapping phase points in 3D reconstruction

“…To evaluate our model, we compare Zhang's method [10], Feng's method [12], Song's method [18], Zhu's method [15], and our method, and the experimental parameters are shown in Table 1. Zhang's method and Feng's method are both classical phase denoising algorithms based on local neighbourhood; Song's method is a global denoising algorithm; Zhu's method is a recently proposed phase denoising algorithm.…”

“…Feng et al [12] used a Gaussian filter to detect invalid points; Wang et al [13] used monotonicity test, RMSE check, and 3D point cloud smoothness check to remove noise object points. Zhang et al [14] used a modulation-level histogram of the image instead of a grayscale histogram to segment the background and shadow points; Zhu et al [15] used Gaussian-weighted Euclidean distance to define the invalid energy function, and removed points with error energy greater than a certain invalid threshold as invalid points. While these techniques can represent the relationship between neighbouring phase points well and have been successful in most scenarios, so far they have not been able to contain information outside of a given neighbourhood, resulting in insufficient global learning of the model, which will lead to unreasonable denoising in real-world complex scenarios, and if the threshold is not selected well, some valid points will also be eliminated.…”

“…Adaptive threshold selection has always been the direction of phase map denoising [8,11,13,15], and excellent threshold selection algorithm greatly improves the generalization ability of the whole model. We believe that the energy minimization function usually keeps the energy at the right point in a low range, while the energy at the wrong point will be high.…”

“…To evaluate our model, we compare Zhang's method [10], Feng's method [12], Song's method [18], Zhu's method [15], and our method, and the experimental parameters are shown in Table 1.…”

“…Some methods introduce additional auxiliary hardware to eliminate errors [9], which increases the complexity of the system. The more mainstream approaches are to use phase neighbourhood processing and assume similar differences within the neighbourhood of phase points [8,[10][11][12][13][14][15]. The fixed spatial extend of these methods limit the range in which information can be propagated.…”