Highlight Removal And Photometric Stereo Reconstruction Of Non-Lambertian Surfaces

The non-Lambertian surfaces not only have diffuse reflection,but also have specular reflection.The highlight regions affect the coherence of underlying textures and colors,which leads to inaccurate evaluation of geometry in image-based 3D reconstructions.Therefore,it is important to effectively analyze and use non-Lambertian reflectance in reconstruction tasks.Most highlight removal algorithms for non-Lambertian surfaces can easily cause image distortion and texture loss which reduce the input data quality for 3D reconstruction;The traditional uncalibrated photometric stereo vision assumes Lambertian surfaces,which causes large error when estimating the normals.In this paper,we focus on the non-Lambertian surfaces.We propose a new clustering algorithm for highlight removal to relieve the texture loss and color distortion,and an uncalibrated photometric stereo algorithm to improve normal estimation accuracy.(1)An improved clustering algorithm for highlight removal.firstly,the single channel image with minimum intensity value is subtracted from the original image based on dichromatic reflection model to obtain an initial specular-free image.Secondly,the minimum and maximum diffuse chromaticity values for each pixel in the highlight area is estimated according to the initial specular-free image.Finally,the distribution pattern of the pixels in the highlight area are analyzed in a minimum-maximum chromaticity space and clustered by x-means method.The specular components of highlight area pixels can be easily separated by using the estimated intensity ratio of diffuse pixels,thereby getting an image without highlights.Experimental results show that,compared with other methods,the peak signal-to-noise ratio increases by 2%to 4%on average,the structural similarity index increases by 1.4%to 4.5%on average.(2)An uncalibrated photometric stereo algorithm based on generative adversarial networks.Existing photometric stereo algorithms usually assume that the light conditions are known or simplify the reflectivity model,which will limit the practical application of photometric stereo algorithms for materials with complex reflection properties.The proposed uncalibrated photometric stereo method based on generative adversarial networks directly learns the mapping between non-Lambertian pixels and its normal,avoiding complex calibration steps of the light source,and does not need to accurately estimate the explicit form of the reflectance function.The generated model only needs to input any number of photometric images to estimate the normals.Experimental results show that compared with other uncalibrated photometric methods,the proposed algorithm reduces the average angle error to 14.98,the mean square error to 9.65,and the accuracy of normal reconstruction is improved.