The image depth estimation based on multi-scale texture features and least-square method
Lizhi Zhang, Tingting Chen, Huadong Sun, Zhijie Zhao, Xuesong Jin, Ju Huang · 2014
The technique of 3D scene reconstruction from 2D scene image can satisfy the demands of 3D display market, and it is the main research contents in the field of image processing. For the 3D reconstruction, the extraction of depth cues is one of the most important contents. Therefore, we proposed a least-square depth estimation method based on multi-scale texture features, which can better describe the relationship between the texture characteristics and the scene depth. first we captured the texture energy of different scales as the features of the image using Laws' filters to image texture gradients, texture variations and colors, respectively, and then we trained these features of the training sample sets to get the relationship parameters between texture cues and scene depth, which can finally be applied to estimate the depth of the test sample sets. The experimental results show that the method we proposed performs well in scene depth information extraction, meanwhile it requires less constraint for scene structure and space layout of 2D images. Consequently, it has good robustness because it need no specific analysis to the images with special depth cues such as linear perspective.