Stereo matching is one of the most active research areas in the field of computer vision. Stereo matching aims to obtain 3D information by extracting correct correspondence between two images captured from different point of views. There are two research parts in stereo matching: similarity measure between correspondence points and optimization technique for dence disparity estimation.
The crux of stereo matching problem in similarity measure perspective is how to deal with the inferent points ambiguity that results from the ambiguous local appearances of image points. Similarity measures in stereo matching are classified as feature-based, intensity-based or non-parametric measure. And most similarity measures in the literatures are based on pixel intensity comparison. When images are taken at different illumination conditions or different sensors used, it is very unlikely that the corresponding pixels would have the same intensity creating false correspondences if it is only based on intensity matching functions alone. Especially illumination variations between input images can cause serious degrade in the performance of stereo matching algorithms. In this situation, mutual information-based method is powerful. However, it is still ambiguous or erroneous in considering local illumination variations between images. Therefore, similarity measure to these radiometric variations are demanded and become inevitable for stereo matching.
Optimization method in stereo matching can be classified into two categories: local and global optimization methods, and most state-of-the-art algorithms fall into global optimization method. Global optimization methods can greatly suppress the matching ambiguities caused by various factors such as occluded and textureless regions. However, They are usually computationally expensive due to the slow-converging optimization process.
In this paper, it was proposed that a stereo matching similarity measure based on entropy and census transform and an optimization technique using dynamic programming to estimate disparity efficiently based on multi-resolution method. Proposed similarity measure is composed of entropy, Haar wavelet feature vector, and modified Census transform. In general, mutual information similarity measure based on entropy about stereo images and disparity map is a popular and powerful similarity measure which is robust to complex intensity transformation. However, it is still ambiguous or erroneous with local radiometric variations, since it only accounts for global variation between images, and does not contain spatial information. Haar wavelet response can express frequency properties of image regions and is robust to various intensity changes and bias. Therefore, entropy was utilized with Haar wavelet feature vector as geometric measure. Modified Census transform was used as another spatial similarity measure. Census transform is a well-known non-parametric measure. And it is powerful to textureless and disparity discontinuity region and robust to noisy environment. A combination of entropy with Haar wavelet feature vector and modified Census transform as similarity measure was proposed to find correspondence. It is invariant to local radiometric variations and global illumination changes, so it can be applied to find correspondence for images which undergo local as well as global radiometric variations.
Proposed optimization method is a new disparity estimation technique based on dynamic programming. A method using dynamic programming with 8-direction energy aggregation to estimate accurate disparity map was applied. Using 8-direction energy aggregation, accurate disparities can be found at disparity discontinuous region and suppress a streaking phenomenon in disparity map.
Finally, the multi-resolution scheme was proposed to increase efficiency while processing and disparity estimation method. A Gaussian pyramid which prevent the ailasing at low-resolution image pyramid levels was used. And the multi-resolution scheme was proposed to perform matching at every levels to find accurate disparity. This method can perform matching efficiently and make accurate disparity map.
And proposed method was validated with experimental results on stereo images.