Due to the position differences of two eyes, there exists a difference between the projections of an object on the two retinas, which is called disparity. Stereo matching aims to obtain the disparity accurately and robustly, which is an important research topics in computer vision. Although some remarkable results have been achieved by those algorithms which are based largely on pure mathematics or computational practicality, they are unable to perform as good as human stereopsis in terms of speed, accuracy and robustness. Therefore, a neurophysiology inspired computational model is solicited. This thesis is to explore newly developed theories about the cortical disparity detection and propose neurophysiologically sound methods for stereo matching. The main contributions and novelties are as follows: 1. We studied and reviewed related neurophysiological progresses about stereo perception, including the low level areas (V1, V2, V3), dorsal areas (MT, MST, IPS) and ventral areas (V4, IT). In addition to these physiological studies, the disparity energy model and its extensions are also reviewed. 2. The disparity tuning responses of V1 neurons have been well described by the disparity energy model. However, this model fails to explain a physiological finding that these neurons should have weaker responses to binocularly anti-correlated random dot stereograms than to random dot stereograms. A weighted disparity energy model is proposed to tackle this problem. Under this model, the responses of the neurons are modulated by making use of the signal differences within the left and right receptive fields. Then, the population responses are computed based on the responses of individual neurons and interaction among them for disparity computation. The main contributions are two-fold: (1) it can adequately describe the weaker responses of the neurons in V1 to anti-correlated random dot stereograms than to random dot stereograms; (2) the obtained disparities are more accurate than existing neurophysiological methods, and even better than some classical computer vision methods. 3. Inspired by the disparity computation in the visual cortex, a hierarchical model is proposed. The merit of the proposed model lies in three aspects: (1) A normalized disparity energy model being consistent with psychological data is proposed to weaken the influence of stereograms contrast and position on the retinas on disparity energy of neurons; (2) Based on properties o...
修改评论