Computational Models And Applications Of The Early Stages Of Biological Visual System

Human visual system has perfect capability for visual information processing and its intelligence outperforms the best current computer vision system. Therefore, building brain-inspired methods to improve computer vision applications has attracted serious concern. This thesis addresses a series of topics related to the computational principles for low- and mid-level vision and explores computational principles of visual information processing especially in the early stages of biological visual system. Based on the visual mechanisms(e.g., receptive field, attention, color processing) in the early vision, we built a series of brain-inspired models for some basic computer vision applications, including edge detection, contour extraction, saliency computation and color constancy.Firstly, we explored the computational models of the color-opponent receptive field from retina to primary visual cortex(V1), and proposed a new framework for color edge detection. This paper explored the functional roles of various color-opponent cells and promoted the understanding of color processing in visual system. We concluded that the proposed method can flexibly capture both structured chromatic and achromatic edges(contours) with additional advantage of quite simple implementation with low computational cost. In addition, we employed the spatial sparseness constraint of neural responses to further suppress the unwanted edges of texture elements and detect the structured contours.In the second part, we built a multifeature-based non-classical receptive field(NCRF)in V1, to improve the performance of perceptually salient contour detection in grey-scale natural images. We proposed a unified framework for NCRF-based models and explored the functional roles of various cues in visual information integration. The experimental results show that combining multiple cues can substantially improve the performance of contour detection compared to the models using single cue. In general, luminance and luminance contrast contribute much more than orientation to the specific task of contour extraction.In the third part, we discussed the topics on visual saliency and visual search. We first defined a visual task of “salient structure detection” to unify the saliency-related tasks like fixation prediction, salient object detection, and detection of other structures of interest from cluttered environments. We also proposed a unified framework inspired by the guided search theory for salient structure detection. We extracted the prior based on the layout of contours and predicted the salient structures with Bayesian inference. Experimental results on six large datasets demonstrate that our system achieves competitive performance for both the tasks of fixation prediction and salient object detection. In addition, our system can also be easily extended to salient edge detection and multi-object search.In the last part of the thesis, we explored the illuminant estimation and color constancy in natural scenes. Inspired by a psychophysical experiment of simulating color constancy, we first proposed a novel hypothesis indicating that most of the natural images include some detectable pixels that are at least approximately grey. We validated our assumption through comprehensive statistical evaluation on diverse collection of datasets,and then put forward a novel grey pixel detection method. The proposed method can be easily extended to multi-illuminant scenes and image sequences. Experimental results on six benchmark datasets show that the proposed method achieves the state-of-the-art performance for illuminant estimation with the inherent merit of low computational cost.