视网膜颜色视觉的计算模型与应用研究

发布时间：2018-06-05 17:18

本文选题：视网膜 + 颜色视觉　；参考：《电子科技大学》2017年博士论文

【摘要】：视觉,是客观世界在视网膜上的投影由视神经系统处理加工后所产生的主观感觉。视觉研究的终极目标是阐明这一主观感觉产生的过程。研究视觉,不仅有利于了解视觉神经系统的工作机制,认识人类自身,也有利于提高图像采集和处理技术,为人工智能机器提供输入。本文从计算模型的角度出发,按照基于生理、面向任务、可计算、系统最简的思路,结合计算机视觉中颜色恒常、有雾图像去雾增强以及高动态范围图像显示等三个具体应用,探讨视网膜完成颜色恒常、图像增强以及亮度动态范围压缩等视觉任务的可能机制。在颜色恒常中,本文引入水平细胞对视锥输出的调节作用和视网膜神经节细胞非经典感受野动态亚区结构,通过模拟水平细胞对视锥细胞的大范围调节,神经节细胞的中心外周颜色拮抗响应,神经节细胞非经典感受野亚区结构去抑制作用,神经节细胞动态适应等视网膜机制中与颜色恒常高度相关的部分,提出一种无需显式估计光源的颜色不变性计算模型。该模型的主要创新在于,通过神经节细胞非经典感受野亚区结构在保持对物体表面响应的同时选择性地抑制光源,通过动态适应机制根据场景特征选择抑制权重保证对不同场景的适应性。该模型与现有颜色恒常算法相比极有竞争力的比较结果说明,水平细胞对视锥信号的调整和神经节细胞动态适应的非经典感受野亚区去抑制作用在生物的颜色恒常能力中发挥了重要作用。在图像去雾增强中,本文部分参考大气散射模型对有雾图像成因的描述,通过模拟双极细胞高斯差感受野、无长突细胞根据视杆输出对视锥双极细胞输出的动态调整、视网膜神经节细胞非经典感受野去抑制作用以及ON/OFF通路整合等视网膜机制中与图像增强有关的部分,提出一种无需对环境进行显式先验假设的有雾图像去雾增强模型。该模型的主要创新在于,利用视网膜不同层次的细胞分别处理有雾图像的不同衰减,为单幅图像去雾增强提供一种简单直接的快速算法。该模型与现有去雾算法极有竞争力的比较结果说明,双极细胞、无长突细胞以及神经节细胞在视觉系统信号增强中有着一定的贡献。在高动态范围图像显示中,本文引入随局部亮度动态变化的水平细胞感受野,通过模拟水平细胞动态感受野和双极细胞高斯差感受野等视网膜机制中与亮度动态范围压缩相关的部分,提出一种在低动态范围输出设备上显示高动态范围图像的色调映射方法。该方法的主要创新在于,通过模仿水平细胞根据亮度动态变化的感受野,有效避免在处理高光区域时出现光晕现象,同时利用双极细胞高斯差感受野滤除场景冗余信息增强局部对比度。与视觉系统在不同自然光照环境下快速获得稳定感知的能力一致,该方法在不同场景中均有着稳定的表现。该方法与现有色调映射算法极有竞争力的比较结果说明,水平细胞对视锥输出的侧抑制和双极细胞带通滤波的高斯差感受野有效地压缩了输入信号的动态范围,有利于视觉系统以有限的资源处理动态范围极大的输入信号。本文通过三个描述不同视网膜机制的计算模型分别探讨了视网膜在完成颜色恒常、图像增强以及亮度动态范围压缩等视觉任务中的作用,整合这些不同生理机制,建立完整描述视网膜功能的统一多任务模型是今后进一步研究工作的方向和重点。
[Abstract]:Vision is the subjective sensation that the projection of the objective world on the retina is processed by the optic nerve system. The ultimate goal of the visual study is to clarify the process of the subjective sensation. From the point of view of the model, three specific applications such as color constancy, fogging enhancement and high dynamic range image display in the computer vision are used to explore the color constancy of the retina, and the color constancy of the retina is discussed. The possible mechanisms of visual tasks such as enhancement and luminance dynamic range compression. In color constancy, this paper introduces the regulation of the horizontal cells to the cone output and the dynamic subregion of the non classical receptive field of retinal ganglion cells. By simulating the wide range regulation of the cones of the horizontal cells and the central peripheral facial features of the ganglion cells. The response of color antagonism, ganglion cells non classical sensing of the structure of the wild sub region, the dynamic adaptation of ganglion cells in the retina mechanism related to the color constancy, and proposed a color invariance calculation model without explicit estimation of the light source. The main innovation of the model lies in the non classical receptive field of ganglion cells. The subregion structure, while maintaining the response to the surface of the object, selectively inhibits the light source and selects the inhibition weight to ensure the adaptability to different scenes by the dynamic adaptation mechanism according to the scene characteristics. The model is very competitive compared with the existing color constancy algorithm, which shows the adjustment of the conical signal and ganglion in the horizontal cell. In the image demogging enhancement, this paper describes the formation of the fog images in the atmospheric scattering model, through the simulation of the Gauss difference receptive field of the bipolar cell, and the non amacrine cell based on the optic pole output to the cone bipolar bipolar disorder. The dynamic adjustment of cell output, the inhibition of the non classical receptive field of retinal ganglion cells and the part of the image enhancement related to the retinal mechanism of ON/OFF pathway integration, propose a fogging enhancement model without the explicit prior assumption of the environment. The main innovation of this model is the use of different retina. The hierarchical cells treat the different attenuation of foggy images and provide a simple and direct algorithm for single image fog enhancement. The comparison results of this model with the existing fogging algorithms show that bipolar cells, no amacrine cells and ganglion cells have a certain contribution to the enhancement of visual system signal. In the state range image display, this paper introduces a horizontal cell receptive field with dynamic changes in local luminance, and presents a high dynamic range image on low dynamic range output devices by simulating the parts related to the dynamic range compression of the luminance dynamic range, such as the dynamic receptive field of horizontal cells and the Gauss differential receptive field of bipolar cells. The main innovation of this method is that by imitating the receptive field of the dynamic change of luminance by the horizontal cells, the halo phenomenon can be avoided in the high light region, and the local contrast is enhanced by using the bipolar cell Gauss difference receptive field to filter the scene redundancy information. The ability of fast obtaining stable perception is consistent, and the method has a stable performance in different scenarios. The comparison results with the existing hue mapping algorithm show that the Gauss differential field of horizontal cells to the cone output side suppression and the bipolar cell band pass filter effectively compresses the dynamic model of the input signal. It is beneficial for the visual system to deal with the dynamic range of input signals with limited resources. In this paper, the functions of retina in visual tasks such as color constancy, image enhancement and luminance dynamic range compression are discussed respectively by three computational models describing different retinal mechanisms, and these different physiological mechanisms are integrated. The unified multitask model describing retinal function is the direction and emphasis of further research.
【学位授予单位】：电子科技大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：R339.14;TP391.41

【参考文献】