【作者】 张悦；

【导师】 方涛；

【作者基本信息】 上海交通大学 ， 控制科学与工程， 2014， 硕士

【摘要】 颜色作为一种非常重要的视觉特征，广泛应用于计算机视觉，尤其是图像处理领域。然而目前很多颜色特征是从计算机视觉角度出发提出的，难以像人类视觉系统那样同时表达出我们所需要的不同颜色属性信息，比如色调、饱和度和颜色恒常性等。有些学者尝试从生物视觉的角度，模拟哺乳动物感知颜色的过程，进行颜色特征的提取。由于视觉皮层中颜色通路各区域处理颜色的机制比较复杂且很多神经机制尚不统一，大部分颜色特征的研究只是集中在视觉通路的低层次阶段，沿整个颜色通路研究颜色信息的编码方法很少。在总结颜色神经机制的基础上，开展颜色神经编码研究，这对于提取符合生物视觉处理的、稳健的颜色特征具有重要理论与应用价值。本文主要从生物视觉的角度出发，通过研究每个区域颜色感知的神经机制，以及各区域之间的信息传递及转换过程，建立了颜色的层次性编码模型。该模型较好地模拟了自然光经眼睛进入视网膜，到侧膝体，再经过V1，V2，PIT及IT阶段的处理过程，最终形成了从低层到高层的颜色层次性特征表达。由此模型提取的颜色特征具有颜色恒常性、稳健性等特点。为了评价提出的颜色的层次性编码模型性能，首先利用颜色恒常性的图像集进行实验，验证了模型提取的特征比普通的颜色特征有更好颜色恒常性。同时，利用USGS遥感数据集与Caltech-101数据集，提取颜色特征并进行分类，与传统的颜色特征进行了对比，发现本文方法能够在分类中取得更好效果。最后用到两种颜色相关的数据集进行了实验分析，验证了本文方法对颜色辨别的有效性。更多还原

【Abstract】 Color is a very important visual feature which is widely used incomputer vision especially image processing. However, most color featuresare extracted from the perspective of computer vision, they are difficult tocharacterize different color attributes simultaneously as the human visualsystem does, such as hue, saturation and color constancy and so on. Somescholars attempt to extract color features by simulating the process ofmammals peceiving color. Since the complexity and disunion of colorvision’s mechanisms in the cortex, most studies only focus on the colorcharacteristics of the low-level stage of visual pathway, rarely take intoaccount the entire color channel. On the basis of summarizing color neuralmechanisms, we carry out research of the neural coding of color which isvery important and valuable for the extracting biological and stable colorfeatures.This article proceeds from the perspective of biological vision. Bystudying the neural mechanisms of color perception for each cortical region,as well as information transfer and conversion process from a cortical regionto the next neighboring one, we establish a hierarchical model of colorcoding (called HMCC). The model well simulates the neural processing of natural light entering through the retina into the eye, the lateral geniculatebody, and then to V1, V2, PIT and IT stage, and eventually forms colorfeature expression from low-level to higher-level vision. The color featureextracted by the HMCC model is with color constancy, stability and othercharacteristics.In order to evaluate the performance of proposed HMCC, firstly we doexperiments using color constancy image sets and verify that the feature hasbetter color constancy than ordinary color features.Meanwhile, The colorfeature extracted by the HMCC model is used to classify the USGS remotesensing data sets and Caltech-101data sets, these experimental results showthat the proposed method can achieve better results than these traditionalcolor features. Finally, by analyzing experiment of two color-related data setswe verify the effectiveness of color discrimination of the proposed method.更多还原