【作者】 杨娴；

【导师】 李勃； 陈启美；

【作者基本信息】 南京大学 ， 通信与信息系统， 2014， 硕士

【摘要】 在工业视觉领域,瓷砖的智能分色以色彩、纹理感知为主,是典型的模拟人眼视觉系统(HVS)的应用案例。但瓷砖花色繁多、差别细微,精准有效的图像预处理是后续特征提取、分色判断的必要前提。论文基于这一应用场景给出基于HVS的智能瓷砖分色系统图像预处理算法,其研究成果已实际应用于视觉传感瓷砖分色产品,在实验室承担的《视觉传感瓷砖分拣系统》、《基于视觉传感的高速公路灾害天气及应急预警决策辅助综合系统》等多项国家级、部省级项目中多有体现。论文分析了预处理在瓷砖智能分色系统中的地位,结合国内外视觉传感预处理技术的研究现状,解析了预处理关键算法存在的问题,改进了多种相关模块算法方案,主要包括图像亮度空间不均匀性校正,基于色彩恒常特征的色光校正、工业相机的色彩校正、图像纹理的视觉仿真、基于掩膜效应的自适应色差阈值的计算等,并于瓷砖产线上验证。在亮度预处理方面,光源、镜头、相机亮度响应各异,图像亮度分布不均。传统的查表法需逐点测量光场亮度,工作量大,论文结合彩色对数图像处理框架CSLIP,利用灰板计算亮度分布的透射率,校正了亮度分布的空间不均匀性,为后续分色提供亮度一致的输入图像。在图像色彩预处理方面,为了提高彩色信息的输出质量,克服由于光照和相机带来的色偏问题,论文模拟人眼的色彩恒常性,给出了基于彩色对数图像处理的色光校正方法,将图像色彩还原到标准光源下；在相机色彩校正方面,论文采用回归校正方法结合自动亮度分级策略来解决相机的色偏问题,提高成像的色彩精确度。在图像纹理预处理方面,论文模拟人眼的色彩同化、对立现象,还原人眼对纹理的处理效果；在提取图像纹理信息时,结合局部背景亮度及空间频率对人眼视觉的影响,改进了色差可视阈值的计算方法,提高了后续的图像高级特征的提取和分析的准确度。经实验验证和实际应用,使用预处理算法后图像特征的检测率显著提升,具有重要的理论意义和应用价值。其创新点或特点在于：● 改进了多参数的亮度不均匀校正算法,将CSLIP应用于亮度校正中,还原亮度均匀的图像；● 改进了色光校正算法,克服对角模型中色彩通道间的相关性,将图像颜色校正到标准白光下；●改进了基于多项式回归的色彩校正算法,使用自动亮度分级校正策略,解决色彩校正中的过曝或过暗问题；●提出了自适应色差阂值的计算方法,结合人眼的掩膜效应,有效降低图像噪声对特征选择的影响。更多还原

【Abstract】 In the field of machine vision applications, intelligent color separation of ceramic tiles, a highly subjective process, focuses attention on color and texture perception. It is a typical application case that simulates human brain recognition and analysis. Image preprocessing is the necessary prerequisite for feature extraction and color separation. With regard to the above application case, the thesis proposes a series of image pre-processing algorithms based on human color perception in color separation system of ceramic tiles. The research findings are highly reflected in several national and provincial projects, such as "Ceramic Tile Sorting System by Visual Sensing" and "Highway Disaster Weather and Emergency Forewarn Decision Supplementary Synthetic System Based on Visual Sensing".Based on analysis of recent research progress at home and abroad, the meaning and role of preprocessing in vision sensing systems are described. For a variety of practical applications, difficulties and encountered problems in key modules and algorithms are analyzed.Then corresponding algorithms or optimized solutions are proposed. Study on key preprocessing algorithms includes brightness correction, color light correction based on color constancy, color correction for industrial camera and texture simulates based on HVS. These preprocessing algorithms have been experimentally verified and applied in multiple application prospects such as vision sensing based ceramic tiles sorting, express ways and abnormal event detection.In terms of brightness pre-processing, to correct the vignetting effect caused by devices or light source, the thesis proposes a correction method in brightness space based on CSLIP logarithm computing framework. The proposed method transfers calculation of image pixel values to that of transmittance, then recovers images whose brightness is uniform.In terms of image color pre-processing, to improve the output quality of color information and overcome the color cast problem caused by illumination or cameras, the thesis proposes a colored light correction method based on logarithmic color image processing framework to recover image color to the standard D65 condition. Moreover, the regression-calibration method is used to improve color accuracy of digital cameras, Combined with automatic brightness classification strategy.In terms of image texture pre-processing, the thesis simulates color assimilation and contrast phenomena of human eye, recover the processing effect of human eye on texture. The thesis puts forward a local adaptive perceptual color difference algorithm during the extracting process of image texture information and then combines the influence of local changes in luminance and spatial frequency to human visual system. In a word, the proposed method lays the foundation for advanced image feature extraction and analysis in the next stage.By experimental verification and practical application, the research results of preprocessing algorithms are credible, and have important theoretical significance and application value. The main contributions of this dissertation are summarized as follows:·A multi-parameter uneven brightness correction algorithm is improved. CSLIP is applied into brightness correction so that images with even brightness are recovered.·A colored light correction algorithm is improved. The correlations between the color channels in the diagonal model are overcome, so as to correct the image color to standard white light.·A color correction algorithm based on polynomial regression is improved. An automatic luminance gradation correction strategy is used to solve the overexposure or dark problem in color correction.·Combined with the mask effect of human eye, a calculation method of color difference threshold is improved. As a result, the influence of image noise on feature selection is effectively reduced.更多还原