【作者】 胥斌；

【导师】 唐远炎；

【作者基本信息】 重庆大学 ， 计算机应用技术， 2013， 博士

【摘要】 特征提取是模式识别和计算机视觉研究的核心技术，数十年来受到学者们的广泛关注。现实环境中图像存在平移、旋转、尺度、光照及视点等多种变换，增强图像特征对这些变换的不变性，提高图像分类和识别的性能是一个日益重要的课题。本文对图像的不变特征提取进行了深入、系统的研究，在总结传统的不变特征提取方法的基础上，重点探讨了一种新的图像不变特征表达——散射变换，并将其应用到人脸识别和关键点检测中。针对遮挡人脸识别，本文将散射变换引入图像梯度域，提出了一种新的基于图像散射表达的抗遮挡人脸识别方案；利用二阶散射系数的特性，本文提出基于二阶散射能量的图像关键点检测方案，实验验证了方案的有效性；针对人脸识别中的变光照问题，结合小波多尺度特性，本文提出了一种新的多尺度光照不变量。本文的主要工作和创新点包括：1．提出了一种新的基于图像散射表达的抗遮挡人脸识别方案——梯度散射脸。本文以“Recognition by the Rest”为思想，将散射变换引入图像梯度域，一方面极大地抑制了遮挡块对原图像频率结构的破坏，另一方面，得益于散射表达的局部平移不变性和弹性形变稳定性，实现了对脸部未遮挡信息的充分提取。传统的抗遮挡人脸识别方法总是试图“重建”被遮挡脸部，这些方法以“局部特征识别”和“稀疏表达分类”为代表：“局部特征识别”用脸部显著性特征的加权来表征人脸，这种方法不仅需要更多的训练样本，而且基图像中不属于显著性的人脸部分也会产生不为零的权值；“稀疏表达分类”将遮挡人脸看作无遮挡部分和遮挡块的和，引入遮挡字典表示遮挡块，而无遮挡部分应仅由训练图像稀疏表示，重构时抛弃在遮挡字典上的投影系数，仅由对应于训练图像的稀疏系数得到重构图像。“稀疏表达分类”存在两个问题：首先，实际图像中的遮挡往往是非线性的，抛掉遮挡字典系数并不能完全消除遮挡的影响；其次，实际中训练图像是有限的，这些有限的训练图像的稀疏表达不足以囊括人脸图像中出现的其它形变。实验结果显示，提出的梯度散射脸优于上述两种主流方法，对遮挡人脸取得了相当高的正确识别率，且梯度散射脸不需要进行训练，这对许多实际应用具有重要意义。2．提出基于二阶散射能量的图像关键点检测新方案——SCD。现在流行的关键点不变特征一般都是在多尺度情况下检测得到，如Harris-Laplace算子和SIFT,而所有在多尺度下检测局部不变特征点的方法都面临一个共同的问题：图像中的局部结构往往存在于一个尺度范围之内，而不是某个固定的尺度，在多尺度下检测特征点会得到许多位置和尺度非常接近的点，它们都代表同一个局部结构，冗余特征点会造成后续错误匹配的大量增加。二阶散射能量是尺度间结构相似程度的一种反映，由二阶散射能量图局部极值点对应的关键点是一种尺度间局部特征，其稳定性和关键程度相比单一尺度上的关键点都更高。实验结果显示，本文提出的SCD方法得到的关键点数量相对更少，具有较强的结构性，对图像视点变化、尺度变换和景深变化及一定程度的非刚性变换具有良好的鲁棒性，取得了较高的可重复率。3．提出了一种新的多尺度光照不变量——MGF。光照问题是人脸识别领域面临的难题之一，由于光照变化使得面部特征不再明显具有分辨能力，这样造成了人脸的类内散度增大，甚至大于人脸样本的类间散度。研究表明，同一个人在不同光照条件下得到的图像之间的差异，可能比不同人在同一光照条件下得到的图像之间的差异还要大。MGF方法对图像的一阶高斯差分进行小波变换，提取其方向特征作为光照不变量，理论分析证实了MGF是一个光照不变量，且保留了更多的图像结构信息。另外，MGF提供了一种用不同特性小波构造光照不变量的途径，可以构造更为丰富的光照不变量。实验结果表明，MGF能有效地提高人脸识别对光照的鲁棒性。更多还原

【Abstract】 Feature extraction is the core technology of pattern recognition and computervision, which is widely concerned by scholars for decades. Since the existing of allkinds of image transformation including translation, rotation, scale, illumination,viewpoint and so on in reality environment, it becomes increasingly important toenhance features invariance to improve the performance of classification andidentificationBased on summarizing the traditional methods, in this dissertation, we study on theimage invariant feature extraction deeply and systematically, focus on a new imageinvariant representation called Scattering transform (SCATT), and apply it to facerecognition and key-point detection. For occlusion face recognition, we propose a newscheme by introducing the scattering transform into the image gradient domain; Takingthe advantage of the characteristics from the second-order scattering coefficients, wepropose a scheme for image key-point detection and, the relating experiments verify theeffectiveness of the scheme; With regard to face recognition under varying lighting, wepresent a new multi-scale illumination invariance using wavelet. In this dissertation, themain work and contributions are as follows:Firstly, a new occlusion robust face recognition scheme called Gradientscatt isproposed based on SCATT. By the idea ‘Recognition by the Rest’, SCATT is appliedinto the image gradient domain. On the one hand, the negative impact on the frequencystructure of the original image due to the occlusion block are greatly suppressed,however, Gradientscatt can fully extracted the information from the parts uncoveredthanks to SCATT’s local translation invariant and elastic deformation stability. It isdifferent from the two kinds of traditional methods--‘Local feature recognition’ and‘Sparse representation classification’, which are always trying to rebuild the occludedparts.‘Local feature recognition’ using the weighted sum of the salient features in theface to characterize the face, which not only needs more training samples and non-zerovalue belonging to non-significant characteristics are produced.‘Sparse representationclassification’ consider the occluded face as the sum of the unobstructed part andocclusion block, which introduce the occlusion dictionary presenting the occlusionblock while unobstructed part should only be represented sparsely by the trainingimages. Face was reconstructed only by the sparse coefficients corresponding to the training image and the coefficients projected on the occlusion dictionary are discarded.There were two problems: Firstly, the occlusion in reality is often non-linear so as to itis impossible to remove the negative impact completely due to occlusion block bythrowing away the occlusion dictionary coefficients; Secondly, training images islimited and sparse representation of limited training image is not sufficient to embracedeformations existing in face. Experimental results show that Gradientscatt is superiorto the above-mentioned two main methods, which obtains a very high recognition ratefor occlusion face recognition.Secondly, we propose an image key-point detector called SCD based on thesecond-order scattering energy. Nowadays popular key-point features are generallydetected in the multi-scale space, such as Harris-Laplace operator and SIFT. However,all of the multi-scale methods to detect key-point face a common problem: the localstructures in image often exist in a scale range instead of a fixed scale. Detecting thekey-point in the multi-scale space will inevitably get many points with similar positionand scale and they all represent the same local structure, undoubtly redundantkey-points would result in a significant increase of mismatch. The second-orderscattering energy is a reflection of structural similarity between the scales. The localextreme points of the second-order scattering energy graph are corresponding tokey-points, which are inter-scale local features, so the stability and criticality of thekey-points is higher than that corresponding one at a single scale only. Experimentalresults show that the number of the key-points got by the SCD method is relatively lessthan Harris and SIFT, and which are more structural and more robust to imageviewpoint changes, scale transformation and projection change.Thirdly, a novel method to extract illumination invariant features is proposed forface recognition under varying lighting conditions called MGF. Varying Lighting is oneof the key problems for face recognition. Due to the illumination change, the facialfeatures are no longer to be distinguished clearly, which cause bigger intra-classdivergence, even greater than the inter-class divergence of face samples. The relatingresearches show that the difference between the images from the same person underdifferent light conditions may be even larger than that from the different people underthe same light conditions. The MGF applies wavelet transform on Gaussian differentialof image to extract the direction as invariants. Theoretical analysis confirms the MGF isan illumination insensitive measure, and which keeps more structure information. MGFprovides an approach to construct more illumination invariant measure using wavelet with more different characteristics. Experimental results indicate that MGF caneffectively improve the robustness to varying lighting for face recognition.更多还原