Human Detection And Face Recognition Based On Computer Vision

Human being perceives and understands the surrounding three-dimension world bymeans of visual, auditory and tactile sensation, etc. Human being can only use visual tounderstand two-dimension image quite accurately. For example, a man can easily telldifferences on the shadow, texture, block between objects, and lightness or darkness in animage. However, no matter how exciting achievements people have made in computer visionin the past two decades, it is still a dream that computer can observe and understand the worldas two-year-old children does. Nowadays, there are still many difficulties and challenges inthe field of computer vision. Image classification and identification is one of the mostimportant issues needed to be researched in computer vision. With the development ofcomputer network, computer image processing capabilities and multimedia resources, imageclassification and identification technology in large number of images, especially humandetection and face recognition technology, are widely concerned by domestic scholars, foreignscholars and research institutions. Human detection and face recognition can be used in citysecurity, driver assistance, advanced user interface, video and image retrieval in Web and et alwhich are closely related to our daily life.In this paper, we use technologies of computer vision, pattern recognition, artificialintelligence and image processing, etc. We research the significance, background, difficultyand hot issues, as well as the main methods in human detection and face recognition in detail.Several new algorithms on human detection and face recognition are proposed in this paper,which got excellent results in main human databases and face databases. Summary andprospect are given at the end of this paper.In the first chapter, we analysis in detail the background and significance of humandetection and face recognition, and show their applications in eight areas of intelligencesurveillance, advanced user interface, driver assistance, biological feature recognition and etal. Then, we elaborate difficulties in human detection and face recognition, which are causedby block between objects, illumination changes, posture or attitude changes, texture changes,camera angle changes and so on. After that, domestic and foreign research institutions,research projects, academic journals and conferences on human detection and face recognitionare introduced. And then, we show the development of human detection and face recognitionin the future. Finally, we summarize the main work, innovation and the structure of this paper.In the second chapter of this paper, we begin with description of image classification andidentification methods. In image classification methods, bag-of-word methods where samplesare supervised in training and part-based methods where samples are semi-supervised in training are introduced. SVM (Support Vector Machine) method is elaborated in detail, whichwill be used in this paper. Then common methods in human detection and face recognition aredescribed in detail. At the end of this chapter, we give a summary.There are shortcomings of fixed-location, fixed-scale and fixed-number in featureregions extracted in human detection, therefore in the third chapter, we use theHog(Histograms of Oriented Gradients) features description and propose fast human detectionbased on united Hogs and cascade classifiers. In this chapter we begin with analyzing theshortcomings of these traditional human detection methods, and show our new algorithm. Theconcepts of regular rectangle, irregular rectangle and their VVR (Vertex-VectorRepresentation) are proposed. After that, intersection detection algorithm between regularrectangles and irregular rectangles, VVR generation algorithm for irregular rectangles and fastscanning algorithm for inner points of irregular rectangles are given. Compared with humandetection based on multi-scale Haar and human detection based on multi-scale Hog, we showcompleted experiment results and analysis. A summary is given at the end of this chapter.Face regions extracted in traditional face recognition are fixed-location, which results inthat eigenvalues in these regions are susceptible to non-significant face regions, such asforehead and cheek. Therefore in the forth chapter, we propose weighted multi-channel Gaborface recognition based on region selection and FFT (Fast Fourier Transform) pretreatment.Firstly, we introduce face recognition based on the traditional multi-channel Gabor features,and propose the concept of relative entropy in the ithGabor channel of certain face region.Secondly, we show the necessity of face region selection by experiments and proposeadaptive multi-channel Gabor face region selection algorithm based on relative entropy. Thenwe present the new algorithm. Thirdly, compared with single-channel Gabor facerepresentation, multi-channel Gabor face representation, ensemble Gabor face representation,weighted multi-channel Gabor face representation, weighted multi-channel Gabor facerepresentation based on FFT pretreatment and so on, we show experiment results of the newalgorithm respectively in time, space, detection rate, false positive rate and common types ofimage noise. We give a summary at the end of this chapter.In the last chapter, we show what has been done and what will be done in this paper.In summary of this paper, we introduce the background, significance, difficult and hotissues, domestic and foreign research status and trends of human detection and facerecognition. We elaborate in detail main algorithms and experiment methods on imageclassification and recognition, especially on human detection and face recognition. Fasthuman detection based on united Hogs and cascade classifiers, as well as weightedmulti-channel Gabor face recognition based on region selection and FFT pretreatment areproposed. We describe their procedure and algorithms derived from them, such as fastscanning algorithm for inner points of irregular shapes and adaptive multi-channel Gabor faceregion selection algorithm based on weight of relative entropy. Each algorithm is comparedwith current main algorithms on human detection and face recognition, and completed experiment results and analysis are given. What has been done in this paper not only broadensresearch areas in human detection and face recognition, expends their research ideas, andcontributes certain meanings and values to human detection and face recognition, but also canbe used in city security, intelligence surveillance, driver assistance, biological featurerecognition and et al.