Research On Face Retrieval Mixed With Multiple Image Feature

Image data is full of people's life with its vivid features. The continuously enriched and growth of image data, strongly demand fast, automatic and effective retrieval to it, image retrieval based on content has been an unprecedented development in this context. In the retrieval process, because face is an important visual information, images that contain the face trend to be more interesting. This article researches aiming at how to find the images that containing a particular character form large number of chaotic images, using the thought that firstly narrow retrieval range then determine the identity of person, the image is divided into two categories—"contains character image" and "do not contain character image", then doing identity authentication (face recognition) to image that contains character. This article has done the following several aspects around the "image classification" and "face recognition" these two key issues:(1) To achieve image classification, firstly analysis the various features description and extraction of the image. As a training sample of images, using the local cumulative histogram method to extract color feature, extraction of texture feature is by means of Gray-level Co-occurrence Matrix(GLCM)method. The feature is expressed as ordered pair of attribute and value, obtained the decision tree of classification images by training, divided the image into two categories.(2) After obtain the character image, using haar-like features to obtain classifier by training based on Adaboost algorithm. It achieves the purpose that the face detection and segmentation. Using the Gabor wavelet and LBP to extract and mixed the face global feature and local feature. Then designing and training BP neural network to achieve face recognition.(3) The decision tree and artificial neural network method all belong to the same areas of machine learning, in order to avoid over-fitting problem and use the training sample efficiently, it needs to make effective assessment to the accuracy of classifier. This text studies the existing assessment methods, design a new assessment method based on it, seeks to consolidate the advantages of existing methods, use the training data more effective, and verify by experiment.This article applied decision tree classification technology to image retrieval, achieving the purpose of narrowing the scope of search, speed up the retrieval speed;and has made better the assessment methods to improve the classification accuracy.