A Research On Face Detection And Recognition Algorithms

Face recognition technology is widely used in identification, authentication and othersecurity areas. In natural scenes, input image contains various confounding factors, mainlyabout the uneven illumination, complex expressions, gestures and facial mask changes andother issues, these interfering factors make sharp decline in the rate of recognition,in thispaper,through the three main steps in face recognition system: image preprocessing, facedetection and recognition, namely the elimination of various interference factors, increaserecognition rate of natural scenes.1. Image pre-processing operations mainly research light compensation and imagefiltering, light compensation can eliminate the interference caused by uneven illumination,comparative analysis of the reference white linear correction, non-linear correction methodand the value of the correction method. Degree from the filtering effect and retain the originalimage detail, comparative analysis of the effect of the mean filter and median filter.2. Using AdaBoost algorithm to detect face image, a cascade of weak classifiers which isobtained by readily available make up strong classifier,realize the face detection. In thedetected face region, using AdaBoost algorithm to locate the human eye once more, justreplace face positive and negative samples into eye positive and negative samples in thealgorithm. Deal the image with geometric correction according to eyes position in the facearea.3. In the face recognition algorithm, the modules dimensional principal componentanalysis (Two-dimensional Principal Component Analysis) first block face image, extractingfeatures of each image sub-block by2DPCA algorithm to identify the distance criterionfunction. This algorithm does not consider the differences between the different sub-blocks ofan image, therefore, the paper by calculating the entropy values of the training sample imagesof different image sub-block is determined weights, increasing the right image sub-blockcontains more than a value corresponding to the information. And calculate test sampleentropy, and entropy of the image sub-block at the same position on the training samples iscompared, may be excluded entropy large interference image sub-block (block face). Inaddition, the information is integrated into the algorithm into test samples to improve theadaptability and robustness of the algorithm. Because2DPCA algorithm does not consider the class information of training samples, while MSD algorithm makes full use of the categoryinformation of training samples, the improved algorithm and2DPCA MSD algorithmcombined. On the ORL database, AR face database and self experiments were conducted facedatabase to verify the improved algorithm of this paper is effective and robust.