Research And Implementation Of Face Recognition System Based On Neural Network In Complex Background

Video surveillance market is now at a stage of vigorous development, all kinds ofadvanced HD monitoring equipment will also be popularized gradually in China. But wecould found that the monitoring still relies on backward artifical monitoring mode. Themarket demand for intelligent video surveillance system with face recognition is now moreurgent with the development of Safety City and Smart City construction. Aiming at thedemand of the market, we adopted the Vialo-Jones method to detect human face rapidly,adopted the PCA technology to extract face feature, classifies and calibrates face images byself organizing map neural network, and finally we implemented one real-time video facerecognition system based on Qt and OpenCV on Windows/Linux platform. The system hasthe advantages of small size, graphical user interface, good portability and easy to use. Thesystem has the functions such as batch cutting face, performing PCA, neural network training,verifying neural network, video face recognition and image face recognition. After testing, thesystem achieved good recognition effect in complex background, and is worth popularizing.The core idea of this paper is to utilize the neural network’s self-organizing ability toconstruct a two-dimensional matrix, which classifies faces in organized order. After extractingthe features of face image through PCA, system projects the feature vector to an element ofthe two-dimensional matrix, and then we could use the coordinate of the element to identifythe face image. So the face recognition could be done by determine if the euclidean distancebetween two elements is less than a threshold.The innovative works of this paper are: We adopted self organizing map neural work inface recognition and achieved good results. We designed a high performance self organizingmap neural network based on matrix operation, simplified the program flow and improved theperformance of the real-time video processing system. We improved the structure of theneural network by building a stereo annular grid in logic, so that each neuron will be activatedfairly in training, the result of training will be distributed more evenly in the neural network,and that will increased the utilization rate of neural network. To improve the recognition rate,we designed the algorithm of real time face correction, rotate the face image according to theangle between a horizontal line and the line of eyes center. We used a mask image to coverparts of face image outside eyes, nose and mouth, to eliminate the influence of differentillumination, angle and covering. And we attached importance to the visual of testing andverifying, impreoved the system by researching the “weight-face” matrix and so on.