Desguised Face Detection And Recognition

As banks enhance the self-service business of saving and drawing money, more and more ATMs have been used. Meanwhile, the crimes associated with the ATMs also increased obviously. The criminals always wear sunglasses, masks or caps to cover the faces. So, in order to prevent these crimes happening, it is very important to detect and recognize the disguised faces.In this paper, we study the method for disguised face detection and recognition under the complex background of self-service banks. We propose a from-dynamic-to-static foreground object detection strategy to detect a person. The method consists of three stages. The first stage is the dynamic detection of the strategy. In this stage, we propose the updatable learning-based codebook model to improve the traditional code book model which can not update the background. The second stage is the static detection of the strategy. In the stage, we propose the LBP + HOG feature-based head-shoulder detection. Combining these two features not only can increase the detection accuracy but also can enhance the computational efficiency. We detect moving targets to eliminate the interference of the static background and decrease the impact of the dynamic backgro und, then further determine whether the target is a person. Meanwhile, this strategy only conducts the static detection when the moving target is detected, which can save the computational time, decrease the impact of the background and increase the detection accuracy. The third stage is to detect and recognize the disguised faces by using the disguise-based Adaboost classifier. We use the images of disguised faces as the training positive samples and use the images of normal faces together with the background images as the training negative samples. In this way, we can decrease the error detection and recognize the class of the disguise. In addition, we use PCA with the nearest neighbor classifier to further determine the disguised-like area that detected in the previous step to get higher accuracy.We execute a lot of experiments on the real video database and laboratory video database. The experiments show that our method can achieve good performance on disguised detection rate, accuracy, error rate and mi ssing rate.