Recently, face recognition systems(FRS) are increasingly used in policing such as criminals tracking, terrorist arrest, security, and surveillance in smart-city. Because, face recognition is less invasive and does not require a subject to be in proxim...
Recently, face recognition systems(FRS) are increasingly used in policing such as criminals tracking, terrorist arrest, security, and surveillance in smart-city. Because, face recognition is less invasive and does not require a subject to be in proximity to or in contact with a sensor. Also, as many video devices are developed, it is now possible to easily acquire images that can be used for investigation of police.
The general FRS is consists of training data acquisition, pre-processing, recognition model training, and prove image classification. In a real-world setting, facial-data acquisition can occur in many different environments, so a sufficient number of facial images are therefore needed to construct a FRS that is reliable under various conditions. However, it is difficult to collect sufficient images in real-world for each subject. Consequently, it is very difficult to recognize suspects found in CCTV when collecting single sample per person, such as police database of criminals.
In this paper, we proposed the FRS for improving performance of face recognition under Single Sample per Person problem. The proposed FRS has added data augmentation and enhancement steps to the conventional FRS.
First, in data augmentation step, we proposed the method of generating illumination and facial expression variation from a neutral face image by using the Bidirectional Integral Features(BIF) and the Weighted Interpolation Maps(WIM). From the noise model, an image including various variations can be generated through a change in contrast and brightness of the neutral image.
The BIF uses bidirectional integration image to extract each contrast factors for six illumination directions. Because the specific position of the integral image represents the distribution of gray-level intensities in the sub-region up to the corresponding position, it is possible to extract pixel change information by the shadow of the illumination. The WIM indicates the degree of variation in the pixel as it changes from neutral to expressive. The pixel value in a position with a large degree of variation can be replaced by the brightness of a neutral image of the auxiliary set. The WIM generates 6 facial expression images such as smile, surprise, angry, disgust, fear, and sad. As a result of the face recognition experiment on a number of face databases, the proposed method showed higher recognition rate than the existing method. Also, we confirmed that the proposed method for the PF07 database has the highest recognition rate when the images generated by BIF and WIM are used for training.
Second, in data refinement step, we propose the Linearly Projection Vector Estimation(LPVE) to refine noisy face images. The projection vectors of all data converge to the point in the L2PCA feature space when retention rate is 0. Also, the projection vector of the loss data converges linearly from the convergence point according to the degree of retention rate from projection vector of the original data. Thus, original projection vector of loss data is linearly estimated from convergence point. The results of face recognition experiments on AR database shows higher face recognition rate than other methods.