Face Identification System in Transform Domains Over Secure Communication Channel

In a Face Recognition (FR) system, the facial pose image is utilized as the information based upon which a person can be identified. Enrollment and recognition are the two modes in FR system where poses are processed after pre-processing step. In FR, it is accustomed to employ the same pre-processing step(s) in both enrollment and recognition modes. In addition, it is crucial to have an accurate, fast, and less resources demanding system. In this paper, a new pre-processing approach is presented in which the training and testing face poses have different dimensions. The proposed technique aims at improving the overall recognition system performance by increasing both accuracy and final decision revealing time and decreasing model storage space requirements. The presented work is exhaustively tested on a FR system. Two Dimensional Discrete Haar Transform (2D DHT), and Cosine (2D DCT) are employed in an adaptive algorithm to calculate the optimal weights of each coefficient in each domain by diminishing residual portion of the input pose signal. After utilizing pre-processing step (that includes image crop and resize), the adaptive algorithm is utilized to calculate the weights in both DWT and DCT domains. Only a predefined number of dominant DWT coefficients, as observed by the corresponding higher weight of that coefficient, are retained. The classifier employed in this work is based on discriminative sparse representation technique for FR with employment of \(\ell_2\) regularization. As the results show, the new technique met three-parameter model criteria (storage, computational complexity, and accuracy rate) and achieved an average of 95.56% for ORL dataset with five training poses. The candidate application of the proposed technique is in communications systems utilized by FR systems.