| Summary: | Face recognition across pose is considered as a critical problem in image processing. Occlusion which is caused by pose leads to a lack of information to template matching. Furthermore, templates should be provided with the ability to restore the prior knowledge of facial shape and appearance. It is useful to integrate template matching with skin-based Models, but skin is unsteady through different aspects. Moreover, utilizing depth maps in 3D models deal effectively with pose issue. However, it requires a good estimation of ethnicity. concluding 3D shape closest to reality, and customizing the depth map. The research methodology is divided into four phases. In the first phase. there are 1721 2D images acquired from PICS, Multi-PIE, CAS-PEAL. Face-Place. and USF Human-ID. In addition, prior facial information is defined including shape, landmarks, appearance, polygonal mesh, and 3D depth information. In the secomd ~c. the methods and models used in implementation are proposed. The implementation includes an explicit skin-based model which analyzes the skin across several color spaces, a skin-based segmentation to extract face and feature, composite cross-correlation to enchase template matching. Rcstricted Shape and Abbreviated Appearance Correlation Model (RSAACM) to detect facial landmarks, Adaptive Neuro-Fuzzy Model (ANFM) for estimating ethnicity, and 3DÂethnical technique to synthesize 3D models based on 2D face images. In the third phase, all the proposed models are analyzed in order to adopt the best structures and implementation. In the fourth phase, the models are evaluated and compared with previous works to ensure their efficiency. The used processes of evaluation and analysis include Active Energy (AE), Algorithm Order (AO), composite crossÂcorrelation, Root Mean Square Error (RMSE), cosine distance, and elapsed time. The analytical and experimental results indicate that (Cb and Cr) of YCbCr and (H) of HSV are the most appropriate for explicit skin models compared to XYZ, RGB, CMYK, and HSV. The skin-based segmentation model reaches 98.69% for precision and 95.68% for accuracy. The RMSE of detecting features is reduced by composite templates compared to standard templates. The proposed RSAACM is compared with recently published models. The comparison includes Boosted Regression and Graph Model, Optimized Part Mixtures and Cascaded Deformable Shape Model, Graphical Model. and Fast Fitting Appearance Model. The comparative results showed that RSAACM was more efficient when it comes to detecting landmarks, where its RMSE is decreased to I 0.33. Thus, the measured cosine distance for 3D-Ethnic return is decreased by 37% compared to Generic Elastic Model (GEM). Evaluation of the proposed models on gold standard data demonstrates the signi ficant accuracy of face recognition across pose using the RSAACM and 3D-ethnical technique.
|
|---|