A Study Of Local Ternary Pattern In Illumination Face Recognition

In recent years, face feature extraction and classification techniques, as a common biological feature extraction technology, is one of the key research direction in pattern recognition and computer vision field. Feature extraction is the key of the classification process. A good feature makes different individuals between classes more discriminative. This paper has the following works.1) Local Ternary whole Descriptor(LTwCD) is proposed in this paper. Concentrating on background of illumination problem and the basic theory of Local pattern, including: LBP, Local Ternary pattern and Local Ternary Contrary Descriptor, we presents a Local Ternary whole Descriptor(LTwCD), the algorithm takes into account not only the symbolic information and amplitude information in the field between the pixel and the center pixel, and the contrast information. The algorithm has coding invariance with partial illumination changes. It’s able to extract illumination invariant feature from face images affected by the illumination and has a better ability to identify.2) This paper discusses the common light processing algorithms, including pixel-based image processing and face modeling methods. Base on the Lambert model, this paper analyzes the advantages and disadvantages of light process algorithm based on the illumination model, For example, Holomorphic filtering, Self-quotient image, Gradient Image, Relative Gradient (RG) for extracting illumination invariant characteristics.3) We propose a novel LTwCD-based face recognition approach, namely Relative Gradient Local Ternary whole Contrast Descriptor (RGLTwCD), in which the relative gradient is first applied to the original face images to extract illumination invariant image. Then, this image be encoded by LTwCD to obtain two binary-image. We use histogram to calculate their gray value information and cascade them. This cascading histogram is acted as image’s feature vector. Finally, the Chi square dissimilarity measure and the nearest neighboring classifier are used for classification.Algorithm proposed has been proved of coding invariant under partial illumination linear change not only in theory, but also experiment on Yale face database, Yale B face database, Extended Yale B face database, CMUPIE face database. It can eliminate the influence of a strong light on face image and extract the image illumination invariant texture.