Boosting facial expression recognition using LDGP - Local Distinctive Gradient Pattern

Appearance based local feature methods are widely used for facial expression recognition because of their simplicity and high accuracy rates of recognition. However, the achieved accuracy rates and running time yet need to be improved. A new appearance based local feature method, called Local Distinctive Gradient Pattern (LDGP) is proposed in this paper. It derives two 4-bit local binary patterns from two different layers for a pixel by comparing the gray color intensity value of the pixel with its neighboring pixels in four distinct directions. Since each face image is divided into equal sized blocks, two histograms for the two 4-bit LDGP patterns of all pixels in each block can be constructed. The histograms of all blocks are then concatenated to build the feature vector for the given image. To evaluate the effectiveness of the proposed descriptor, experiments were conducted on the popular JAFFE dataset using Support Vector Machine (SVM) as the classifier. Extensive experimental results with seven prototype expressions show that proposed LDGP descriptor is superior to other appearance-based feature descriptors in terms of accuracy rates of recognition.