Image Classifier Using Modified Binary Support Vector Machines Structures and 2-D Gabor Filters

This paper proposes an image classification system based on modified binary SVM (Support Vector Machines) structures. We design a 40 two-dimensional Gabor filter bank in order to perform convolution between each one of the filters and a given input image. This pre-processing step will result in a 40-feature vector for each image. These feature vectors will then be fed into the system for the training and testing processes. We also use a dataset of images that differ in background, illumination, scale and rotation conditions to test the effectiveness of our methods with real world data. After data standardization, we model a set of binary SVM to perform multi-class classification using the two major approaches found in the literature: One vs One SVM and One vs All SVM. This work provides a comparison between the two classification approaches and shows a classification accuracy rate above 90% for each class of objects and global accuracy rate results above 95% using three different types of SVM kernels.