A Finite Difference Approximate Fractional-Order Gradient Operator for Improving Image Classification Performance

Abstract

This study presents a finite difference fractional order gradient operator to enhance CoHOG feature extraction for image classification problems. A finite difference Euler approximation of fractional-order derivative (FOD) filter structure is used for the fractional-order gradient calculations for derivative orders around the value of one. The proposed gradient operator is employed for CoHOG feature extraction and performance improvements obtained in image classification are demonstrated over six image sets containing about 15000 images. It is observed that roughly 5% improvement is possible in average correct classification percentage (CCO) at the fractional gradient order of 0.4 when compared to average correct classification percentage of the conventional gradient with the integer order of 1.0. In the case of proper fraction-orders selection particular to dataset, classification accuracy improvements can be obtained up to 14.8% depending on dataset.