Integral Image based Fast Convolution Operations via Separable Kernels

In this paper a method for accelerating the convolution processes, which is performed with different size of kernel filters, is proposed. The proposed method is basically based on the integral image representation approach. Integral image approaches are not directly usable with kernel filters. The convolution process is performed over sub-matrices by defining kernel filters as linear combinations of sub-matrices. Comparing the running times between standard convolution and the proposed method under different sizes of kernel filters and different resolutions of images, it has been observed that the convolution processing times for 19 × 19 and larger kernel matrices are in favor of the proposed method.