High Throughput Variable Size Non-square Gabor Engine with Feature Pooling Based on GPU

Abstract

Increasing application of Gabor feature space in various computer vision tasks and its high computational demand, encourages using parallel computing technologies. In this work we have designed a high throughput GPU based Gabor kernel that mimics the function of initial biological visual cortex layers namely ‘Simple’ and ‘Complex’ cells. The kernel is basically a Gabor filter bank with adjustable number of orientations and scales, supporting ‘Non-Square’ and ‘Variable Size’ filter masks on different channels. Consequently our GPU based Gabor kernel tends to be adjustably more accurate, more flexible for different applications, with optimum computational cost at lower resources. The second important task of our high throughput engine is ‘Gabor Feature Pooling’ with Max and Histogram methods, similar to biological visual ‘Complex cells’. This part of our ‘Gabor Engine’ makes it very practical for computer vision applications, since in addition to massive Gabor features, it also provides more abstract spatial invariant orientational information based on image Gabor features. We have optimised the Engine design to take maximum advantage of all GPU parallel resources and maximum bandwidths of all memories.