Аlgorithm of fast computation of local image histograms on video card1

Abstract

An algorithm of parallel computation of image histograms of different types, including brightness and oriented gradient ones, on video cards of various types is presented. Now local histograms are used for solution of some tasks of image processing and recognition, but their application is restricted due to the long computational time. One of the difficulties appearing during parallel computations of this vector feature is the large number of conflicts of simultaneous access to video memory sells. In the developed version, the number of conflicts of simultaneous access are many times reduced. It accelerated the computations. For instance, 9D vectors of histograms of oriented gradient for all 256×256 windows of a HD image are calculated on the GPU NVIDIA GeForce GTX 1060 within 1,9 msec, whereas the same computations made by the CPU Intel Core i7-6700 take 151 msec.

Keywords

• images
• local histograms
• algorithm
• parallel version
• cuda