Real-time implementation and performance optimization of local derivative pattern algorithm on GPUS

Abstract

Pattern based texture descriptors are widely used in Content Based Image Retrieval (CBIR) for efficient retrieval of matching images. Local Derivative Pattern (LDP), a higher order local pattern operator, originally proposed for face recognition, encodes the distinctive spatial relationships contained in a local region of an image as the feature vector. LDP efficiently extracts finer details and provides efficient retrieval however, it was proposed for images of limited resolution. Over the period of time the development in the digital image sensors had paid way for capturing images at a very high resolution. LDP algorithm though very efficient in content-based image retrieval did not scale well when capturing features from such high-resolution images as it becomes computationally very expensive. This paper proposes how to efficiently extract parallelism from the LDP algorithm and strategies for optimally implementing it by exploiting some inherent General-Purpose Graphics Processing Unit (GPGPU) characteristics. By optimally configuring the GPGPU kernels, image retrieval was performed at a much faster rate. The LDP algorithm was ported on to Compute Unified Device Architecture (CUDA) supported GPGPU and a maximum speed up of around 240x was achieved as compared to its sequential counterpart.