Computationally Efficient Content Retrieval for Large Databases of Complex Images

Abstract

newline significantly. This has led to an exigent demand for highly efficient image retrieval systems to satisfy human needs. Content Based Image Retrieval (CBIR) helps to organize the digital image archives by their visual content and finds application in many important fields like medical diagnosis, fingerprint recognition, crime prevention, digital libraries etc. However, there are several challenges researchers face when developing efficient CBIR systems and these challenges need to be addressed carefully to attain maximum performance. Some of the challenges faced by the CBIR systems are the increased computational complexity of the algorithms, difficulty to search for an image in large databases, and the semantic gap problem. Further, finding an efficient image descriptor and the efficacy of the similarity measure used also pose challenges to the CBIR system designed. newlineThe first phase of our research aims at addressing the most important challenge, the increasing computational complexity of the CBIR algorithms. Over the period of time due to the extensive efforts of the researchers to increase the efficiency of the CBIR algorithms the size of the image feature vectors extracted have become tremendously large. This has led to an increase in the execution time of the algorithms making them computationally very expensive. We propose a framework for parallelizing the CBIR algorithms using many core processors so that the computational cost can be greatly reduced. In this phase of the research, efficient techniques for extracting parallelism from a pattern based algorithm, Local Derivative Pattern (LDP), and the strategies for implementing it on General Purpose Graphic Processing Units (GPGPUs) is presented. LDP algorithm though very efficient in retrieval does not scale well when capturing features from high resolution images as it becomes computationally very expensive. With the optimal configuration of GPGPU kernels the image retrieval is done at a much faster rate and a substantial amount of speedup i