Research Of Image Segmentation In Web Image Search Based On GPUs

The complexity of visual information makes image processing time-consuming. The amount of visual information users submit is much more than that of text information when retrieving. To reduce time consumed, CBIR system has to use those algorithms of high speed but low effect, or process images after they are reduced to a smaller size to get a high speed. But both methods degrade the precision of retrieval results and satisfaction of users. Speeding up the image processing module using GPU can well solve the problem.The framework of GPU based CBIR system consists of 5 layers as follows, user interaction layer, cluster management layer, CPU based module processing layer, GPU based module processing layer and data center layer. The image processing module includes thumbnail generating sub-module and image feature extracting sub-module, both of which are bottle-necks in this real-time system. This paper does a study on some important image processing algorithms.Effect of image segmentation directly determines whether the image feature extracted is meaningful or not. And more, it seriously affects the precision of retrieval results. Image segmentation is complex and time-consuming, so it is the key factor of those which affect the retrieval efficiency and the speed of back-end data updating. This paper implements several GPU based image segmentation algorithms, including c-means clustering, Canny edge detection, watershed segmentation, Mean Shift and fuzzy c-means algorithms, etc. The system first uses Mean Shift on GPU to pre-segment images, and then uses fuzzy c-means to segment images in detail. This method not only enhances the effect of segmentation, but also speeds up the segmentation considerably.This paper uses the feature of cell automata machine and Bellman-Ford shortest-path to parallelize watershed algorithm. This paper use the feature of texture memory to degrade the number of non-aligning access of global memory in Mean Shift algorithm. In C-means clustering, this paper proposes a novel method to make clustering centers be not need to transfer back to CPU when updating them. The whole procedure is completed in GUP. This method can obtain a much better speedup. In Canny edge detection, this paper sacrifices more memory space to win a better time efficiency, which enhances the computing intensity of each thread. At last, the paper applies these GPU based algorithms in image processing module, make CPU and GPU cooperate concurrently and asynchronously and get a final speedup of 2 times in thumbnail generating sub-module, tens of times in feature extracting sub-module.