REGION GROWING IMAGE SEGMENTATION ON LARGE DATASETS USING GPU

Abstract

Image segmentation is an important image processing, and it seems everywhere if we want to analyze what inside the image. There are varieties of applications of image segmentation such as the field of filtering noise from image, medical imaging, and locating objects in satellite images and in automatic traffic control systems, machine vision in problem of feature extraction and in recognition. This paper focuses on accelerating the image segmentation mechanism using region growing algorithm inside GPU (Graphical Processing Unit). In region growing algorithm, an initial set of small areas are iteratively merged according to similarity constraints. We have started by choosing an arbitrary seed pixel and compare it with neighboring pixels. Region is grown from the seed pixel by adding in neighboring pixels that are similar, increasing the size of the region. When the growth of one region stops we simply choose another seed pixel which does not yet belong to any region and start again. This whole process is continued until all pixels belong to some region. If any of the segment makers has the fusion cost lower than the maximum fusion cost (a given threshold), it is selected to grow. Avoid information overlapping like two threads attempting to merge its segment with the same adjacent segment.  Experiments have demonstrated that the proposed shape features do not imply in a significant change of the segmentation results, as long as the algorithm’s parameters are properly adjusted. Moreover, experiments for performance evaluation indicated the potential of using GPUs to accelerate this kind of application. For a simple hardware (GeForce 630M GT), the parallel algorithm reached a maximum speed up of approximately 20-30% for different datasets. Considering that segmentation is responsible for a significant portion of the execution time in many image analysis applications, especially in object-oriented analysis of remote sensing images, the experimentally observed acceleration values are significant. Two variants of PBF (Parallel Best Fitting) and PLMBF (Parallel Local Mutual Best Fitting) have been used to analyze the best merging cost of the two segments. It has been found that the PLMBF has been performed better than PBF.  It should also be noted that these performance gains can be obtained with low investment in hardware, as GPUs with increasing processing power are currently available on the market at declining prices. The parallel computational scheme is well suited for cluster computing, leading to a good solution for segmenting very large data sets.