Image Salient Region Detection And Its GPU Parallel Computing

The visual attention mechanism is of great significance for the survival and reproduction of the animals relying on their visual perception. For instance, animals can capture prey or evade predators efficiently as noticing them instantly. Besides, it also has important significance in the bionics. With the amount of explosive growth image data nowadays, processing the digital images equally is not only slow, but also a waste of resources. Visual attention mechanism enlightens us to simulate human visual in the area of computer vision, so as to extract the salient region or region of interest in the image for post-processing.Salient region detection belongs to visual saliency analysis category. It introduces the visual attention mechanism into the image processing, which reduces the amount of data by screening and selecting the important area of the image. Reliable estimation of salient region allows screening the important information of images without prior knowledge of their contents. Thus, salient region detection is an important step in many computer vision tasks including image segmentation, object recognition, adaptive compression, content-aware image editing and image retrieval etc. Therefore, effective salient region detection can greatly improve the overall efficiency of these tasks.Humans routinely and effortlessly judge the importance of image regions, and focus attentions on important parts instantly, no matter how complexity the background is. The main goal of this paper is to gain some real-time image region detection algorithms by accelerating the salient region detection with GPU parallel computing, which is becoming one of the most popular HPC (high performance computing). Based on the analysis of four salient region detection algorithms:Itti, FT, MSSS and HC, we design the MSSS and HC’s GPU parallelization algorithms and improve the MSSS algorithm by incorporating a factor of center prior. Compared with the original sequential algorithm in our implementation, the parallelized MSSS and HC algorithms achieve speedups of over40x and60x respectively using one of the largest publicly available data sets. Lots of experiments show that the improved parallel MSSS algorithm not only achieves a speedup of over30x, but also predicts salient region better.