The Salient Region Detection Based On Superpixel

With the rapid improvement of network technology and the popularity of multi-functional electronic devices,the rich information carried in the image makes people's daily life and work more convenient,but the redundancy of multimedia information also presents new challenges.With the explosive growth of images and the uneven quality of images,how to faster and more efficiently organize,extract and present key information through image processing techniques has become more and more essential.The effectiveness of a saliency detection method is that it can automatically locate the most salient region that can attract a visual attention of human.However,the previous techniques prefer to give noisy and fuzzy saliency maps,which will be a crucial limitation for the performance of subsequent image processing.Thus,this thesis presents a superpixel-based framework of bottom-up saliency region detection,combining a variety of priors and bias to enhance visual saliency detection.Firstly,the coarse salient map is obtained by improving the existing method.The superpixel segmentation algorithm is used to aggregate similarity regions in order to generate the superpixel image,and then the initial salient map is obtained by combining the color contrast,center prior and the spatial correlation.Different from the traditional methods which calculate the mean color of each superpixel,this thesis directly selects the most frequently occurring color as the dominant color of the corresponding superpixel,which can effectively reduce the artifact introduced by segmenation and then improve the robustness.Based on that,this paper establishes a color palette to re-quantize the original picture,and quantized image is similar to the original image but has less color information,greatly decreasing the computational complexity.Finally,this paper employs the color histogram smoothing algorithm to recalculate the global context,resulting in a coarse salient map.Secondly,this thesis presents an iterative framework to further optimize the coarse saliency map.On the one hand,considering that the saliency region is usually not directly connected with the image boundary,the image saliency can be defined as the contrast between the current superpixel and the boundary superpixels,and then the proposed approach can directly use the boundary knowledge of the image to greatly suppress the image non-salient area.On the other hand,in view of the fact that the saliency regions are usually aggregated and most of the significant pixels have similar colors,there is a higher possibility of greatly highlighting the significant regions by imparting a similar saliency value to all the superpixels with similar color.This framework can iteratively refine the coarse saliency map to effectively increase the contrast between the salient and non-salient regions until a high resolution,less noise and almost binary salient map is reached.The saliency map can be viewed as a priori constraint on the original image,providing sufficient foreground and background information for subsequent images.In this paper,saliency map of the proposed algorithm is directly applied to the content-based image retargeting,automatic alpha matting and other image editing procedures.The experimental results demonstrate that the proposed approach is suitable for combining with practical applications and greatly improving the performance of the original application.