Image Saliency Detection Method Based On Fusion Region Contrast

The saliency of the image is expressed in the form of the region,which is the most noticeable and most expressive part of the image.Saliency detection is to extract the significant parts of the image by simulating the human visual system.Image saliency detection plays an important role in the field of computer vision,and is the basic step to solve the computer vision problems such as object recognition,image retrieval,scene description and so on.At present,there are many saliency detection methods.Depending on the implementation mechanism,these methods can be broadly divided into three categories: the class of biological visual simulation,the class of pure math computation and the integration of two classes.According to the contrast way,these methods can also be roughly divided into three categories: the class of global contrast,the class of local contrast and the integration of two classes.And depending on the processing granular,these methods can be divided into two categories: the processing method based on pixel and the processing method based on pixel block.In this paper,several representative methods are selected and analyzed in detail.And based on the experimental results,the characteristics and limitations of each method are summarized.The simple computational method based on biological simulation takes a long time,and uses a lot of conjecture theory,so the result may be not reliable.The early pure mathematical methods are simple and fast.But because they are too "rational" and ignore the visual characteristics,the test results are often poor.Mathematical methods in recent years have expanded the feature space to measure the saliency of the image,and the quality of the detection has been improved.The detection results of their combination are good,but need to pay a greater computational cost.The method based on local contrast can well detect the edge information of the image,but can not highlight the full significant goal.The method based on global contrast can detect the significant target completely,but the edge of the target is usually not very clear.The detection method combining the two contrast ways can not only maintain the saliency of the target but also preserve the good edge information.Compared with the pixel level processing method,the saliency map based on the pixel block processing method is more in line with the human visual attention characteristics.To overcome the limitations of existing methods,this paper proposes an image saliency detection method based on fusion region contrast.Firstly,the input image is segmented by the improved SLIC algorithm,and the superpixel segmentation image with good edge information is obtained.The superpixels with similar color feature are fused to get the basic region segmentation image through the DBSCAN clustering method.Through this step,the maximum connectivity of superpixels is ensured.Thirdly,the significant value of every fusion region is computed by the color and distance features,and a salient threshold is adaptively determined by using OTSU algorithm.These regions are included in the region of interest if their salient values exceed this salient threshold,otherwise they are identified as the region of background.At last,the salient scores of the regions in the region of interest are calculated by local contrast so as to better project a salient object,thus saliency map of high quality has been gotten.The method in this paper and contrast method are tested in MSRA-1000,SED2 and ECSSD,and evaluated in subjective and objective two aspects.The experimental results show that the subjective effect of saliency map obtained by this method is more consistent with human vision,and the results of the Precision-Recall curves and the values of F comparison show that the method in this paper has excellent performance.