A Method For Image Salient Regions Detection Based On Layers And Dynamic Threshold

In resent years, salient regions detection has become a extremely active research field. At present, there exist some methods for salient regions detection, such as: Contrast-based image attention analysis by using fuzzy growing, SVM-based salient region(s) extraction method for image retrieval, and so on. These methods generally deal with pertinence among the involved pixels in order to fulfill our goal of salient regions detection. But, these methods ignore the relationship among pixels, and can not provide satisfying results.This dissertation improves the model for detecting salient regions based on layers. We first divide the image into layers. And for every layer, we calculate saliency, and then add the results of every layer. We use dynamic threshold to extract salient regions.For every layer, extracts basic characteristics and calculates saliency characteristics including intensity saliency, saturation saliency and color saliency. And then, we synthesize and analyze every characteristic to eliminate the effect of noise pixels on salient regions detection.Because of the possibility characteristics of images, the planar normal distribution function can take charge of the pertinence of pixels accurately and reasonably. This step is the key of salient regions detection. On one hand, it can make salient regions detection coherently. On the other hand, it can take the pertinence of pixels into consideration and eliminate the effect of noise pixels on salient regions detection.We add the results of layers. Out of consideration of the statistics feature, we use the dynamic threshold, and then reach the final saliency regions. The dynamic threshold tends to adapt well getting rid of the effect of the original images. We apply this method to real images, and finally get a satisfying result.