Reduced Methods Of Image Information

With the development of high-speed computer and integrated circuit, digitalimage processing has made great break through and been widely used in many areas,such as biomedicine, industrial manufacture, space exploration, public security,literature and art. However, there are some problems that prevent further developmentof image processing. Mass data problem is one of the difficulties among them, whichcauses lots of trouble in image store, transport and computation. To solve this problem,many reduced methods of image information have been proposed, such as imagecompression, image thresholding, and region of interest detections. But the problemhas not been totally solved yet. Therefore, in this paper, we do some researches onimage thresholding and region of interest detections. The main work can besummarized as follow:Firstly, we find that existing global image binarization methods cannot ensurepreservation of most image detail information, which results in poor robustness andgenerality. Therefore, a novel histogram, which is called histogram of cumulativeedge gray-level transition range, is proposed. Since it measures the amount of imageinformation that different gray-level can preserve, an optimal threshold is selectedfrom this histogram, which ensures maximal image information will be preserved.Because the binary results of images with variable background of intensity are notreliable by global methods, a local adaptive binarization method is proposed.Compared with other methods, our local adaptive algorithm shows great superiority inspeed and good tolerance to noise.Secondly, we find that most of existing image multi-thresholding methods cannotchoose the number of threshold automatically and are set manually. Therefore, amulti-thresholding method based on cumulative edge removing strategy is proposed.After one threshold is set, the edge points that belong to this threshold are removed,and histogram of left edge is re-calculated. The algorithm ends automatically whenRPC reaches the set value. This strategy ensures each level of threshold is optimal andproves better robustness. Compared with other methods, our method sets the numberof threshold automatically and restores original image better under the same numbersof threshold. Finally, we find that regions of interest detected by the classic Itti model which isbased on two dimensional image do not match the result of human binocularstereoscopic vision very well. Therefore, a method based on three dimensionalinformation is proposed. The depth information is acquired by Kinect developed byMicrosoft. By using center-surround strategy and combination information of depth,luminance, orientation and color, our method generates a saliency map, whose regionof interest coincides with human visual system much better.