Research On Automatic Image Segmentation Technology By Simulating Human Vision

The capacity of human visual system processing information far exceedsthe machine vision system. Simulating the human visual mechanism is an importantresearch direction for performance breakthrough in machine vision. We model thehuman visual system by visual attention algorithm in this study. This study proposedan automatic image segmentation algorithm on the basis of visual saliency accordingto the theory of ‘fixation-saccade’. This thesis’s main idea is that: Firstly, we calculatethe image saliency map through image saliency detection algorithm; Then, simulatinghuman active vision, we only focus on significant pixels and ignore other pixels, sothat improving the efficiency and the ability of anti-interference of the proposedmethod; Last, we extract the most salient pixels as the samples, and build the neuralnetwork classifier through sample-learning to come true segmenting imageautomatically. We try to get stable outputs through classifier integration by simulatingmicro-saccade.The main contents and innovations in this paper are as follows:(1) We proposed an image saliency detection method (Hyper-complex Fast FourierTransform based Spectral Residual, HFFT-SR) by improving a traditional visualattention method (spectral residual algorithm, SR). Since SR algorithm extractssaliency map from the gray image, that losing the color information of images andsaliency points located on the object edges. HFFT-SR method extracts saliency mapfrom the color image reserved the color spectrum information of image, and saliencypoints located around the whole object, matching with human perception basically.(2) We apply visual attention algorithm to multi-focus image fusion field, andpropose a method of multi-focus image fusion based on visual saliency. Experimentsshow that focus area is more salient than defocus area, so we can get focus areathrough comparing the saliency of pixels between original images, to come truefusing images. Experimental results show that the proposed method can choose theclearest pixels from the focus area, obtain over37db PSNR and need no parametersettings. It also exhibits strong robustness in different type of images.(3) We proposed an algorithm of image segmentation based on visual saliency byextracting samples, active learning and classifier integration. Firstly, the algorithm can calculate the image salience automatically, and extract a few important samples fromthe most salient area by simulating human active vision. Then, we build efficientELM (extreme leaning machine) classifier rely on sample learning, and get the stableoutput through classifier integration, to come true segmenting images. Experimentsshow that the proposed algorithm performs very well for general natural scenes basedon standard image library.