The Study On Image Segmentation Based On Visual Saliency Of Image And Improved SVM

Image segmentation is a process of dividing an image into different regions such that region of interest, but the union of any two adjacent regions isn’t homogeneous. Image segmentation is the first step and preprocess in image understanding and object recognition. The basic development process of image segmentation is as follows:the early classical image segmentation algorithms based on image intensity and gradient, the eighties active contour models, the image segmentation algorithms with prior knowledges, color image segmentation. From the development process of image segmentation, we can observe that the machine intelligence and pattern recognition ability of the image segmentation become better and better.The thesis mainly aims at researching the related task of image segmentation, object segmentation including in computer vision. The image segmentation task is to divide the image into a plurality of regions with similar characteristics as the basis, the results of image segmentation is the foundation of image understanding and object recognition, it can also be understood as the primary stage of image understanding and object recognition. In image segmentation, fusing together the consistency of image space and unsupervised clustering algorithm to perform image segmentation process by constructing a probability tree structure, proposing the global segmentation method of vector valued image by constructing the edge detection operator of vector valued image, putting forward a new model of variational form and calculation method of vector valued image with convex property, the level set is applied to a new color image segmentation process, eventually. In the part of image preprocessing, improving and enhancing the existing visual saliency metric method, putting forward the improved visual saliency metric method and multi-scale visual saliency detection algorithm, and applying the improved image preprocessing methods into image segmentation.In the basic methods of image segmentation and object segmentation, putting forward the incremental support vector machine learning methods, including the online incremental learning and online reduction learning method; the incremental support vector machine learning method proposing the variable learning support vector machine, including the incremental learning algorithms and decremental learning algorithms based on the variable learning support vector machine; variable learning support vector machine is improved to adaptive variable learning support vector machine, and these improved support vector machine theories in image segmentation and object segmentation for applied research.The main work and research achievements are as follows:(1) For applying the image space coherence and unsupervised clustering related information in the image spatial structure into the image segmentation process, proposing the application of the combination of the two to achieve the rapid implementation of a new method of image segmentation. First of all, the image segmentation process with probability tree structure has formed many overdivided regions, and these regions can achieve the ideal edge part. The unsupervised image segmentation framework based on probability is used to deal with the segmentation regions.(2) According to the vector valued image boundary detection operator, the vector valued image global segmentation method is proposed, and putting forward a new model of variational form and calculation method of vector valued image with convex property. The new model combines the excellent characteristics of many existing models. This method does not suffer from the limitations of the initial value selection, and it can prove that the energy functional has reached the global minimum in theory. The dual theorem has been introduced into the numerical calculation process, thus greatly improved the operation speed and the instantaneity of the model.(3) Compared with the existing image segmentation methods of computer vision attention mechanism, the system saliency map of improved visual saliency metric generated can more accurately express the saliency value of each pixel in the original image, the threshold segmentation according to the system saliency map can distinguish between the foreground objects and background regions, need not join the other methods, this method not only can reduce the time complexity, and can make the image segmentation results more consistent with human visual habit.(4) Multi-scale visual saliency detection algorithm can quickly generate the meticulous visual saliency result maps which are the same size as original input image. Multi-scale method can generate the visual saliency result maps which can apply into the highly instantaneity picture’s sequence based on image content. Multi-sccale method can be used for fast visual salient region detection.(5) The support vector machine algorithms of online incremental learning OFV-SVM and decremental learning ODV-SVM are aimed at the training samples, especially the large capacity training samples, OIV-SVM can effectively save the training time for the premise of ensuring its generalization ability.(6) To solve the related problems of long execution time and low execution efficiency of classical support vector machine method, the support vector machine method based on incremental learning vector is proposed. The initialization classifier can reduce the initial training set to a reduced narrow set under some conditions. Then use the reduced narrow set aim at the initial support vector machine classifier to reverse process to obtain the final classifier of support vector machine. This method can greatly reduce the learning time of large data set on support vector machine, and has good generalization ability. In order to validate the applicability of this learning method, choosing the related color image to proceeding simulation experiment from Berkeley image segmentation data set BSDS500and Internet network, the experimental results show that the process of obtaining segmentation results of this method is less time consuming than the traditional support vector machine, and can obtain better segmentation results than the manual annotation results in the Berkeley image segmentation data.(7) The object image segmentation algorithms of using visual saliency features and adaptive variable learning support vector machine. The algorithm executed the visual saliency measurement process based on the image local features, global features and special features, and the visual saliency feature and color feature are represented as object feature. We have used particle filtering algorithm to intialize particle swarm and update particle weights operation. Conducting the experiments in many pictures’ sequences, the algorithm can achieve good effective and real-time object image segmentation. The method can overcome the instability of object image segmentation caused by using single color feature, and it can effectively solve the detection difficulties caused by object deformation, illumination changes and the background with similar color distributions, it has good robustness.