Research On Image Segmentation Method With Support Vector Machine Using Multi-features

Image segmentation is a basic step of image analysis, meanwhile, is also one of difficultand hot research issues in image information processing. The quality of image segmentationresults has important influence on subsequent image analysis and understanding. At present,the common image segmentation methods include threshold segmentation, edge detection,regional growth, neural network, support vector machine(SVM) and fuzzy clustering, etc. TheSVM not only uses structural risk minimization principle, but also comprehends the neuralnetwork, the statistical learning theory, etc. It has been proved that the SVM has theadvantage of effectively solving pattern recognition issues with characteristics of smallsamples, high dimension and nonlinearization. In recent years, many scholars have paidattention to image segmentation methods using the SVM.The author investigates image segmentation methods based on the SVM in detail andfurther find that their used features of samples are extracted only from gray-level information,rather than image texture information and edge sharp change. However, for rich textureinformation and target areas on the edge of images with low contrastness, we can notcharacterize the overall target only relying gray-level features. Therefore the classical imagesegmentation methods based on the SVM are hard to get a satisfactory segmentation effect.For the above problem, the main work is as follows:(1)After analyzing the importance of frequency domain phase information and textureinformation in characterizing image features, a new image segmentation method which usesphase consistency and texture features is proposed. The new method combines phaseconsistency statistic characteristics, texture features and gray-level features into a trainingeigenvector and segments images with the SVM classification technique. Phase consistencystatistic characteristics are described by the mean, variance, skewness, kurtosis and entropy.Texture features are characterized by the energy and frequency domain directional. Gray-levelfeatures are formed by gray pixel values.(2)Comparing the new method with the classical image segmentation methods based onthe SVM and Canny edge detection method in experimental analysis. The experimentalresults show that the new method is more effective than the classical methods, especially in the situation when there is low edge contrast and rich textural information in target areas ofimage. At the same time, the new method avoids the case where target’s shadow is treated asedge mistakenly, to which Canny edge detection method is prone when there is low edgecontrast in target areas of image.