Study On The Multi-scale And Multi-feature Image Segmentation Algorithms

Image feature extraction and segmentation is the basis of computer vision and image processing.Due to the complexity of image space arrangement,as well as the spectrum of similar features internal heterogeneity,makes the traditional image segmentation based on spectral characteristics of single,unable to obtain the better segmentation effect,which needs the effective use of the image texture information,scale and object features in the process of image segmentation.Multi-scale geometric transformation is a kind of method of image sparse representation in multi-scale and multi-direction,which is a simulation image expression of human visual system on the basis of Fourier transform and wavelet transform.Multi-scale geometric transformation can also express the characteristics of the image in the spatial domain and frequency domain,for subsequent research on the multi-scale feature extraction of image laid a foundation.In this paper,we study on image segmentation from the perspective of multi-feature and multi-scale.The main results as follows:(1)Multi-scale feature extraction and analysis.Extract the image texture feature based on Contourlet transform.First of all,the image is decomposed into high frequency and low frequency part by Laplacian pyramid;then,extract the image multi-direction features of high frequency through a directional filter bank;finally,make high frequency multi-direction features to vector stacking,get the multi-scale and multi-direction image texture features based on Contourlet transform.The experimental results show that: Multi-scale texture features are better than the traditional texture feature on the visual or accuracy;Contourlet texture can reflect more directional information than the gray level co-occurrence matrix(GLCM)texture and Wavelet texture.(2)Study on the algorithm of image multi-feature segmentation.Improve the mean shift kernel function,and apply it to image multi-feature space,then get the improved multi-feature mean shift image segmentation algorithm.(3)High-resolution remote sensing image segmentation based on multi-scale and multi-feature mean shift image segmentation algorithm.At the same bandwidth mean shift scales,integrate with multi features,do high resolution remote sensing image segmentation experiment,compared with traditional mean shift image segmentation results.The experimental results shows: multi-features image segmentation can distinguish more image details than single feature image segmentation,makes segmentation results more accurate.(4)Study on object oriented normalized cut image segmentation after the mean shift segmentation.View the object after mean shift segmentation as the node of graph,calculate the weight coefficient of the normalized cut algorithm,do normalized cut for object-oriented image segmentation experiment.The experimental results show that: the object-oriented normalized cut algorithm optimize the over-segmentation phenomena of the mean shift,improve the speed and efficiency of traditional normalized cut algorithm.