Research On Natural Image Segmentation Algorithm Based On Subspace Clustering

In the related researches of computer vision,image segmentation occupies an irreplaceable important position as a separate application direction or as one of the preprocess procedures of algorithms such as object detection,object tracking and image content understanding.Nowadays,there is a huge amount of mature segmentation methods have been developed.These existing methods can be roughly divided into two classes: traditional segmentation methods and deep learning segmentation methods.Although those two kinds of methods both contain specific algorithms to deal with unfavorable situation which has a bad influence on segmentation result such as partial visible noise,small area of defect,unbalance illumination within the image,they are helpless when encounter internal abstract information or redundancy information and complex noise in high dimension.Subspace clustering algorithm may be image information from the high-dimensional to low-dimensional mapping,so release redundant information between the dimensions and complexity of the noise,the noise was removed in the abstract sense,so this sparse subspace clustering algorithm is applied to this paper's image segmentation task.Based on natural images as the segmentation target set,this paper first performs a series of preprocessing operations on the target data set: the superpixel segmentation algorithm is used to reduce the calculation complexity and space occupation of the subsequent segmentation algorithm.And according to the characteristics which natural images contains such as obvious difference among most objects' color,clear texture and so on,color histogram feature,local binary pattern and deep feature extracted from trained neural network model are extracted in units of superpixel blocks and used.Secondly,in the process of image segmentation based on clustering method,this paper embeds the classic sparse subspace clustering algorithm,and analyzes its defects based on its segmentation results.Thus,a re-weighted multi-task sparse subspace is proposed,and further experiments are carried out to verify that this paper's method has better advantages than classic sparse subspace clustering algorithm and other existing sparse subspace clustering methods.However,due to the fact that clustering method is essentially an unsupervised method,and the features used for classification are relatively monotonous and shallow,even if deep features are added,the segmentation results are still not optimal.In contrast,deep learning method as a convenient and effective,end-to-end and supervised method,has become the main research direction in the field of image segmentation.Therefore,based on the above research,this paper further encapsulates the sparse subspace clustering algorithm and its improved algorithm into a module with learnable parameters,embeds them into the classic segmentation network u-net,and proposes SSCUnet and RSSCUnet,which both achieves better segmentation effect.At last,Conditional random field algorithm is used to optimize the contour of the segmentation result,and further improvement has been achieved.