Research On Image Structure, Texture And Partial Field Cooperative Decomposition Method

Image decomposition addresses the well-studied problem of factorizing an input image into its different components that separate the prominent structures of depicted objects from residual image details such as texture.Solutions to this problem have been shown to be of importance in both computer vision and medical image analysis.However,on one hand,most of existing approaches such as the popular bilateral filter(BF)and bilateral texture filter(BTF)are lack of functions to represent special structures of image objects(such as the vascular structure).On the other hand,they often regardless of bias field which is an undesirable low-frequency image artifact and can highly degrade the performance of image decomposition.In this paper,we propose a novel image decomposition model which can be implemented by cooperatively separating a single image into its tubular structure,texture and bias components.For a given image,our cooperative scheme is carried out by combining the bias filed estimation with a novel image decomposition operator,which is successful at separating tubular structures from texture details without getting distracted by bias field.The novel decomposition operator is conducted by determining two image filtering kernels for the traditional BF: a spatial kernel of optimally Line Spread Function(OLSF)and a range kernel based on Patch Shift(PS)operator.Specifically,our contributions can be summarized by the following details:(1)We provided a novel image decomposition approach for separating tubular structures from image details,then applied it in the task of retinal image denoising.This approach can effectively distinguish between vascular structures and image noise by employing OLSF which is able to extract some special features of the vascular structures,such as the orientations and scales of local vessels.Experimental results on both synthetic images and retinal images show that this method is superior to the classical BF in preserving thin vessels with low contrast.This work provides a possibility of image decomposition on retinal images.Moreover,it can also work equally well for any image that contains small,low-contrast and straight tubular structures,which can be severed as a foundation for our following study.(2)We proposed a new filtering approach for decomposing tubular structure from texture details in one image.This approach is based on a novel BF which is a fusion of the PS operator and the OLSF we proposed before.In particular,the PS operator is established by representing the texture feature of the pixel as local statistical features of the patch.It has been proved to be effective in separating structure from texture details.We employ PS operator and OLSF to establish the range kernel and spatial kernel of BF,respectively.Compare with the classical BF and BTF,the outstanding performance of our approach has been verified by results of both retinal images and nature images.(3)We proposed a collaborative decomposition model for accurately decomposing one image into tubular structure,texture and bias field,and implement it by adding a bias field estimation.For the estimation of bias field,we employed the sparseness property of the gradient to separate the bias field from the image,then separate the texture and structure component by the method we provided before.Comparative experimental results on both nature and retinal images show that the collaborative decomposition model is more rigorous and practical than the traditional model.Compare with the classical BF and BTF,collaborative decomposition approach can obtain outstanding performance in preserving tubular structure without disturbed by image bias.