Research On Watershed Image Segmentation Method Based On Non-separable Wavelets

Image segmentation is an import field in image processing and pattern recognition.Its purpose is to separate the interested target area from the image,and the segmentation result directly affects the subsequent information processing.In order to obtain the accurate position of the target image and its complete contour,a variety of segmentation methods have been developed by various scholars.The watershed algorithm is widely used because of its fast calculation speed,accurate positioning of the object's edges,and the ability to obtain continuous,closed,single-pixel wide edges.It is a commonly used morphological segmentation method based on gradient picture.However,in this algorithm,the gradient operator is usually affected by other local irregularities such as noise and quantization errors,and it's easy to generate a large number of scattered regions,resulting in over-segmentation.The separable wavelet can only extract the horizontal,vertical and diagonal information of the image when it is used in image decompose,which is easy to cause under-segmentation when it is applied to image segmentation.To solve those problems,an watershed image segmentation method based on non-separable wavelet is proposed in this thesis.The main research work is as follows:In this thesis,we study the theory of two-dimensional wavelet,and construct two-channel non-separable wavelet filters with tight support and orthogonality,three-channel and four-channel non-separable wavelet filters with tight support,orthogonality and symmetry by using the construction of high-dimensional non-separable wavelet and filter banks,and propose an image segmentation method based on the combination of non-separable wavelet and watershed transform.In this method,the first step is using the two-dimensional wavelet theory to construct an non-separable wavelet filter bank,and utilizing the constructed filter to perform multiscale non-separable wavelet decomposition on the original image.Secondly,each low-frequency sub-picture decomposed is labeled with foreground and background,and the watershed algorithm is used to obtain label matrixs,then the label matrixs are used to regionally average the low-frequency sub-pictures to reduce the impact of quantization errors and other local irregularities on the segmentation results.After that,the low-frequency sub-pictures obtained after the regional average is used to replace the original low-frequency sub-pictures by performing the non-separable wavelet reconstruction transform operation to get the gray-level average image.Finally,the marked-watershed algorithm is used again to acquire the final segmentation result.The experimental results show that the method can effectively solve the watershed over-segmentation and the separable wavelet under-segmentation problem,and retain the contour information of the image target well.