Image Segmentation Based On Active Contour Models With Structural Prior Information

Image segmentation is an essential task of image processing that connects low and high-level computer vision.Image segmentation strives to divide an image into some unique sub-regions so that they can be served as a pre-processing step for image processing applications.Image segmentation continues to be the most demanding issue because of feature artefacts of images,e.g.,low contrast,noise and intensity variations or inhomogeneities.Active contour model(ACM)is one of the distinguished energy-based approaches in image segmentation.In this dissertation,we have designed four Active contour models to segment various types of images with low contrast,fuzzy borders,complicated background and intensity variations or inhomogeneities.The key contributions presented in this dissertation are precisely summarized below.Firstly,we proposed a new Signed Pressure Force(SPF)function named as Hyperbolic Trigonometric Signed Pressure Force(HTSPF)function that can be used to detect the contour of ROI of diverse intensities,even at weak and blurred borders.Our HTSPF function utilizes the harmonic mean intensities of the image that result in effective segmentation of low contrast homogeneous-intensity images.Besides,a global region-based ACM is proposed,which integrates the HTSPF function in the level set framework of "Active Contour Model using Arithmetic Mean"(ACMAM).In this way,the ability of the prior model(i-e,ACMAM)is improved by utilizing the harmonic mean intensities into our designed HTSPF function,by which the segmentation results of homogeneous-intensity images are promoted.Experimental results confirm that our designed global region-based ACM is rapid and more effective than other related models for segmentation of homogeneous-intensity images which have multiple objects with diverse intensities,fuzzy borders and low contrast.Secondly,we designed a hybrid ACM named as HMG to improve the segmentation of medical images with low contrast,hazy edges and ultrasound images which have multiple objects with dissimilar distributions.Our proposed hybrid ACM uses the edge-based level set framework of Distance Regularized Level Set Evolution(DRLSE)model and replaces its region information with global harmonic mean based SPF function.This integration of global harmonic mean intensities of the image in the framework of the DRLSE model delivers robust and effective segmentation results by preserving the subtle difference between the desired object intensities and its background.The experimental results demonstrate that our proposed hybrid ACM is more efficient,rapid and accurate as compared to the other related models.Thirdly,we designed a hybrid ACM "Global,Local Bias-corrected ACM"(GLBACM)to cope with noise,complex background and intensity variations or inhomogeneities in various types of images.Our designed model GLBACM is motivated by global and local bias-fitting energy and its energy function is comprised of both the global fitted image(GFI)and local bias-fitted image(LBI).Due to the introduction of GFI,the progression of the contour speeds up over smooth intensity regions,while LBI contributes in the segmentation of objects with intensity inhomogeneity by retaining the fine local details combined with bias correction.During the minimization of this energy function by the gradient descent method,global image difference is substituted by the novel "Global Hyperbolic Trigonometric Signed Pressure Force"(GHTSPF)function and local image difference is substituted by the novel "local Hyperbolic Trigonometric Signed Pressure Force"(LHTSPF)function.Our GHTSPF function is designed by utilizing the difference between GFI and the observed image which results in segmenting the fuzzy and weak border objects with a higher convergence rate as compared to other related models.In contrast,our LHTSPF function is designed by utilizing the difference between LBI and the observed image which makes it capable of segmenting complex objects in the presence of noise and intensity inhomogeneity.Experimental results on various types of images with noise and intensity variations or inhomogeneities demonstrate that the designed GLBACM outperforms the related segmentation models.Finally,we developed a new hybrid ACM named as "Global-Local Harmonic Mean based ACM"(GLHMACM)to improve the convergence speed with accuracy for the segmentation of images with intensity variations or inhomogeneities.Our GLHMACM model improves the framework of ACMAM model by utilizing an improved SPF function in its energy function.The proposed SPF function named as "Global-Local Harmonic Mean based SPF"(GLHMACM_SPF)function is based on a global-local harmonic mean fitted image.The global-local harmonic mean fitted image integrates both the global harmonic mean intensities and local mean intensities of the image.In this way,the ability of the ACMAM model is improved by incorporation of both the local and global region image details in our GLHMACM model.The experimental results demonstrate that the designed GLHMACM model reveals desirable segmentation results with faster convergence speed.Furthermore,the designed GLHMACM model is less susceptible to placement of initial curve and noise as compared to other related ACMs.