Research On Image Segmentation Based On Level SET

In the field of computer vision, segmentation is dividing a digit image into several parts.The objective is to simplify a digit image or acquire the meaningful parts for the purpose ofanalysis. After segmentation, each part has a similar property. For example, all pixels in a partshare a similar color, gray scale or texture. As a fundamental step of further processing,segmentation has numerous applications, such as medical image analysis, remote sensingimage processing, robot navigation, agricultural engineering and video surveillance, to name afew of them. This research is theoretically significant and plays an important role inautomization. This thesis summarizes the theory of level set method, calculus of variationsand active contour model; conducts research and proposes improvements for imagesegmentation by reviewing some typical active contour methods:(1) This thesis proposes a new model based on C-V model and prior information, whichhas several merits in comparison with traditional active contour models: First it does notrequire knowing the number of parts in images and since has a strong adaptivity to thebackground change. Second it is easy to extend the proposed model since it supports multiplepriors. Finally, the new model is computationally efficient compared with the equations fromclassical models, such as C-V, LBF. This new model is applied to segment agriculturalimages. Experimental results demonstrate that the model proposed by this thesis can get quitegood results. So this model is an efficient way to segment complex agricultural images.(2) The derivation of segmentation algorithm based on tracking distributions is verycomplex. It involves Green's Theorem, differential geometry and transformation betweencurve evolution equation and level set evolution equation. Moreover, the curve evolution isjust an intermediate product and a tool for the deduction of level set evolution equation. Sothis makes the deduction hard to be understood by readers. In order to overcome this defect,this thesis proposes a new derivation of segmentation algorithm based on trackingdistributions. Here our new derivation does not introduce the curve evolution and onlyinvolves variational theory, which makes the segmentation algorithm more tractable. Inaddition, we acquire the same level set evolution equations after verifying two types ofdensity matching criterion, showing that our derivation is correct.(3) This thesis applies the improved segmentation algorithm based on tracking distributions to segmentation of animal images and extends to animal tracking. This algorithmcan estimate the boundary of animals correctly, it is a new method for intelligent monitoringand behavior analysis of animals.