Please use this identifier to cite or link to this item:
|Title:||Automated retinal vessel detection||Authors:||Zhang, Xinyi||Keywords:||DRNTU::Engineering::Electrical and electronic engineering||Issue Date:||2017||Abstract:||As the first step of automatic retina image analysis, automatic retinal vessels segmentation is a meaningful and important topic for auto diagnosis. There are several algorithms currently can realise automatic retinal vessels segmentation with a relatively high accuracy. Based on the existing research, this thesis chooses artificial neural network (ANN) as the main segmentation algorithm followed by post-processing method to achieve vessels segmentation. This thesis chooses DRIVE  and STARE  as the experiment databases. After comparing several different neural network structures, this thesis selects out the most suitable structures to generate probability image for each fundus image. Tests different threshold selection methods on different database and chooses Otsu's method  at last as the way to find the most suitable threshold to get the segmentation result in the form of binary image. Applies two post-processing methods on the binary image based on the knowledge of morphology  to improve the segmentation performance. Explores the different influence of truncation filter  and data source (green channel or luminance) on the final segmentation results which is related to the database. This thesis also displays the performance of cross training to show the robustness of the method. The final accuracy can reach 0.9516 and 0.9608 in DRIVE and STARE separately. The performance can be improved further is we adjust the data source or other procedures according to the input images' quality. More details are discussed in the experiments and conclusion sections.||URI:||http://hdl.handle.net/10356/72618||Fulltext Permission:||restricted||Fulltext Availability:||With Fulltext|
|Appears in Collections:||EEE Theses|
Items in DR-NTU are protected by copyright, with all rights reserved, unless otherwise indicated.