Research On Object Segmentation Evaluation Based On Cluster Learning

Image segmentation plays an important role in computer vision,which is a preprocessing step of target recognition,scene analysis,target detection and other tasks.The quality of the segmentation will affect the outputs of these tasks.And present,subjective evaluation is the most commonly used segmentation evaluation method while it is tedious.Objective evaluation methods include supervised and unsupervised evaluation methods,among which supervised evaluation requires a manually segmented or pre-processed reference image.Unsupervised evaluation,without reference segmentation,is crucial to online segmentation evaluation.In order to get better unsupervised segmentation evaluation results,this thesis studies the method of object segmentation evaluation based on cluster learning.The main content is,based on edge features,design and implement an unsupervised evaluation method,which finds the local spatial structure characteristics of segmentation region by clustering on the basis of segmentation information of patch.And the evaluation is based on the segmentation information of patches.Firstly,we introduce the background of the research status of image segmentation and image segmentation evaluation.Then we present related techniques involved in this thesis,including supervised evaluation,unsupervised evaluation,edge features extraction,cluster and saliency detection.Next,we propose a segmentation evaluation method based on image edge feature clustering,where the construction of training data and the calculation of evaluation scores are introduced in detail.We experiment on segmentation evaluation database and three different segmentation granularity results.The results show that the proposed method has better performance than the unsupervised evaluation methods of F,Hw,Zeb,and Frc,and match the performance of supervised methods of Variation of Information(VOI),Global Consistency Error(GCE).While it lacks the ability to evaluate over segmentation.Then we propose the segmentation evaluation method based on saliency detection and an improved segmentation evaluation method based on image edge features clustering by introducing visual saliency and setting segmentation boundary coefficient respectively.The improved methods are also verified by experiments,and the experiment results show that the improved methods make up for the shortcoming of the original method and make the comprehensive evaluation performance.The performance of the improved method with the segmentation boundary coefficient exceeds the Probability Rand index(PRI),VOI and GCE.