A Learning-based Co-Evaluation Framework For Image Segmentation Evaluation

Image segmentation is a fundamental task in automatic image analysis. It consists of subdividing an image into its constituent parts and extracting these parts of interest. Many image segmentation approaches and algorithms have been proposed for particular uses. However, there is still no generally accepted effectiveness measure which is suitable for evaluating the segmentation quality in every application. Ensemble learning is the process by which multiple models, such as classifiers or experts, are strategically generated and combined to solve a particular computational intelligence problem. It is primarily used to improve the performance of a model for classification or prediction.In this thesis, we first summarize the existing image segmentation and evaluation methods. Then we discuss and analyze the supervised and unsupervised evaluation methods in experimental evaluation. We also introduce some machine learning methods, such as Naive Bayes, Real AdaBoost, etc. Besides, we propose two ensemble learning-based evaluation methods (supervised and unsupervised) and a learning-based co-evaluation framework which derives from ensemble learning technology and benefits from multiple stand-alone measures, including supervised (BDE, RI, VOI and GCE) and unsupervised (Intensity, Gradient Direction and Texture) methods. To this end, different segmentation evaluation measures are chosen to evaluate segmentation separately, and the results are effectively combined using machine learning methods. We train and implement this framework in our new segmentation dataset which contains images of different contents with segmentation ground truth using different machine learning methods. We also use Weizmann Segmentation Database for unsupervised evaluation methods. In addition, we provide human evaluation of image segmentation pairs which contains two segmentations of one certain image to benchmark the evaluation results of the measures. Experimental results show a better performance than the stand-alone methods.