Research On Convolutional Neural Network Based Image Segmentation Quality Assessment

With the rapid development of Internet information technology,the degree of digitization of modern scientific research and daily life has continuously increased,and media data is becoming a major source of information.As image is the main carrier of media data,the technology of efficient image intelligence semantic analysis has become an indispensable attribute of the improvement of production efficiency and living standard.As the foundation of image processing,image segmentation is the first step of image analysis,aiming to provide basic visual content analysis for high-level semantic image understanding.However,due to the variety and complexity of image data,segmentation is an ill-defined problem.Meanwhile,segmentation is a challenging task which covers the steps of feature extraction,prior modeling,segmentation model construction and segmentation quality assessment.The segmentation quality assessment(SQA)is the last step of image segmentation,which undertakes the task of measuring the quality of segmentation results,and is of positive significance for obtaining high-quality segmentation results consistent with human perception system and optimizing segmentation algorithm performance.Segmentation quality assessment is devoted to the measurement of segmentation quality,which is challenged by how to semantically describe the quality of segmentation results.Existing methods model the SQA task as a regression problem,and implement SQA by regressing the quality features of input segmentation results.However,there are still some challenges to be overcome: 1)how to extract the segmentation quality features more semantically,2)how to construct SQA regression models more effectively,3)how to optimize segmentation methods based on SQA network.To this end,this thesis carries out the studies of CNN based segmentation quality assessment method.The main contributions are as follows:1.This thesis studies the CNN based segmentation quality assessment method aiming at the quality assessment of single image segmentation result.This thesis proposes a multi-branch network to capture various types of quality cues of segmentation results,which combines the original image and three down-sampling methods,and measures the quality of the segmentation results by using more context cues to semantically describe the quality features of the segmentation results.2.This thesis studies the quality assessment for correlated segmentation results.Since the similarities between them leads to the difficulty of quality assessment,this thesis constructs a CNN based SQA network by introducing the quality ranking information,which evaluates the quality of correlated segmentation results by utilizing both distance and comparing supervisions,and significantly improve the performance of quality assessment of correlated segmentation results.3.This thesis studies the bounding box based segmentation optimization method based on SQA to tackle mis-segmentation problem caused by fixed parameters in bounding box based segmentation methods.This thesis measures the quality of segmentation results obtained by different settings and selects adaptive parameters through using SQA network,and further constructs an optimization framework for bounding box based segmentation.Moreover,the bounding box based segmentation method can be improved efficiently by using the segmentation quality assessment.