Research On The Theory And Method Of Image Semantic Content Segmentation

As a fundamental task and research hotspot in the field of computational vision,image semantic content segmentation has important theoretical significance and research value,and has been widely applied in many fields,such as autonomous driving,video monitoring,and biomedical,etc.In recent years,with the rapid development and application of artificial intelligence technology and deep learning in the field of computer vision,image semantic content segmentation has developed from a simple task of distinguishing only foreground objects and background content in images to a complex task of segmenting different semantic content in images.However,due to the fact that real-world images usually contain a large number of semantic content with semantically diverse,diverse scales,and complex relationships,how to accurately and quickly identify and segment the semantic content in images still faces enormous challenges.Therefore,in order to meet the urgent demand for computer vision applications to improve the accuracy of image semantic content segmentation,this dissertation conducts research on the theory and method of image semantic content segmentation.With the overall goal of improving the performance of image semantic content segmentation algorithm and reducing the computational consumption of image semantic content segmentation algorithm,this dissertation will study the three main tasks of semantic segmentation,instance segmentation,and panoptic segmentation in image semantic content segmentation.The specific research content and innovation include the following aspects:(1)Aiming at the problem of the influence of the accuracy of object detector on instance segmentation performance in instance segmentation,that is,because the object detector cannot completely detect the foreground object,the instance segmentation algorithm cannot completely segment the foreground object,a study of instance segmentation based on foreground object optimization is conducted in this dissertation.This dissertation constructs a mathematical model of the relationship between foreground pixels and their corresponding foreground objects and proposes an instance segmentation algorithm based on foreground object optimization to assist the instance segmentation based on object detection framework to segment the region of the foreground object that is not detected by the object detector,so as to further improve the performance of instance segmentation algorithm.(2)Aiming at the problem of confusion among foreground objects in instance segmentation,that is,because the instance segmentation algorithm based on object detection framework will transitionally respond to the features of non-target foreground pixels belonging to the same semantic category as the target foreground object in the object detection bounding box,a study of instance segmentation based foreground object confusion processing is conducted in this dissertation.This dissertation regards the problem of confusion among foreground objects in instance segmentation as the problem of transition response of non-target foreground object features and proposes an instance segmentation algorithm based on foreground object confusion processing to assist instance segmentation algorithm to distinguish the target foreground object and the non-target foreground area in the object detection bounding box,so as to further improve the performance of instance segmentation algorithm.(3)Aiming at the problem of how to simultaneously segment multiple foreground objects in the same image in instance segmentation,a study of instance segmentation based on collaborative attention is conducted.This dissertation finds that there are multiple foreground objects that need to be segmented in the same image,these foreground objects share the same background information and have complex semantic and spatial relationships with each other,especially when these foreground objects belong to the same semantic category.Therefore,this dissertation proposes an instance segmentation algorithm based on collaborative attention,which strengthens instance segmentation features of the corresponding semantic category of the foreground object by sharing appearance features and semantic features of foreground objects of the same semantic category in the same image to,so as to further improve the performance of instance segmentation algorithm.(4)Aiming at the problem of how to segment the same semantic content in different images in semantic segmentation,a study of semantic segmentation based on feature memory is conducted in this dissertation.This dissertation finds that there are also semantic and spatial relationships between the same semantic content in different images in semantic segmentation.Therefore,this dissertation proposed a semantic segmentation algorithm on feature memory,which strengthens semantic segmentation features by memorizing the semantic and spatial information of the same semantic content in different images,so as to further improve the performance of semantic segmentation algorithm.(5)Aiming at the problem of computing time and resource consumption due to the complex network structure in panoptic segmentation,a study of panoptic segmentation based pixel relationship prediction is conducted in this dissertation.This dissertation firstly builds a mathematical model based on the relationship between pixels to judge whether two pixels belong to the same foreground object; then proposes a panoptic segmentation algorithm based on pixels relationship prediction,which achieves fast panoptic segmentation while ensuring the performance of panoptic segmentation.(6)Aiming at the problem of computing time and resource consumption due to the complexity of the foreground pixel clustering method,a study of panoptic segmentation based on foreground pixel clustering is conducted in this dissertation.This dissertation first builds a mathematical model of three-dimensional space vector,which integrates the judgment of foreground pixels and background pixels,the prediction of the geometric center of foreground objects,and the prediction of the relationship between foreground pixels and corresponding foreground objects into a mathematical model; and then uses the constructed three-dimensional space vector mathematics model to propose a panoptic segmentation algorithm based on foreground pixel clustering,which achieves fast panoptic segmentation while ensuring the performance of panoptic segmentation.The research on the theory and method of image semantic content segmentation carried out in this dissertation discusses how to use the complex semantic and spatial relationship information between different semantic content in images to achieve high-quality image semantic content segmentation,and provide technical support for other computer vision tasks.The theoretical results of this dissertation can be widely used in real-world fields,such as automatic driving,medical diagnosis,and image retrieval,which have important theoretical and application values.