The Research Of Contextual Information Fusion Algorithms In Semantic Segmentation

The image is an important information carrier.With the popularity of mobile devices and the Internet,more and more people share their daily life through images.How to make machines automatically and efficiently understand the semantic information in the images has attracted increasing attention in the artificial intelligence community,promoting the emergence and development of semantic segmentation technology.Semantic segmentation plays an important role in many real applications,such as autonomous driving,human behavior analysis,and medical diagnosis.However,the variation of image quality(e.g.,different illumination condition,occlusion,and resolution)and the complexity of image content(e.g.,different scenes and object categories)pose a huge challenge for the semantic segmentation task.As a result,the segmentation results usually suffer from the issue of inconsistency.In recent years,more and more work has focused on using contextual information to improve the performance of semantic segmentation and has made significant progress.In this paper,we analyze the weakness of previous methods,and propose efficient contextual information fusion algorithms for two important semantic segmentation tasks,i.e.,the scene parsing task and the human parsing task,respectively:1.We propose an efficient contextual information fusion method to make use of the long-range contextual information for the scene parsing task.We design a spatial pyramid sampling mechanism and embed it into a standard non-local block,leading to greatly reduced computational overhead.Compared with standard non-local block,our method reduces the running time by around 6 times,and at the same time reduces the GPU running memory occupation by 28 times.Extensive experiments on three scene parsing benchmark datasets demonstrate the effectiveness and efficiency of our approach.2.We design a flexible contextual information fusion method for the human parsing task.Our approach effectively utilizes the contextual information from neighboring regions for each pixel in the feature map,and can easily adjust the size of the neighboring regions considered to avoid introducing interference information.In both the single-person and multi-person parsing datasets,our fusion method shows superior performance and efficiency.In addition,cross-dataset experiments on a challenging video surveillance dataset verify the generalization ability and robustness of the proposed approach.In this paper,we propose two contextual information fusion algorithms,which are suitable for the scene parsing task and the human parsing task,respectively.The proposed methods can accurately identify and segment objects in the image,and effectively address the issue of inconsistency in the semantic segmentation task.