Research On Image Hierarchical Semantic Segmentation Based On Superpixel

Semantic segmentation is a crucial link in image understanding.In order to achieve high-quality semantic segmentation,two problems usually need to be solved:1)how to design an effective feature representation to distinguish objects of different categories;2)how to use context information to ensure consistency between pixel labels.In order to solve these two problems,this paper researches on three aspects:superpixel segmentation,two-dimensional contour accurate matching and analysis,and image hierarchical semantic segmentation based on superpixel,and proposes corresponding processing methods,which are verified by experiments.The main research work and results of this article include:1.In order to achieve an accurate,fast and easier-to-apply superpixel segmentation method,this paper proposes a clustering-based fine-to-coarse superpixel segmentation algorithm FCSS.By introducing color thresholds and depth thresholds with physical meaning as algorithm parameters,high-quality segmentation with fewer superpixels is achieved,which reduces the upper layer.The complexity of the application provides an easy-to-understand interface.The experimental results on the BSD and NYU-Depth V2 datasets show that when the number of superpixels is 100,FCSS can obtain higher segmentation performance,which is superior to the existing mainstream algorithms.2.Aiming at the problem of precise contour matching and part recognition in object recognition,this paper proposes a method for precise contour matching and analysis based on the minimum point-pair cost.By combining the prototype knowledge base,two strategies,coarse-to-fine secondary matching and minimal point-pair cost exact matching,are introduced for contour matching.Among them,the coarse-to-fine secondary matching strategy can effectively reduce the sensitivity of the matching process to changes in contour details.Experimental results on MPEG-7 and Animal data sets show that the proposed method is effective and feasible in object segmentation,contour recognition and part recognition in object recognition,and all have high accuracy.3.Semantic segmentation is a key step in image understanding,and its intelligibility is very important.Aiming at this problem,this paper proposes an interpretable superpixel-based image hierarchical semantic segmentation method.First,a class feature library is constructed from the data set using a feature extraction algorithm.Then,a superpixel hierarchical tree is constructed by a superpixel segmentation algorithm and a superpixel fusion algorithm.Finally,the graph model is used to decode and reason about it to obtain the final category label.The experimental results on the SiftFlow dataset confirm the effectiveness of the strategy.Although its accuracy is not as high as the deep learning method,it has advantages such as interpretability and growth,which are worth exploring.The main research contributions of this article:Introduce the priority strategy,build a superpixel hierarchical tree based on the superpixel segmentation algorithm and superpixel fusion algorithm,propose a semantic segmentation method with the advantages of interpretability and growth,and verify it through experiments.The method is effective and feasible.