Parsing Natural Scenes Based On Hierarchical Region Merge

Natural scene understanding has been one of the central tasks in computer vision,because it can help us to use the image resource efficiently. The goal of sceneunderstanding is to recognize what objects in the image and where the objects arelocated. With the research, we found hierarchical structure is commonly found in thenatural scene images. The whole complicated object is made up of several simpleparts, and these simple parts also consist of several small parts. The procedure formsthe hierarchical structure. This structure not only can help us to identity the objectsbut also how the small parts interact to form the whole objects.Our algorithm is based on the hierarchical structure. We merge the neighboringsegments continuously until they combined into the whole object. The result is aforest which contains several trees, one tree commonly represents one object. Nowevery tree has its possible labels which computed when merged, so we only tocompute the most possible label in these labels, and the place of the segments whichmerged into the trees are the object’s location. Our algorithm can be seen as a deeplearning method for Image semantic understanding.The work in this paper as follows:(1) introduce a machine learning model to describe the merge process，includinglabel consistency、set conditions to merge、compute the possible labels of the supersegment and so on.(2) There are more than exponentially many possible to merge, so it’s feasible tocompute directly, we introduce the greedy inference to compute the best merge trees.(3) Use the max margin to learn the parameters in the model.(4) Cluster the segments features hierarchically to initialize the parameter.Through the hierarchical cluster, we can get the general characteristics of everycategory, and then our parameter has the feature to predict category.(5) We compute three different features to learn and test our model, and comparethem.