Image Editing Based On Contextual Semantics

This paper put forward an image editing algorithm based on context semantic, which analyzed the context semantic of image, and was applied to image editing techniques. Image semantics is intuitive feelings of people for images, including people's understanding of objects contained in the image and one of the relation between implicit objects. It can describe the image information completely, and reduce the lack of details in the image editing greatly. The traditional image editing method mainly utilizes basic features such as color, texture and shape, while the application of semantic features appears less. The major difficulties are enormous semantic gap between visual representation of images and entities description. The reason is that visual representation is difficult to be associated with semantic, In addition, the definition of semantic itself exists big uncertainty and ambiguity, which is easy to produce ambiguity. To solve the above problem, a model of image editing based on context semantic is built based on the discussion of shape feature in the bottom in this thesis. The model consists of image representation based on context semantic—Shape Nets and jigsaw algorithm based on context semantic.Shape Nets is a image editing model with appearance characteristic of shape, compact structure, hierarchical representation of image structure and object contour. In a ShapeNets, firstly, detect the image contained in the object by genetic objectness proposal, and select the best candidate window according to training by two stages cascaded SVM. Object contour is extracted through classic contour extraction algorithm based on Canny; the spatial relationship between objects is described by using adaptive sampling granularity model, which is described with a typical node. ShapeNets is a hierarchical structure from coarse to fine. When user search expected object on the target image, the system automatically query image library to find candidate area of semantic compatibility. Apply the algorithm to a large number of such as experiments image matching, image understanding, behavior analysis, which is proved the feasibility, efficiency and practicability.The superpixel segment image and saliency propagation composite image were used by the novel designed image editing based on context semantic. The clustering objective function combining entropy rate, balance term, homogeneity, size and numbers was used to fulfill the image segmentation. When considering the great effects of the background complexity on the visual attention, the pretreatment of the images was conducted by the image editing based on context semantic. Therefore, the synthesis of images were completed by the global propagation procedure based geodesic distance. Results show that the thoughts and the framework provided by us is valuable to further research, and has a certain reference value for the future image editing and algorithm design of image analysis.