Image Semantic Segmentation Based On Probability Graph Model

In computer vision, image segmentation is the process of partitioning a digital image into multiple non-overlapping segments, witch based on characters of pixels. The goal of image segmentation is to express directly different regions which include different information and inter-region relationship. Our main task is to find an expressive model which could mine for the image inter-object communication to efficiently guide the semantic segmentation and solve semantic understanding problems.Many semantic segmentation tasks are boiled down to a labeling problem that tries to assign a label to each pixel of an image. These discrete labels may vary depending on the task, for example they may correspond to different object classes, or to depths or to intensity. These labeling problems are typically formulated as a probability graph model.In this thesis, we built the graph model by a multi-layer condition random field based on different partition granularity, our energy function composed by unary potential, pairwise potential and inter-layer potential, we use graph cut based move making algorithms to find an approximate optimal solution. This thesis main content is as follows:(1) We use vari-grained images as the foundation for multi-class condition random field model, including pixel layer and three super-pixel layers with different quantization. We adopt the classic Mean-shift algorithm for unsupervised image segmentation, the layer with super-pixel nodes provide effectively guide for image segmentation during the process of model reasoning.(2) Unary potential and pairwise potential. On the pixel layer, pixel-based dense features are used to calculate the probability of labels taking by pixels, we used a number of shape filter to capture spatial layout information of the observation point, let the filters as weak classifiers of Boosting algorithm and train the strong classifier used to calculate label of the observation point. Strong classifier outputs are converted to the form of probability, this is, unary potential we are concerned. Pixel pairwise potential is calculate by classical contrast sensitive potential. On the segments layers, we counted the dominant label as unary potential and calculated the histogram distance of neighboring segments as pairwise potential.(1) Inter-layer potential. In order to ensure the label consistency of the segment, this paper proposed an inter-layer potential based on the quality of image segmentation, witch encourage individual pixels within the single segment take the consistent label. We also introduced the truncate threshold to tolerant wrong labels on a small-scale. Such improvements make the model more reasonable and robust.On the MSRC-21database, our average labeling accuracy has increased18.1%compared with lower-order CRF, and increased6.3%compared with traditional higher-order CRF.