Research On Image Understanding Algorithm By Embedding Prior Information In Conditional Random Field Framework

Scene Image Understanding is an important research content in the field of computer vision and it is the base for the computer to understand the semantic content of vision media. Now it plays a key role in the application of image retrieval and automatic drive. In recent years, scene image understanding has received widely attension and lots of image understanding algorithms have been designed on the base of the CRF framework. Though existing algorithms have got a good performance by embedding the context of local smooth, position and co-occurrence into the CRF model, they also have some disadvantages and fail to improve the accuracy of foreground. So our paper focuses on the research of mining and exploiting the prior knowledge in the scene image. The main contributions are as follows:1. Propose a CRF based image understanding algorithm by embedding saliency prior (SPCRF). SPCRF segments the foreground object by saliency detection and builds the unified model on the foreground area and background area. Particularly, SPCRF uses the super-pixel as the classification element and builds full connected relations on the foreground area. The full connected relations have a positive effect to the problem of inconsistent classification due to wide difference of texture and color between sub-regions in the foreground area. Experiments show that SPCRF improves the classification accuracy of foreground effectively.2. Introduce a CRF based image understanding by embedding object confidence prior (OCPCRF). OCPCRF parses the image through embedding the object confidence context into the CRF framework. Particularly, OCPCRF adopts the one-to-many strategy to train the object classifier and compute the object confidence which is different with the traditional methods by simply statistical counting. Experiments show that OCPCRF improves the pixel labeling accuracy.3. Propose a CRF based weakly supervised scene image understanding by embedding alignment prior (APCRF). On the basis of Multi Image Model, APCRF mines the structure context through image registration algorithm to build the relation between super-pixels on two images and furtherly treats the relation as the pair-wise potential in the MIM. Experiments show that the usage of the registration prior improves the classification accuracy notably.