Study On Contextual Image Classification And Texture Segmentation Based On Markov Random Field

Image classification is a very critical aspect for image interpretation process. The results of classification, which is more correct and accurate, can greatly facilitate the follow-up works, such as target identification, urban planning, environmental monitoring and land utilization etc. Research on image classification should consider not only the feature information of pixels itself, but also the related information between adjacent pixels. Markov random field (MRF) method is widely used for the reason that it provides a convenient and consistent modeling method to the correlation of spatial context[1]. In recent years, MRF method becomes a hot research neighborhood in image analysis and computer vision because it can well characterize the feature of spatial context correlation of the local image.This paper discusses a series of theoretical and practical issues relating to MRF, including:MRF-Gibbs distribution equivalence problem, MAP-MRF framework construction problems, energy minimization problem, observation field modeling problem, Label field MRF modeling problem, wavelet multi-scale feature extraction sequence problem, accuracy assessment algorithm problems etc.Around the theoretical issues, firstly, contextual MRF image classification model is established based on the MAP-MRF framework in the original resolution scale; Secondly, in order to further describe the image of the non-stationary characteristics and observed field characteristics, use of multi-scale wavelet transform characteristics based on contextual MRF image classification model, the observation field is built on a series of multi-scale wavelet, the corresponding resolution label field take advantage of the characteristics of the scale field in order to establish a multi-scale contextual MRF image classification model.Based on contextual MRF and multi-scale contextual MRF model, this paper sets up five experiments from two aspects, Landsat TM remote sensing image classification and synthesis texture image segmentation from Brodatz texture library. Analysis of experimental results, the writer obtains the following important experiment conclusions.(1)From the experiments of different initial label on the southern region of Dali Erhai 5Basin:Different initial label has some influence for the results of image classification, the overall accuracy of FCM+contextual MRF model is84.33%while Rand+contextual MRF model is up to93.67%; Classification accuracy of random label further illustrate the advantages of MRF spatial context in image analysis.(2)From the experiments of different energy weight parameters on the Hongkong Tai Po area:When the weight parameter λ is small, label field is dominant, so that the classification results focus on the area of the local properties; On the other hand, the observation field is dominant so that the classification results focus on the area of consistent properties. When λ=0.9～1.2, the classification model can achieve the maximum classification accuracy of93.07%.(3)From the experiments of different clique potential parameters on the Lijiang Lashihai region:Smaller clique potential parameters β corresponding classification result is more details information, but as much as possible to keep the real location of the boundaries; With the β increasing, the classification results details gradually reduced, the regional better; and when β is too large, the classification results boundary deformation, between classes can not distinguish more.(4)By the segmentation comparative experiments in synthetic texture binary and three-valued image:The contextual MRF model segmentation not only consider gray characteristics of pixels, but also consider the contextual information of pixels, so that the segmentation results are much better than FCM; Multi-scale contextual MRF model segmentation makes use of the contextual information, while the original resolution texture image wavelet multi-scale sequence expression, so that the segmentation results become ideal, by visual inspection segmentation accuracy can reach100%.In this paper, the practice and experimental data of MRF modeling, can be used to carry out the research and application of image analysis, such as:remote sensing data classification, synthetic aperture radar signal processing, computer vision problems (such as stereo matching) can provide important references.