A Bilevel Contextual MRF Model for Supervised Classification of High Spatial Resolution Remote Sensing Images

By labelling high spatial resolution (HSR) images with specific semantic classes according to geographical properties, scene classification has been proven to be an effective method for HSR remote sensing image semantic interpretation. Deep learning is widely applied in HSR remote sensing scene classification. Most of the scene classification methods based on deep learning assume that the training datasets and the test datasets come from the same datasets or obey similar feature distributions. However, in practical application scenarios, it is difficult to guarantee this assumption. For new datasets, it is time-consuming and labor-intensive to repeat data annotation and network design. The neural architecture search (NAS) can automate the process of redesigning the baseline network. However, traditional NAS lacks the generalization ability to different settings and tasks. In this paper, a novel neural network search architecture framework—the spatial generalization neural architecture search (SGNAS) framework—is proposed. This model applies the NAS of spatial generalization to cross-domain scene classification of HSR images to bridge the domain gap. The proposed SGNAS can automatically search the architecture suitable for HSR image scene classification and possesses network design principles similar to the manually designed networks. To obtain a simple and low-dimensional search space, the traditional NAS search space was optimized and the human-the-loop method was used. To extend the optimized search space to different tasks, the search space was generalized. The experimental results demonstrate that the network searched by the SGNAS framework with good generalization ability displays its effectiveness for cross-domain scene classification of HSR images, both in accuracy and time efficiency.

A novel fuzzy decision tree is proposed in this paper (the FDT-support vector machine (SVM) classifier), where the node discriminations are implemented via binary SVMs. The tree structure is determined via a class grouping algorithm, which forms the groups of classes to be separated at each internal node, based on the degree of fuzzy confusion between the classes. In addition, effective feature selection is incorporated within the tree building process, selecting suitable feature subsets required for the node discriminations individually. FDT-SVM exhibits a number of attractive merits such as enhanced classification accuracy, interpretable hierarchy, and low model complexity. Furthermore, it provides hierarchical image segmentation and has reasonably low computational and data storage demands. Our approach is tested on two different tasks: natural forest classification using a QuickBird multispectral image and urban classification using hyperspectral data. Exhaustive experimental investigation demonstrates that FDT-SVM is favorably compared with six existing methods, including traditional multiclass SVMs and SVM-based binary hierarchical trees. Comparative analysis is carried out in terms of testing rates, architecture complexity, and computational times required for the operative phase.

For decades, Markov random fields (MRF) have been used by statistical physicists to explain various phenomena occurring among neighboring particles because of their ability to describe local interactions between them. In Winkler (1995) and Bremaud (1999), an MRF model is used to explain why neighboring particles are more likely to rotate in the same direction (clockwise or counterclockwise) or why intensity values of adjacent pixels of an image are more likely to be the same than different values. This model is called the Ising model. There are a large number of problems that can be modeled using the Ising model and where an MRF model can be used. Basically, an MRF model is a spatial-domain extension of a temporal Markov chain where an event at the current time instant depends only on events of a few previous time instants. In MRF, the statistical dependence is defined over the neighborhood system, a collection of neighbors, rather than past events as in the Markov chain model. It is obvious that this type of spatial dependence is a common phenomenon in various signal types including images. In general, images are smooth and, therefore, the intensity values of neighboring pixels are highly dependent on each other. Because of its highly theoretical and complex nature, and intensive computational requirements, practical uses of MRF were extremely limited until recently. Due to the dramatic improvements in computer technologies, MRF modeling has become more feasible for numerous applications, for example, image analysis because many image properties, such as texture, seem to fit an MRF model, i.e., intensity values of neighboring pixels of images are known to be highly correlated with each other. The Markovian nature of these textural properties has long been recognized in the image processing community, and has widely been used in a variety of applications (e.g. image compression and image noise removal). However, these applications were limited to empirical studies and were not based on a statistical model such as the MRF model.

The accurate discrimination of distinct thematic classes using classification techniques developed for medium/low resolution images is not effective when apply to very high spatial resolution (HR) data (e.g. Quickbird, IKONOS) due to the spatial heterogeneity issue. In this paper, Markov random field (MRF) models, which are useful tools for integrating contextual (considering spatial dependence within and between pixels) information into classification process is used to model spatial heterogeneity for improving the classification accuracy. Two novel MRF approaches are evaluated using a Quickbird HR image covers Xixi National Wetland Park, Hangzhou, China. The experimental results show this method is effective to exact segmentation of land boundaries and suppress classification noises. In addition, the improved MRF models outperform than conventional method in terms of classification accuracy and time-efficiency.

Abstract. For high spatial resolution Remote Sensing images, it is very important to investigate the transformational methods between background and target characteristics. Only in this way rich details in images under different imaging conditions can be well extracted. Amongst the characteristics of imagery targets, texture is a visual feature that reflects the homogeneity of images and the inner attributes of different objects. What's more, it includes important information which describes the structural arrangement of objects and the connection with the surrounding environment. This paper regards texture as the major feature and investigates the transformational methods of texture feature description under different imaging conditions. This paper mainly consists of three parts:(1) Construct a wavelet filter based on Gabor wavelet, which describes texture features obtained under different imaging conditions;(2) Process and analyze the different object's texture features jointly by the relationship which is built by the wavelet description;(3) Build the transformation between the wavelet descriptions of the different object's texture features based on the characteristics of the band and direction.

In this paper, we propose a sequence labeling method by using SVM posterior probabilities with a Markov Random Field (MRF) model for colorectal Narrow Band Imaging (NBI) zoom-videoendoscope. Classifying each frame of a video sequence by SVM classifiers independently leads to an output sequence which is unstable and hard to understand by endoscopists. To make it more stable and readable, we use an MRF model to label the sequence of posterior probabilities. In addition, we introduce class asymmetry for the NBI images in order to keep and enhance frames where there is a possibility that cancers might have been detected. Experimental results with NBI video sequences demonstrate that the proposed MRF model with class asymmetry performs much better than a model without asymmetry.

The present paper is concerned with the study of parallel approaches for the relaxation algorithms associated to Markov Random Fields (MRF) models in image analysis, [2]. MRF-based relaxation algorithms are known to be intrinsically parallel and massive parallelizations by updating at once independent variables of the field have been early described, [4]. We present here a different approach based on relaxation algorithms running in parallel at different and interacting periodically. The proposed approach is based on parallel Markov chains which have been studied by Aarts et al., [l], and Graffigne, [3], for global optimization using simulated annealing. In our method, parallel Markov chains are coupled with a new multiscale exploration of the configuration space for MRF. This multiscale exploration can be implemented efficiently by considering a sequence of multiresolution MRF models whose parameters and neighborhood structures are obtained from the original MRF model. The designed algorithm is suited for MIMD computers and can be outlined as follows. First MRF models at different scales are computed from the original MRF model. The relaxations at different scales are stochastic and cooperative : every p iterations - one iteration corresponding to a full sweep on the image - a processor attempts to transfer local estimates to the processor controlling the next finer scale. With the low resolution levels of the hierarchy are associated high temperatures in the stochastic relaxation, [3]. A high temperature allows to escape from local minima of the energy function. With the intermediate resolution levels are associated lower temperatures. At these levels, the relaxation process becomes more sensitive to local minima and visits the large or medium size valleys of the energy landscape. At the finest resolution level deterministic relaxation is adopted : the estimation is ultimately refined at that level. The proposed parallelization approach is compared to sequential deterministic and stochastic relaxation in the context of optical flow estimation. The multiscale algorithm exhibits fast convergence properties, comparable to multigrid deterministic relaxation, and the obtained estimates are close to the near optimal solutions obtained by time consuming sequential stochastic methods.

Scale selection for anisotropic diffusion filter by Markov random field model

In order to overcome the poor anti-noise performance of traditional fuzzy C-Means (FCM) algorithm in image segmentation, a novel improved FCM algorithm was proposed in this paper based on Particle Swarm Optimization (PSO) algorithm and Markov Random Field (MRF) model, which can make full use of the global searching ability of PSO and the spatial information integrating ability of MRF for image segmentation. In this algorithm, the image segmentation is converted to a PSO optimization problem, in which the fitness function is set up to containing the spatial information based on the spectral value and the neighboring pixels modeled by MRFs. And segmentation results can be iteratively obtained during the PSO iterations according to the newly designed membership function of FCM in which the spatial information is integrated. The experiments herein reported in this paper illustrate the better performance of this algorithm than the traditional FCM algorithm and the PSO algorithm for image segmentation.

This paper presents a novel texture segmentation method using neural networks and a Markov random field (MRF) model. Multi-scale wavelet coefficients are used as input for the neural networks. The output of the neural network is modeled as a posterior probability. Initially, the multi-scale texture segmentation is performed by the posterior probabilities from the neural networks and MAP (maximum a posterior) classification. Then the MAP segmentation maps are produced at all scales. In order to obtain the more improved segmentation result at the finest scale, our proposed method fuses the multi-scale MAP segmentations sequentially from coarse to fine scales. This is done by computing the MAP segmentation given the segmentation map at one scale and a priori knowledge regarding contextual information which is extracted from the adjacent coarser scale segmentation. In this fusion process, the MRF prior distribution and Gibbs sampler are used, where the MRF model serves as the smoothness constraint and the Gibbs sampler acts as the MAP classifier.

We present a new method using the Bayesian approach and a Markov random field (MRF) model of integrating several low level visual modules. Using this approach, we have shown how results from an edge based stereo module can be integrated with an intensity based stereo algorithm. In another example, results from a shape from shading module are combined with intensity based stereo. We first derive the intensity based stereo algorithm using the MRF model. The integration is then performed by coupling the results from other modules to the energy functional of the MRF associated with the intensity based stereo. The maximum a prosteriori (MAP) estimate of the resulting MRF is obtained using the mean field annealing algorithm. Results from real and artificial images show a consistent improvement in the accuracy after integration.

Image segmentation method combined with Markov random field (MRF) model and regional growth image in this paper is proposed to achieve the accuracy of sonar image segmentation. First of all, the characteristics of field and the label field of sonar image is established through the MRF model; secondly, the further image segmentation is processed by region growing based on the initial image segmentation by ICM algorithm, and this reduced segmentation plaques caused by the MRF method on image segmentation. Finally, experimental results show that the method in this paper can obtain the image segmentation effectively.

Markov Random Field (MRF) theory has a significant potential role in image segmentation field. It uses (pixels, regions, edges)-based on MRF models to detect objects, boundaries and other relevant information in an image. This paper proposes an extension of Unified Markov Random Field (UMRF) model to include edge-based features. Firstly, the proposed technique employs the likelihood function to combine the advantages of the pixel-based, region-based and edge-based MRF model, by computing the product of the pixel likelihood function, regional likelihood function and edge likelihood function. Secondly, the region-based macro texture features are extracted using the UMRF model. Then the edge-based features are extracted using the maximum gradient method to recover any significant lost information. A principled probabilistic inference is implemented to integrate various types of likelihood information and spatial constraints by iteratively updating the posterior probability of the proposed model. The segmentation process is completed when the iterations converge. The proposed enhanced UMRF technique which combines pixel-based, region-based and edge-based features achieved a higher skin lesion segmentation accuracy than MRF model which combines pixel-based and region-based only.

A valid multiscale classification method of synthetic aperture radar (SAR) imagery is proposed based on multiscale technology and Markov random field (MRF) mode. Firstly, we employ multiscale autoregressive model for extracting the feature of SAR image. which is modeled by Markov random field (MRF) Model that relies on the Gaussian distribution. Secondly, using the joint probability distribution in terms of an energy function, estimation of parameters can be performed by the stochastic relaxation algorithm. Then the maximum posteriori (MAP) is designed as the optimal criterion and the final labels are obtained by the simulated annealing algorithm. Experimental results show that this method is accurate, efficient and robust.

We present a new approach for the discriminative training of continuous-valued Markov Random Field (MRF) model parameters. In our approach we train the MRF model by optimizing the parameters so that the minimum energy solution of the model is as similar as possible to the ground-truth. This leads to parameters which are directly optimized to increase the quality of the MAP estimates during inference. Our proposed technique allows us to develop a framework that is flexible and intuitively easy to understand and implement, which makes it an attractive alternative to learn the parameters of a continuous-valued MRF model. We demonstrate the effectiveness of our technique by applying it to the problems of image denoising and in-painting using the Field of Experts model. In our experiments, the performance of our system compares favourably to the Field of Experts model trained using contrastive divergence when applied to the denoising and in-painting tasks.