Remote sensing image denoising faces many challenges since a remote sensing image usually covers a wide area and thus contains complex contents. Using the patch-based statistical characteristics is a flexible method to improve the denoising performance. There are usually two kinds of statistical characteristics available: interior and exterior characteristics. Different statistical characteristics have their own strengths to restore specific image contents. Combining different statistical characteristics to use their strengths together may have the potential to improve denoising results. This work proposes a method combining statistical characteristics to adaptively select statistical characteristics for different image contents. The proposed approach is implemented through a new characteristics selection criterion learned over training data. Moreover, with the proposed combination method, this work develops a denoising algorithm for remote sensing images. Experimental results show that our method can make full use of the advantages of interior and exterior characteristics for different image contents and thus improve the denoising performance.
Analyzing composite behaviors involving objects from multiple categories in surveillance videos is a challenging task due to the complicated relationships among human and objects. This paper presents a novel behavior analysis framework using a hierarchical dynamic Bayesian network (DBN) for video surveillance systems. The model is built for extracting objects' behaviors and their relationships by representing behaviors using spatial-temporal characteristics. The recognition of object behaviors is processed by the DBN at multiple levels: features of objects at low level, objects and their relationships at middle level, and event at high level, where event refers to behaviors of a single type object as well as behaviors consisting of several types of objects such as "a person getting in a car." Furthermore, to reduce the complexity, a simple model selection criterion is addressed, by which the appropriated model is picked out from a pool of candidate models. Experiments are shown to demonstrate that the proposed framework could efficiently recognize and semantically describe composite object and human activities in surveillance videos.
Fields of experts (FoE) image denoising is one of the most promising high-order Markov random field (MRF)-based image denoising methods. However, the original algorithm by Roth and Black did not consider the parameter selection problem in its iteration, so it cannot be directly applied to real image denoising tasks. An automatic stopping criterion in FoE image denoising is introduced, through which the denoised image can be obtained without reference image and noise variance estimation. Experimental results validate its better performance than the classic FoE method, both on synthetic and real noisy images.
We propose to fuse an image's local and global information for scene classification. First, the image's local information is represented by context information exploited using spatial pyramid matching. Images are segmented to patches by a regular grid, and scale invariant feature transform (SIFT) features are extracted. All the patch features are clustered and quantified to get visual words. The visual word pair and visual word triplet are neighboring and different visual words. By an analogy between image pixel space and patch space, we also get visual word groups, which are the continuous occurrence of the same visual words. The spatial envelope is employed for extracting an image's global information. The spatial envelope is a holistic description of the scene, where local information is not taken into account. Finally, a stacked-support vector machine (SVM) fusion method is used to get the scene classification results. Experimented with three benchmark data sets, the results demonstrated that our methods could get better results than most popular scene classification methods presented in recent years.
Superresolution reconstruction produces a high-resolution image from a set of low-resolution images. Accurate subpixel image registration is critical in image superresolution reconstruction. The existence of outliers, which are defined as data points with different distributional characteristics from the assumed model, will produce erroneous image registration estimates that lead to undesirable results. Several solutions have been proposed to handle registration errors as a part of the regularized solution in the reconstruction step; however, they are invalid for videos that contain localized outliers, such as moving objects in the frames. We present a new robust image superresolution method to handle the localized motion outliers. We first separate the low-resolution image into several layers. After identifying the motion models of the layers, we calculate these separately. Then, we can obtain an accurate subpixel image registration of the background that contains important information. Finally, we fuse them into a high-resolution image. The effectiveness of our model is demonstrated with results from superresolution experiments with both synthetic and real sequences in the presence of localized motion outliers.
This paper presents a novel multiresolution approach for texture classification based on the moment features, which are extracted in the x direction and y direction, of a histogram in each image resolution. Then we propose a weighted multiresolution moment feature for supervised texture classification. Moreover, because of their capability of expressing image information using moment methods, these features achieved better performance on rotation invariance and noise robustness than most of popular texture features presented in recent years.
In this paper, a method to detect whether the behavior of a single person in video sequence is abnormal is proposed.
Firstly, after the pre-processing, the background model is gotten based on the Mixture Gaussian Model(GMM), at the
same time the shadow is eliminated; then use the color-shape information and the Random Hough Transform (RHT) to
abstract the zebra crossing and segment the background; Lastly, we use the rectangle and the centroid to judge whether
the person's behavior is abnormal.
Segment-based stereo matching algorithms are not able to deal with the difficulty that disparity boundaries appear inside the initial color segments. To solve this problem, we propose a novel algorithm that segments the reference image by combining color and depth segmentation information. Then we construct the energy function in segment domain, which embodies the smoothness and visibility constraints to penalize the discontinuities of edge pixels and potential occluded regions, respectively. Finally, the optimal disparity plane labeling is approximated by applying loopy belief propagation. Experimental results on the benchmark images demonstrate that our algorithm is comparable to the state-of-the-art algorithms.
We present a new image segmentation algorithm based on Markov random fields (MRF) with adaptive neighborhood (AN) systems. First, a new criterion function is developed to adaptively select the neighborhood system of the MRF. Second, thanks to the appropriate format of the new criterion function, a general iterated conditional mode (GICM) is deduced to incorporate the AN selection into the inference process in the segmentation of images. Third, an end points detection step is introduced to preserve the ends of line structures in the images. The proposed algorithm has the following advantages over previous works: the convergence of the new algorithm is much faster than classical algorithms; the AN configuration of each site is iteratively optimized in the inference process, which leads to more accurate segmentation results; and no extra knowledge is needed in the AN selection. Numerical experiments on a wide range of images demonstrate that the proposed image segmentation algorithm performs better, in terms of speed and detail preserving, than the algorithms based on classical MRFs.
Motion detection and tracking are both fundamental steps in video surveillance. This paper describes two improved algorithms for motion detection and tracking, respectively. Firstly, the shadow and ghost detection processing is fused with updating the background subtraction model, after the frame is transformed to HSV color space from RGB color space in order to reduce the effect of illumination changes, shadows, and ghosts. And the result of detection is used as initialization for tracking. Secondly, the gradient field is combined with region information to locate the boundary of the object accurately, instead of the traditional level-set method, which only utilizes the gradient field to propagate a front. Experimental results show that the algorithms decrease the computational complexity and provide accurate location of the object boundary.
A novel algorithm for shape detection based on mathematical morphology is presented. Two stages are involved. In the first stage, a shape model is learned automatically from learning examples belonging to the same object class. It is a collection of subparts with the description of relations among subparts, represented by a fuzzy graph. In the second stage, the generated model is used to detect similar shapes from images of complex real scenes. Subparts of the shape are detected in sequence based on their saliency, and then the geometric configuration among those detected subparts is checked. A morphological component detector is proposed to detect each subpart by using a soft structuring element, derived from the shape model. Satisfactory results are shown when testing the algorithm on synthetic and real images.
In this paper, we propose a novel and automated line extraction algorithm in multi-spectral images, which fully utilize the complementary information among multi-spectral images. It consists of three main aspects: Firstly, edges are extracted from every spectral image. Then, the edge points from all spectral images are grouped into combined line-support regions according to certain fusion rules. Finally, fits the regions and generates the fused lines. The new algorithm is applied to some real multi-spectral images. The experimental results show that the new algorithm is effective.
This paper presents a line extraction algorithm in SAR (Synthetic Aperture Radar) images. The algorithm is designed based on the statistical characteristics of the speckle in SAR image. Three steps are involved. Firstly, a new edge detector, which combines the Canny operator and Ratio operator, is used to detect the edge points and calculate their directions, then the edge points are grouped according to their edge direction to form the initial lines. Finally, a high-level grouping step connects the fragmental lines. The proposed new edge operator is CFAR (Constant False Alert Rate) and prevents the line from cleavage. The algorithm has been applied in the X-band airborne SAR images, and the results are presented at the end of this paper.
The process of relative radiometric calibration (RRC) is an important step for detecting change and monitoring environment through analyzing multi-temporal satellite images. Two key issues that focus on the RRC are how to extract invariant targets that have little or no variation in their reflectance between two images and how to acquire a linear function that expresses the relationship between the digital counts (DNs) of the two images. In this paper, an automatic method for selecting invariant targets which cover the range of bright, midrange, and dark data values is developed, and a robust estimator, which can be effective in tolerating with up to 50 percent outlier contamination, for calculating the gain and the offset of the linear function is described. The proposed methods are automatic and robust. We have applied the proposed method experimentally to synthesized images and real SPOT images, and experiment results have shown the feasibility of our algorithms.
Due to the complex signal-dependent nature of the speckle in SAR image, it is more reasonable to use different speckle descriptions and despeckling filters for different kinds of regions. A multi-description despeckling approach is presented in this paper. Based on the local statistics, the x2 test is introduced to segment the SAR image into the Gamma-distributed homogeneous regions and the more fluctuant heterogeneous regions. Then a MAP filter is used in the homogeneous regions, and a modified median filter is utilized in the heterogeneous regions. X-band airborne SAR image and synthetic image are used for illustration and comparison.
Multi-sensor information fusion plays an important pole in object recognition and many other application fields. Fusion performance is tightly depended on the fusion level selected and the approach used. Feature level fusion is a potential and difficult fusion level though there might be mainly three fusion levels. Two schemes are developed for key issues of feature level fusion in this paper. In feature selecting, a normal method developed is to analyze the mutual relationship among the features that can be used, and to be applied to order features. In object recognition, a multi-level recognition scheme is developed, whose procedure can be controlled and updated by analyzing the decision result obtained in order to achieve a final reliable result. The new approach is applied to recognize work-piece objects with twelve classes in optical images and open-country objects with four classes based on infrared image sequence and MMW radar. Experimental results are satisfied.
Face recognition has wider application fields. In recurrent references, most of the algorithms are with simple background in the static images, and only used for ID picture recognition. It is necessary to study the whole process of multi-pose face recognition in a complex background. In this paper an automatic multi-pose face recognition system with multi-feature is presented. It consists of several steps as following: face detection based on skin-color and multi-verification, detection and location of the face organs based on the combination of an iterative search by a confidence function and template matching at the candidate points with the analysis of multiple features and their projections, feature extraction for recognition, and recognition decision based on a hierarchical face model with the division of the face poses. The new approach is applied to 420 color images that contain multi-pose faces with two visible eyes in a complex background, and the results are satisfactory.
In this paper, a new method for extracting ellipses from an image is presented. It includes two steps, finding the candidate regions and rectifying the detected regions' position, estimating the ellipses' parameters. It's calculation amount is less than many other methods, and it is very immune to the noise. Specially, it has strong ability to extract multi incomplete and concentric ellipses from an image.
The paper presents a new approach to match an image to a model described by straight lines. It utilizes both straight lines and multi-properties of the line hierarchically, and works efficiently under arbitrary translations, rotations and scale changes in noise conditions. It is practical for many applications.