Showing papers by "Subhasis Chaudhuri published in 2012"

A Novel Domain Adaptation Bayesian Classifier for Updating Land-Cover Maps With Class Differences in Source and Target Domains

Abstract: This paper addresses the problem of land-cover map updating by classification of multitemporal remote-sensing images in the context of domain adaptation (DA). The basic assumptions behind the proposed approach are twofold. The first one is that training data (ground reference information) are available for one of the considered multitemporal acquisitions (source domain) whereas they are not for the other (target domain). The second one is that multitemporal acquisitions (i.e., target and source domains) may be characterized by different sets of classes. Unlike other approaches available in the literature, the proposed DA Bayesian classifier based on maximum a posteriori decision rule (DA-MAP) automatically identifies whether there exist differences between the set of classes in the target and source domains and properly handles these differences in the updating process. The proposed method was tested in different scenarios of increasing complexity related to multitemporal image classification. Experimental results on medium-resolution and very high resolution multitemporal remote-sensing data sets confirm the effectiveness and the reliability of the proposed DA-MAP classifier.

An Optimization-Based Approach to Fusion of Hyperspectral Images

Abstract: In this paper we propose a new approach for visualization-oriented fusion of hyperspectral image bands. The proposed technique has been devised to generate the fused image with a certain set of desired properties for a better visualization. The fusion technique should provide a resultant image with a high local contrast without driving individual pixels into over- or under-saturation. We focus on these desired properties of the resultant image, and formulate a multi-objective cost function for the same. We have shown how we can incorporate the constraint of spatial smoothness of the weight vectors, as opposed to the smoothness of the fused image. The solution of this optimization problem has been provided using the Euler-Lagrange technique. By using an appropriate auxiliary variable, we show how the constrained optimization problem can be converted into a computationally efficient unconstrained one. The effectiveness of the proposed technique is substantiated from the visual and quantitative results provided.

Haptic rendering of dense 3D point cloud data

Abstract: This work is aimed at rendering an object described by a dense point cloud data without a pre-computed polygonal mesh and surface normal information. We use a proxy based rendering technique with proxy collocated with the HIP in free space. To avoid sinking of proxy into the object during collision a sphere of sufficiently large radius is selected as the proxy so as to avoid it going through the point cloud. Once collision is detected, we minimize a cost function depending on the current HIP and proxy positions and find a new goal position for the proxy corresponding to the local minimum of the cost function. Our rendering algorithm continuously evaluates a tangential vector for the proxy to move over the object surface during collision. The penetration depth is calculated from the proxy to the HIP and is used to calculate the reaction force for the haptic device. We used our technique to render several dense point cloud based models. We also use the surface normal evaluated at the point of contact to shade the object surface when shown visually.

Haptic rendering of cultural heritage objects at different scales

Abstract: In this work, we address the issue of virtual representation of objects of cultural heritage for haptic interaction. Our main focus is to provide a haptic access of artistic objects of any physical scale to the differently abled people. This is a low-cost system and, in conjunction with a stereoscopic visual display, gives a better immersive experience even to the sighted persons. To achieve this, we propose a simple multilevel, proxy-based hapto-visual rendering technique for point cloud data which includes the much desired scalability feature which enables the users to change the scale of the objects adaptively during the haptic interaction. For the proposed haptic rendering technique the proxy updation loop runs at a rate 100 times faster than the required haptic updation frequency of 1KHz. We observe that this functionality augments very well to the realism of the experience.

Joint MAP estimation for blind deconvolution: when does it work?

Abstract: Blind deconvolution aims at reconstructing an image from its blurred and noisy version, when the blur kernel is not known. It has been acknowledged that the naive maximum aposteriori probability (MAP) algorithm favors a no-blur solution [3]. In [8] the failure of the direct MAP approach is addressed and it is proved that a simultaneous MAP estimation of the image and the point spread function (PSF) fails, providing a trivial solution. In contrast, we show that an appropriate choice of PSF prior during joint MAP estimation does provide a non-trivial solution. We provide the feasible range for the PSF regularization factor which would prevent a trivial solution.

Accelerating non-local denoising with a patch based dictionary

Abstract: Nonlocal means (NLM) image denoising algorithm is not feasible in many applications due to its high computational cost. High computational burden is due to the search of similar patches for each reference patch in the entire image. In this paper, we present a novel technique of preselecting and grouping the similar patches in the form of a dictionary and hence speeding up the computation of NLM denoising method. We build a dictionary only once, with a set of training images of all possible classes of objects, in which patches with similar photometric structures are clustered together. For each noisy patch, similar patches are searched in the global dictionary. In contrast with previous NLM speedup strategies, our dictionary building approach preclassifies similar patches with the same distance measure as used by NLM method. We achieve a substantial reduction in computational time than the original NLM method especially when search window of NLM is large, without much affecting the PSNR. The proposed algorithm is shown to outperform other prefiltering based fast NLM algorithms computationally as well as qualitatively.

Media for Distance Education

Abstract: Education can be thought of as the process of adding value to life. Apart from its fundamental goal of gathering knowledge, it has manifold long term objectives. One may say that education focuses at (i) value creation, (ii) economic viability, (iii) broadening of perspective, and (iv) enhanced social outreach. It should help an individual mold his/her personality, earn a living, which in turn accounts for his/her social health and lifestyle. It is hard to define the efficiency of educational systems from the value creation point of view since objective measures of value are not possible. In order to devise some means to quantify the efficiency of education, it would have to be treated more like a commodity rather than an abstract concept of knowledge. In this aspect, educational systems may be modeled as commodity markets whose performance measure is given by the throughput of the system. We define the throughput to be the average number of students graduating per year per instructor for a given curriculum.

Video Analysis and Repackaging for Distance Education

Abstract: At the confidence level, a classifier outputs a numerical value for each class indicating the belief or probability that the given input pattern belongs to that class. At the rank level, a classifier assigns a rank to each class with the highest rank being the first choice. Rank value cannot be used in isolation because the highest rank does not necessarily mean a high confidence in the classification. At the abstract level, a classifier only outputs a unique class label or several class labels (in which case, the classes are equally good). The confidence level conveys the richest information, while the abstract level contains the least amount of information about the decision being made. Some of the popular combination schemes are maximum, minimum, voting, mean, median [67], Dempster-Shafer, and Bayes belief integrals. We have used Bayesian belief integration technique for fusing the individual decisions. Methods of combining decisions like minimum, maximum, average and majority vote [67] do not take into account the errors that have occurred in a segment. They just consider the output labels. But in the Bayesian formulation, errors for each segment are captured by a matrix called confusion matrix [162], represented as PTk = ⎛ ⎜⎜⎜⎝ n 11 n (k) 12 · · · n 1M n 21 n (k) 22 · · · n 2M .. . . . .. n M1 n (k) M2 · · · n MM ⎞ ⎟⎟⎟⎠ k = 1,2 . . .K, where each row i, corresponds to class Ci, and each column j corresponds to an event sk(x) = j, meaning that in segment sk, bit x is given a label j. Thus n i j means that n (k) i j bits of class Ci have been assigned a label j in segment k. Thus the total number of samples/bits in a segment are N(k) = M ∑ i=1 M ∑ j=1 n i j (7.7) in which the number of bits in each class Ci are, n i = M ∑ j=1 n i j , i = 1 . . .M (7.8) and the number of samples that are assigned a label j in sk are n j = M ∑ i=1 n i j , j = 1 . . .M. (7.9) If a segment sk has an error, then the corresponding event sk(x) = j will have certain uncertainty. With the help of the corresponding confusion matrix, such an uncertainty is expressed with a probability 7.5 Illustrative results 119 P(x ∈Ci|sk(x) = j) = n i j n j , i = 1 . . .M. (7.10) This is also known as the belief in a proposition bel(x ∈Ci|sk(x) = j) = P(x ∈Ci|sk(x) = j), i = 1 . . .M. (7.11) With K segments, s1, . . .sK , we have K confusion matrices, PT1, . . .PTK . When we consider the same bit in all these segments, then we have K events, sk = jk, k = 1 . . .K. It means that in each of the segments, the same bit can have a different class label. Using the above definition, each event along with the corresponding confusion matrix will give a belief value for each class. To arrive at a single label we need to combine these individual decisions. These decisions can be fused using the Bayesian formulation. By using a standard naı̈ve Bayes formulation, and by assuming that events s1(x) = j1, . . .sk(x) = jk are independent, an expression for belief can be derived [162]. It may be given by bel(i) = η K ∏ k=1 P(x ∈Ci|sk(x) = jk) (7.12) with η being a constant that ensures, ∑i=1 bel(i) = 1 which is calculated as

Matte based generation of land cover maps

Abstract: A novel supervised technique for the generation of spatially consistent land cover maps based on class-matting is presented in this paper. This method takes advantage of both standard supervised classification technique and natural image matting. It adaptively exploits the spatial contextual information contained in the neighborhood of each pixel through the use of image matting to reduce the incongruence inherent in pixel-wise, radiometric classification of multi-spectral remote sensing data, providing a more spatially homogeneous land-cover map besides yielding a better accuracy. In order to make image matting possible for N-class land cover map generation, we extend the basic alpha matting problem into N independent matting problems, each conforming to one particular class. The user input required for the alpha matting algorithm in terms of initially identifying a few sample regions belonging to a particular class (known as the foreground object in matting) is obtained automatically using the supervised ML classifier. Experimental results obtained on multispectral data sets confirm the effectiveness of the proposed system.

Fast computation of edge model representation for image sequence super-resolution

Abstract: Edge model based representation of Laplacian subbands has been demonstrated to be useful in single frame high resolution image generation. A reconstruction based multiframe super-resolution algorithm yields a better super-resolved image if high resolution estimate of individual frame is given as input, instead of original low resolution frames. Fast computation of edge-model based representation enables fast single frame high resolution image generation for multiple frames and in turn helps in speeding up reconstruction based super resolution. In the present work, efficient multiframe edge model computation is achieved by computing edge model for the reference frame and then computing successive models by adapting it on the remaining frames.

PCA based video denoising in a non-local means framework

Abstract: Nonlocal means (NLM) video denoising algorithm though provide very competitive results, suffer from high computational cost. We propose to reduce the computations through the concept of dimensionality reduction using principle component analysis (PCA). Image neighbourhood representations are projected onto a lower dimensional subspace determined by PCA and weights are computed in this reduced subspace. Principle components are computed globally for an entire video shot having similar frames, which reduces computations drastically. We have used a technique of histogram difference to group the frames with similar visual content. We have achieved an improvement in accuracy in addition to reducing the computation. The proposed method is shown to outperform all other nonlocal means related video denoising methods.

Haptic rendering of variable density point cloud through local kernel bandwidth estimation

Abstract: Haptic rendering of a point cloud data without a precomputed mesh structure is always a difficult problem. This paper tries to solve the problem of real time rendering a variable density 3D point cloud data as an optimization problem with a local kernel bandwidth estimation. In order to avoid the ambiguity of deciding the direction of reaction force while rendering the data we prefer a proxy based renderer. To avoid the proxy sinking into the object during collision the proxy point is always kept at minimum distance away from the object surface approximated by the point cloud. This minimum distance is computed during the rendering process depending on the local density of the points in the data using the kernel bandwidth estimation. Once collision is detected, we minimize a cost function depending on the current haptic interaction point (HIP) and proxy positions and find a new goal position for the proxy corresponding to the local minimum of the cost function. We validate the proposed technique by comparing the rendered force with the reaction force computed using a known spherical object.