M Tech Dissertations

Permanent URI for this collectionhttp://ir.daiict.ac.in/handle/123456789/3

Browse

Has files: YesSubject: search.filters.subject.Image reconstruction

Search Results

Now showing 1 - 6 of 6
  • Thumbnail Image
    ItemOpen Access
    Practical approach for depth estimation and image restoration using defocus cue
    (Dhirubhai Ambani Institute of Information and Communication Technology, 2011) Ranipa, Keyur R.; Joshi, Manjunath V.
    Reconstruction of depth from 2D images is an important research issue in computer vision. Depth from defocus (DFD) technique uses space varying blurring of an image as a cue in reconstructing the 3D structure of a scene. In this thesis we explore the regularization based approach for simultaneous estimation of depth and image restoration from defocused observations. We are given two defocused observations of a scene that are captured with di erent camera parameters. Our method consists of two steps. First we obtain the initial estimates for the depth as well as for the focused image. In the second step we re ne the solution by using a fast optimization technique. Here we use the classic depth recovery method due to Subbarao for ob- taining the initial depth map and Wiener lter approach for initial image restoration. Since the problem we are solving is ill-posed and does not yield unique solution, it is necessary to regularize the solution by imposing additional constraint to restrict the solution space. The regularization is performed by imposing smoothness constraint only. However, for preserving the depth and image intensity discontinuities, they are indenti ed prior to the minimization process from initial estimates of the depth map and the restored image. The nal solution is obtained by using computationally e client gradient descent algorithm, thus avoiding the need for computationally taxing algorithms. The depth as well as intensity edge details of the nal solution correspond to those obtained using the initial estimates. The experimental results indicate that the quality of the restored image is found to be satisfactory even under severe space-varying blur conditions.
  • Thumbnail Image
    ItemOpen Access
    Super-resolution of hyperspectral images
    (Dhirubhai Ambani Institute of Information and Communication Technology, 2011) Bhimani, Amitkumar H.; Joshi, Manjunath V.
    Hyperspectral (HS) images are used for space areal application, target detection and remote sensing application. HS images are very rich in spectral resolution but at a cost of spatial resolution. HS images generated by airborne sensors like the NASA’s Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) from satellites like NASA’s Hyperion. We proposed a principal component analysis (PCA) based learning method to increase a spatial resolution of HS images. For spatial resolution enhancement of HS images we need to employ a technique to increase the resolution. We used PCA based approach by learning the details from database which consist of high spatial resolution satellite images. Super-resolution, is an ill-posed problem, and does not result to unique solution, and therefore it is necessary to regularize the solution by imposing some additional constraint to restrict the solution space. To reduce the computational complexity, minimization of the regularized cost function is done using the iterative gradient descent algorithm. In this report the effectiveness of proposed scheme is demonstrated by conducting experiments on both Multispectral (MS) and Hyperspectral real data. The HS and MS images of AVIRIS and Digital airborne Imaging spectrometer (DAIS) respectively used as input for super resolution (SR).
  • Thumbnail Image
    ItemOpen Access
    New learning based super resolution using contourlet transform
    (Dhirubhai Ambani Institute of Information and Communication Technology, 2009) Singh, Vineet P.; Joshi, Manjunath V.
    new learning based super-resolution reconstruction using contourlet transforms is proposed. contourlet transform provides high degree of directionality. It captures geometrical smoothness along multiple directions and learns the edges present in an image normal to the contour. For learning purpose, training set of low resolution (LR) and high resolution (HR) images, all captured using the same camera, are used. Here two and three level contourlet decomposition for LR images (test image and training image dataset) and HR training images respectively. The comparison of contourlet coeffcients of LR test image from the LR training set using minimum absolute difference (MAD) criterion to obtain the best match contourlet coeffcient. The finer details of test image are learned from the high resolution contourlet coefficients of the training data set. The inverse contourlet transform gives super resolved image corresponding to the test image.
  • Thumbnail Image
    ItemOpen Access
    Practical approaches for photometric stereo
    (Dhirubhai Ambani Institute of Information and Communication Technology, 2007) Sharma, Swati; Joshi, Manjunath V.
    In this thesis work, we aim to propose approaches for photometric stereo that are less time consuming and have low computational requirements. Many applications of computer vision require high resolution 3-D structure of the object and high resolution image in order to take better decisions. However, in practical scenario, the major requirement is to obtain these within the time limit specified by the application. In this thesis, we first consider the problem of estimating the high resolution surface gradients, albedo and the intensity field using photometric stereo. Assuming a Lambertian model, the surface gradients and the albedo are estimated using Least Square (LS), Constrained Least Square (CLS) and Total Least Square (TLS) approaches. High resolution depth map as well as image is then obtained using generalized interpolation. These methods are computationally less taxing and hence fast. A comparison of these methods with a regularization based approach is presented, which is computationally very expensive.

    We also propose a simple approach for estimating the light source position (which is generally assumed to be known) from a single image. The proposed method is based on the shading information in the images and does not require any calibration.

    Next we formulate a suitable regularization scheme for simultaneously estimating light source position, surface gradients and albedo given a number of images of a stationary object captured from a stationary camera by changing the light source position. The optimization is carried out using simulated annealing and graph cuts minimization techniques, to get better estimates for the illuminant position, surface gradients and albedo. Although simulated annealing guarantees global minima for any arbitrary energy function, it takes hours for convergence. This makes it inappropriate in practical situations. So, we also look into the use of graph cuts minimization method which converges very fast. We present a comparison of the performance of simulated annealing and graph cuts optimization methods.

    Our first approach for obtaining high resolution depth and image using LS, CLS and TLS minimization methods with generalized interpolation, does not use any regularization which becomes essential while estimating high resolution image intensities and depth values, which is an ill-posed problem. Finally, we estimate the superresolved image and depth map using photometric cue and graph cuts optimization using a discontinuity preserving prior. Our results show that the high resolution images reconstructed using our approach that uses graph cuts minimization are much superior as compared to the approaches that use general interpolation methods.

  • Thumbnail Image
    ItemOpen Access
    Active contours in action
    (Dhirubhai Ambani Institute of Information and Communication Technology, 2005) Shah, Pratik P.; Banerjee, Asim
    There was considerable success in converting images into something like line drawings without resorting to any but the most general prior knowledge about smoothness and continuity. That led to the problem of “grouping” together the lines belonging to each object which is difficult in principle and very demanding of computing. Two terms that describes this problem in image processing tasks are edge detection and segmentation. Active contour models are proven to be very effective tools for image segmentation. The popularity of this semiautomatic approach may be attributed to its ability to aid segmentation process with apriori knowledge and user interaction. For more detailed application domain study for active contours, problem of converting a frontal photograph into a line drawing is taken up along with lip tracking based on Gradient Vector Flow force field (GVF) active contours. In images with gaussian and salt-pepper noise, segmentation process becomes difficult for gradient based methods. This work gives a solution to this problem. A novel break n’ join technique is presented and simulated for various images ranging from synthetic to real with convex and concave regions. And as an outcome, encouraging results are observed.
  • Thumbnail Image
    ItemOpen Access
    Fractal based approach for image segmentation
    (Dhirubhai Ambani Institute of Information and Communication Technology, 2004) Londhe, Tushar; Banerjee, Asim
    In this thesis, we have proposed an algorithm for image segmentation, using the fractal codes. The basic idea behind this algorithm is to use fractal codes for the image segmentation. This method uses compressed codes instead of the gray levels of the image. Therefore it is cost effective in the sense of storage space and time as no decoding is performed before using the segmentation algorithm. Moreover, the proposed scheme can directly use on the images accessed from the image database where images are kept in fractal-compressed code.