Deepika Hazarika

Blocking artifacts reduction using adaptive bilateral filtering

In this paper, we propose a simple, non iterative blocking artifacts reduction method for block discrete cosine transform (DCT) compressed images, using adaptive bilateral filter. Bilateral filter when applied on a input image, smooths out the blocking artifacts by weighted averaging of the pixel values without smoothing the edges. The proper selection of the bilateral filter parameters is very important and affects the filtering results significantly. We select the bilateral filter parameters optimally, through empirical study. The bilateral filter parameters are made adaptive to different decompressed images using different quantization tables. Our main contribution in this paper is to select the bilateral filter parameters optimally and adaptively through empirical study, in application to image deblocking. The proposed method shows highly encouraging results both objectively and subjectively, when compared to many state of the art image deblocking schemes including a recent method based on adaptive bilateral filter.

Despeckling SAR images in the lapped transform domain

In this paper, a novel lapped transform (LT) based approach to SAR image despeckling is introduced. It is shown that LT coefficients of the log transformed, noise free SAR images, obey Generalized Gaussian distribution. The proposed method uses a Bayesian minimum mean square error (MMSE) estimator which is based on modeling the global distribution of the rearranged LT coefficients in a subband using Generalized Gaussian distribution. Finally the proposed algorithm is implemented in cycle spinning mode to compensate for the lack of translation invariance property of LT. Experiments are carried out using synthetically speckled natural and SAR images. The proposed Bayesian based technique in LT based framework, when compared with several existing despeckling techniques, yields very good despeckling results while preserving the important details and textural information of the scene.

Comparison of Generalized Gaussian and Cauchy distributions in modeling of dyadic rearranged 2D DCT coefficients

The dyadic rearrangement of block two-dimensional Discrete Cosine Transform (2D DCT) coefficients when used with zero tree quantizers show comparable performance with that of discrete wavelet transform based methods for image compression applications. Recently we have shown that Generalized Gaussian distribution better models the statistics of subband rearranged 2D DCT coefficients compared to Gaussian, Laplacian and Gamma distributions. This paper presents results of distribution tests that indicate that for most of the natural images Cauchy distribution models the subband coefficients more accurately than Generalized Gaussian distribution. The knowledge of the suitable distribution helps in design of optimal quantizers that may lead to minimum distortion and hence achieve optimal coding efficiency.

Lapped transform-based image denoising with the generalised Gaussian prior

We introduce a new image denoising method based on the statistical modelling of dyadic rearranged lapped transform LT coefficients. Based on Kolomogrov-Smirnov KS goodness of fit test, we have shown that the statistical distribution of the dyadic rearranged LT coefficients in a subband is best approximated by the generalised Gaussian distribution. A Bayesian minimum mean square error MMSE estimator is used to obtain the estimate of noise free coefficients, which is based on modelling the global distribution of the dyadic rearranged LT coefficients using generalised Gaussian distribution. The LT-based image denoising method with generalised Gaussian prior shows highly encouraging both objective and subjective results when compared to several well-known image denoising methods.

Blocking artifacts suppression in Wavelet transform domain using local Wiener filtering

In this paper, a new non iterative method is proposed in the wavelet domain for the reduction of blocking artifacts in block discrete cosine transform (DCT) compressed images. The method employs a local Wiener filter in wavelet domain to smooth out the blocking artifacts. The local variance is used in the same domain for pixel adaptive processing. We have demonstrated that the noise standard deviation has a strong correlation with the average of first 3×3 values from the quantization matrix. The proposed method outperforms several well know image deblocking methods both objectively and subjectively.

A novel approach to color image compression using 3-d Discrete Cosine Transform (DCT)

Most of the color image compression schemes transforms the highly correlated RGB color space into a decorrelated color space like YUV, YCrCb etc. in order to reduce the redundancies between the color components, then the decorrelated color components are coded individually by monochrome image compression schemes. In this paper, we present a novel approach to color image compression scheme using 3-Dimensional Discrete Cosine Transform (3-D DCT). We use 3-D DCT to reduce the intercolor redundancy. In the proposed scheme the 3-D block DCT is directly applied on the highly correlated RGB color components and the coefficients are quantized and entropy coded. We also propose a wavelet structure for 3-D block DCT coefficients. We use Lagrange multiplier method for optimal bit allocation between the different slices of 3-D DCT coefficients. Our proposed coder gains competitive results comparing to the performance of JPEG baseline and JPEG2000 standard. Experimental results shows that the proposed coder performs close to the JPEG2000 standard in terms of R-D performance and outperforms the JPEG baseline standard by a large margin.

A lapped transform domain enhanced lee filter with edge detection for speckle noise reduction in SAR images

This paper describes a method which uses Lapped orthogonal transform (LOT) domain adaptive enhanced Lee filter for despeckling SAR images. Since LOT is block transform, the transform coefficients are first remapped into octave type form and then the enhanced Lee filtering is applied to subband LOT coefficients. The motivation of using Lapped transform lies in its ability to preserve oscillatory type of features present in the images like textures. For more edge preservation during despeckling process, the modified ratio of averages (MROA) edge detector is applied to the approximation subband to obtain edge information which is then employed in the proposed framework to obtain edge information in other subbands. Based on this edge information we classify the edge and non edge coefficients in all oriented subbands at various decomposition levels. Experiments using true SAR images show that the LOT domain enhanced Lee filter in proposed edge preserving framework smoothes the speckle very well in homogeneous regions while preserving more edges and texture information. The proposed despeckling filter shows significant improvement over enhanced Lee filtering in spatial and wavelet domain and also outperforms one recent undecimated wavelet domain method.

Video Noise Reduction in 3-D Mixed Transform Domain Using Its Efficient Wavelet Structure

In this paper, we propose a novel 3-D mixed transform based video denoising scheme for additive white gaussian noise. The 3-D mixed transform consists of lapped biorthogonal transform (LBT) in vertical and horizontal directions and discrete cosine transform (DCT) in temporal direction. The efficient wavelet structure of 3-D mixed transform coefficients is used for denoising the video data. The concept of wavelet structure of 3-D mixed transform coefficients is taken from our recent successful lossy color image compression scheme. The wavelet structure of 3-D mixed transform coefficients allow to use efficient wavelet based image denoising algorithms without wavelet transform. We use the approximate minimum mean square error (MMSE) estimation procedure to denoise each individual slice of 3-D mixed transform coefficients. We propose to use the simple selective recursive temporal filtering with the proposed denoising scheme which improves the average peak signal to noise ratio (PSNR) further by 1dB - 2dB. Experimental results demonstrate that the proposed scheme shows encouraging results when compared to many existing complex wavelet based denoising schemes both in terms of PSNR and visual quality.

A 3-D block transform based approach to color image compression

In this paper, we present a novel approach to color image compression using 3-D block transforms. Unlike, most of the other color image compression schemes, the proposed scheme do not transform the highly correlated RGB color components into a decorrelated color space like YUV, YCrCb etc but uses 3-D block transforms which is applied directly on RGB color components to reduce the intercolor redundancy, then the 3-D block transform coefficients are quantized and entropy coded. Two different transforms are used in this work, the first is 3-D Discrete cosine transform (DCT) and the second is Mixed transform. The Mixed transform consists of the Lapped Bi-orthogonal Transform in vertical and horizontal directions and DCT in temporal direction. The 3-D DCT exhibits blocky artifacts at low bit rates but Mixed transform gives better visual results. Both transform produces NxNx3 transform coefficients when applied on a NxNx3 3-D data block. We propose a wavelet structure for 3-D block transform coefficients which enables a very efficient encoding of the 3-D block transform coefficients at exact desired bit rate. Our proposed coder shows encouraging results when compared to the Rate Distortion performance of JPEG baseline coder and JPEG2000 coder. The R-D performance of the proposed coder outperforms the JPEG baseline coder consistently by a large margin and is roughly comparable to the JPEG2000 coder.

An Efficient Multiscale Wavelet Local Binary Pattern for Biomedical Image Retrieval

A method for biomedical image retrieval using multiscale wavelet local binary pattern (LBP) is presented in this paper. The method first decomposes a biomedical image into approximation and oriented detail subbands using discrete wavelet transform (DWT). Since the oriented detail subbands at each scale exhibit distinct directional features the proposed method employ a new 4-point LBP with selected non-diagonal neighbors in horizontal and vertical subbands, and a 4-point LBP with selected diagonal neighbors in diagonal subband to extract the LBP histogram. An 8-point LBP is employed in approximation subband to extract the LBP histogram. The biomedical image is finally represented by a single feature histogram that is formed by concatenation of all the LBP histograms. The proposed method provides significantly reduced feature vector size while maintaining same or most of the times better retrieval efficiency compared to original LBP and other relevant wavelet-based LBP schemes. The Euclidean distance measure is used for query matching and retrieval is performed based on the least matching distance. The method is tested using OSIRIX image data sets and experimental results validating the efficiency of the proposed method over other relevant schemes, are presented.