Restoration of Digital Image using Blur Removal Technique

DOI : 10.17577/IJERTV8IS100341

Download Full-Text PDF Cite this Publication

Text Only Version

Restoration of Digital Image using Blur Removal Technique

Amreen K Kazi1

Department of Electronics and Telecommunication

Datta Meghe College of Engineering Navi Mumbai, India

Dr.S.D.Sawarkar2 Department of Electronics and Telecommunication

Datta Meghe College of Engineering Navi Mumbai, India

Dr. D.J.Pete3

Department of Electronics and Telecommunication

Datta Meghe College of Engineering Navi Mumbai, India

Abstract Image restoration is the process of recuperating an image from a blurred and noisy image. Image restoration is a rudimentary problem in image processing. In recent years, image restoration algorithms are using patch-processing. Images in the actual world are exposed to various types of degradation during image capture, acquisition, storage, transmission and reproduction. Different attributes that degrade an image are blur, noise, contrast and saturation. In general, the image quality measures is classied as subjective or objective techniques. Self- determining evaluation of image quality is expensive and requires a lot of time and the consequence of such an experiment would also depend on viewing angle. Objective image quality metrics are faster than subjective metrics and harvest immediate results while requiring minimum human involvement. Here we shall apply image sharpening to the previously blurred images. We resize the image and do the RGB channel separation. In this paper Gaussian blur will be analysed. As blind deconvolution deconvolutes without the knowledge of the impulse function. Iterative image restoration is been used in this paper. Our system uses MATLAB in which real images are used for performance testing. This papers idea is very eective and will yield to advanced results in denoising, deblurring, segmentation, and other different applications.

Keywords Image restoration, image deconvolution, image deblurring, kernel , image enhancement.


    The edge sharpness wont be retrieved in blind image restoration process with prior information. De-blurring is that the method of removing blurring artefacts from pictures, like blur which is caused by defocus irregularity or speed or motion blur. In the process of image transmission noise can be added, similarly in image acquisition. Noise can be categorised as Gaussian, Rayleigh, exponential, uniform, impulse, etc. The apparent streaking of objects during a still image occurs in motion blur. A Gaussian blur is that the results of blurring a picture by a Gaussian operate. The success of recent single- image strategies partially stems from the employment of varied distributed priors, for either the latent pictures or motion blur kernels. It additionally finds smart kernel matrix approximation to hurry up blurring and accomplish through home the good de- blur performances on digital datasets. The blurred pictures area unit sometimes in low resolution and suffer severe loss of edge info, that solid nice challenge to existing blind deblurring strategies. In motion image blurring, the blur kernel is linear uniform convolution. Verification of the angle of the kernel supported the observation that the recovered image has the primary distributed illustration once the kernel angle

    corresponds to the real motion angle. The motion kernels length is then figured out. Experimental results demonstrate the

    extensiveness of our planned approach in terms of effectiveness.

    If image deblurring could be automatically achieved then it would be of great practical interest for the enhancement of images in individual photos or their video. In humans visual perception is the most trusted source of information than other perceptions. While, image gives the pictorial information it provides minute details. During the process of retrieving the information and then analyzing that pictorial information by a digital computer different steps are taken and this is known as digital image processing.

    The processing for machine perception may corrupt the pictorial information at a large. The different attributes in images include noise, blur, saturation and contrast, which cause image degradations. The image quality measurements are bifercated as subjective or objective techniques. Spatial domain techniques centralize on the change in behaviour of high contrast sharp features of the image like edges, corners and textures that undergo significant changes during image degradation. Frequency domain techniques centralize on evaluating the statistical features of the Subjective techniques are costly, time consuming and dependent on the angle of viewing it. While, objective image quality metrics have faster processing rate also requires minimum human intervention. Objective techniques are further bifurcated into full reference, reduced reference and no reference. The no-reference methods are taken into consideration in spatial domain and the frequency domain techniques. Singular value decomposition (SVD) is included in modified Haar algoritm This resulted in the very low number of missed detections with convalescence of such blurred image is done by using non blind restoration and blind restoration. In non-blind restoration the information about the kernel is known, whereas in blind image restoration the kernels information is unknown.

    In this paper, we propose a innovative structure founded on sparse illustration to identify the blur kernel. Then, we estimate the length of the blur kernel using Radon transform of Fourier domain. Large motion blur can be with this process. Our evaluation is based on real world images and can be compared to numerous blind image deblurring algorithms. Our proposed approach is pre-eminence as it provides effective and robust output.


    users requirement. When in video technology, the or enhancement of digital material is known as up scaling or

    .jpg/.png input image

    image resize & R-G-B

    Channel Seperation

    Gaussian blur addition

    Blur removal – Angle


    resolution enhancement. The geometric transformation scaling can be used for vectored graphic image, without any loss of image quality. When scaling a raster graphics image, a new image with different number of pixels which can be higher or

    Fig. 1. Flowchart of Image Restoration Process

    The module consist of the input image which can be a .jpg or

    .png image. This read image is then resized in 256*256 or 512*512 size and channel separation is being performed later with red, green, and blue channel on it which is known as image preprocessing. Later blur addition is done with respect to gaussian blur, which provides us with the dimension and sigma values. Next deblurring takes place of both Gaussian and motion blur giving initial PSNR, final PSNR and mean square error.

    1. Input image

      A quadrilateral array of pixels giving minute details of image closely aligned together is known as digital image. The pixels in an image signifies the dimension of the property of a scene measured over a definite area. This property could be many different things, which one can usually measure either in the average brightness (single value) or the brightness of the image filtered through red channel, green channel and blue channel filters (triplet values). The values are normally represented by an eight bit integer (octal), having a range of 256 brightness levels. Resolution of an image is quantitative value of pixels in an image and the its brightness values.

      lower must be generated. In the case of decreasing the pixel number which is also known as scaling down this recurrently reults in a visible quality loss. From the perspective of digital signal processing, the scaling of raster graphics is a 2-D example of sample-rate conversion. This conversion is of a discrete signal from a local sampling rate to another. Image scaling can be construed as a form of image resampling or image reconstruction with respect to the Nyquist sampling theorem. According to the Nyquist theorem, down sampling to a smaller resoluted image from a higher-resolution original can only be accepted after applying a suitable 2-D anti-aliasing filter which prevents the aliasing effect. The image is reduced to such an amount of information that can be carried by the smaller image.

      Fig. 2. Input image

    2. Image Preprocessing

      • Image Resize:

        As the images obtained are not of the exact size always as required, hence it is of utmost importance to understand how to properly resize an image according to our requirement. And so we need to understand how resizing works. The pixel information is changed, when an image resizing occurs. In this process if an image is reduced in size, any unwanted pixel information will be discarded which can be performed by different photo editor program like Photoshop. Similarly if an image is enlarged, the photo editor must create pixel information and add new pixel information which is purely based on its best guesses. And hence a larger size image is obtained which typically results in either a over pixelated or very soft or blurry looking image. And so image size reduction is much simpler than enlarge an image. If an image is being used for high-quality (publishing) or large banner prints, it needs to be captured using the highest resolution and quality possible because of the problems faced in enlarging. Fundamentally in graphical imaging and digital imaging, image scaling is a process of resizing a digital image according to the

        Fig. 3. Resized image to 512*512

      • Channel Separation:

        Colour digital images consist of pixel and these pixels are made of amalgamations of primary colors. A channel in this framework is the grayscale image of the same size as that of color image, made of only one of these primary colours. For example, an image from a typical digital camera will have a red, green and blue channel. Ideally grayscale image has only one channel. An RGB image has these three channels: red, green, and blue. RGB channels roughly follow the colour receptors in the human eye, and their applications are found in computer displays and image scanners. If the RGB image is of 24-bit standardized, each channel having 8 bits, for red, green, and blue. This means the image is composed of three images ,one for each channel. Here each image stores discrete pixels with predictable brightness intensities between 0 and 255. If the RGB image is 48-bit which is very high in color-depth, each channel will be made of 16-bit images. Different display devices commonly use the Red, Green, Blue (RGB) additive color system. Subtractive color system is used in printers which is the Yellow, Magenta, Cyan, Black (YMCK). Professionals working with colors generally utilize in Hue, Saturation and Intensity (HSI).

        Fig. 4. Channel separation of Red.

        Fig. 5. Channel separation of green.

    3. Blur Addition

      • Gaussian Blur:

        In image processing, a Gaussian blur typically known as Gaussian smoothing is the outcome of blurring an image by a Gaussian function. It is a extensively used effect in graphics software, typically to reduce image noise and reduce detail. The graphic effect of this blurring procedure is a smooth blur resembling that of viewing the image through a luminous screen, distinctly different from the bokehs effect produced by disburst focus lens or the shadow of an object under usual brightness. Gaussian smoothing is a pre-processing stage in computer hallucination algorithms so as to enhance different scales of image structures. Statistically, applying a Gaussian blur to an image is the same as convolving the image with a Gaussian function which is known as a two-dimensional Weier strass transform. By divergence, convolving by a circle result in accurately reproduction of the bokeh effect. It is known that the Fourier transform of a Gaussian is another Gaussian. Gaussian blur reduces the image's high-frequency components thus Gaussian blur acts as low pass filter.

        Theoretically, at each point on the image the Gaussian function will be non-zero, stating that calculation of a particular image would need the entire image. In preparation, during computation of discrete approximation of the Gaussian function, pixels distanced at greater than 3 are negligible. And hence information of pixel exclusive of the range is ignored. In actual, a matrix with 6 dimensions is used in image processing program . This resultant is satisfactorily as obtained by when the entire Gaussian distribution is performed.

        Though gaussian blur finding application in circularly symmetric, the Gaussian blur can also be applied to a two- dimensional image. This is done with two self-determining one-dimensional calculations which is named as separable filter. And hence a series of single-dimensional Gaussian matrices are applied in horizontal direction and then reiterating the process in the vertical direction achieving two-dimensional matrix. This helps in as simple calculative steps are performed.

        When an image is convoluted with a kernel of Gaussian values it will generate a Gaussian blur effect. This process is divide into two processes or passes. A single dimensional kernel blurring of image is done only in the horizontal or vertical direction in first pass. Next, this single dimensional kernel is later blurred in remaining directions. This reduces the amount of calculation.

        The midpoints of each individual pixels are taken, then samling the Gaussian filter kernel at these discrete points is done to achieve discretization. This eventually reduces the computational cost. But if small filter kernels are available then point sampling the Gaussian function generates huge amount of error. So both needs to be considered during this integration of the Gaussian function. And this will result in darkening or brightening of the actual image. To avoid this, regularization can be performed on values which can be obtained by division of each term in the kernel by the summation of the same.

        Fig. 6. Gaussian blur added to the image

    4. Blur Removal

      • Angle Estimation:

        Angle estimation is a process of transferring a pictorial image from spatial domain to the frequency domain. By presenting sparse representation within angle estimation algorithm is performed. First fourier transform of pre- processed image is taken. Then its log spectrum is being analysed followed by inverse fourier transform on the same. Then edge detection is being done followed by finding minimum and maximum values for both gaussian blur and motion blur respectively.

      • De-convolution:

    For deconvolution, the indistinct convolution data of the impulse response is used. And hence appropriate assumptions of the input that estimate the impulse response by checking the outputs are performed. In iteratively blind deconvolution continuous iteration for estimation of psf is performed which further improves the image quality. In non-iteratively it can be applicated on the algorithm, depending on exterior information, and extracts the PSF.

    Fig. 7. Gaussian deblurring with final PSNR 27.1301 for red, 27.8249 for green, 26.7856 for blue channel.


    The images show sequential output starting with a basic .jpg image as shown in Fig.2. Later the input image is resized by 512*512 for processing as shown in fig.3. We can also resize the image with 256*256 size. Channel separation is done on this resized image simultaneously as red and green channel separation. The Fig.4 and fig.5 shown red and green channel separation respectively. Then as shon above in fig. 6 Gaussian blur is being added to this channel separated image. Fig.6 shows the deblurred image of the gaussian blurred input. And hence we observed that Gaussian blur can be removed by iterative blind deconvolution.


The proposed system has come up with a method to deblurring the image. The gaussian blur are being formulated. Here information is restored by estimating a blur kernel. Additionally in angle estimation coarse and fine filtering can be used. For length estimation Fourier and radon can be used. The performance should be evaluated using the SNR, PSF and score verses angle ratio. The system retains the pictorial information in a readable form that can be easily viewed by human eye.


  1. K. Dabov, A. Foi, V. Katkovnik, and K. Egiazarian, Image restoration by sparse 3d transform-domain collaborative ltering, in Electronic Imaging 2008. International Society for Optics and Photonics, 2008, pp. 681 207681 207.

  2. M. Elad and M. Aharon, Image denoising via sparse and redundant representations over learned dictionaries, Image Processing, IEEE Transactions on, vol. 15, no. 12, pp. 37363745, 2006.

  3. G. Yu, G. Sapiro, and S. Mallat, Solving inverse problems with piecewise linear estimators: from gaussian mixture models to structured sparsity, Image Processing, IEEE Transactions on, vol. 21, no. 5, pp. 24812499, 2012.

  4. I. Ram, M. Elad, and I. Cohen, Image processing using smooth ordering of its patches, Image Processing, IEEE Transactions on, vol. 22, no. 7, pp. 27642774, 2013.

  5. J. Mairal, F. Bach, J. Ponce, G. Sapiro, and A. Zisserman, Non-local sparse models for image restoration, in Computer Vision, 2009 IEEE 12th International Conference on. IEEE, 2009, pp. 22722279.

  6. P. Chatterjee and P. Milanfar, Patch-based near-optimal image denois- ing, Image Processing, IEEE Transactions on, vol. 21, no. 4, pp. 1635 1649, 2012.

  7. J. Mairal, M. Elad, and G. Sapiro, Sparse representation for color image restoration, Image Processing, IEEE Transactions on, vol. 17, no. 1, pp. 5369, 2008

  8. Y. Romano, M. Protter, and M. Elad, Single image interpolation via adaptive nonlocal sparsity-based modeling, Image Processing, IEEE Transactions on, vol. 23, no. 7, pp. 30853098, 2014.

  9. W. Dong, L. Zhang, G. Shi, and X. Li, Nonlocally centralized sparse representation for image restoration, Image Processing, IEEE Transactions on, vol. 22, no. 4, pp. 16201630, 2013.

  10. S. Roth and M. J. Black, Fields of experts, International Journal of Computer Vision, vol. 82, no. 2, pp. 205229, 2009.

Leave a Reply