Blur is produced in a digital image due to low pass filtering, moving objects or defocus of the camera lens during capture. Image viewers are annoyed by blur artefact and the image's perceived quality suffers as a result. The high-quality input is relevant to communication service providers and imaging product makers because it may help them improve their processes. Human-based blur assessment is time-consuming, expensive and must adhere to subjective evaluation standards. This paper presents a revolutionary no-reference blur assessment algorithm based on re-blurring blurred images using a special mask developed with a Markov basis and Laplace filter. The final blur score of blurred images has been calculated from the local variation in horizontal and vertical pixel intensity of blurred and re-blurred images. The objective scores are generated by applying proposed algorithm on the two image databases i.e., Laboratory for image and video engineering (LIVE) database and Tampere image database (TID 2013). Finally, on the basis of objective and subjective scores performance analysis is done in terms of Pearson linear correlation coefficient (
The computational models accept the challenging task of image quality assessment. For digital image processing systems, it is also a necessary task to measure the blur, noise and other degradations in an image and assess its quality [
The metrics based on the edge analysis, which firstly examine the total number of blurred edges in an image and then gradient threshold was applied to remove the faint edges. Sobel edge detector was used for the edge detection [
The wavelet transform was also used to obtain the detail of horizontal/vertical directions of the image and then the Average cone ratio (ACR) method was used for the calculation of blurriness metric [
The Screen image quality evaluator (SIQE) approach was used to estimate the visual quality of screen content images using a training-based Support vector regression (SVR) model [
We proposed an improved no-reference blur detection algorithm for distorted images, based on the re-blurring concept. According to the first step of the proposed method, a special mask is developed by using the concept of Markov Basis and the Laplace filter. In the second step, the distorted image is re-blurred by using the special mask. For each test and re-blurred image, the local variation in horizontal and vertical pixel intensity is computed. The difference between test and re-blurred images in terms of vertical and horizontal variance is assessed. The maximum value from the horizontal and vertical blur score has been considered as the final blur score.
The rest of the paper is split into four sections. The second section provides an overview of existing approaches. The third section elaborates the Markov basis and proposed no-reference blur assessment method. Fourth section discusses the performance parameters and how they are being used to evaluate the proposed algorithm. The whole work is concluded in fifth section.
A review on the concepts of Markov Basis, basic moves and contingency tables is presented in this section, which will use in the development of proposed algorithm.
Consider, a finite set
For the self-supporting model of two-way contingency tables, similar tests are performed by the set of
A move is defined by
For all
A one-dimensional column vector represents all the elements of
if
if
As explained earlier, the contingency table
If the matrix
The MATLAB software has been used for the development of a proposed no-reference blur detection algorithm. A two-step process is followed by the proposed algorithm to detect blur in distorted images. The first step is to create a special mask based on a Markov basis and a Laplace filter. In the next step, the distorted image is re-blurred by using special mask and the final blur scores are calculated from the local variation of pixel intensities. The explanation of proposed method is given below:
The mask
The flow chart as shown in
The above algorithm is used to determine the blur score. The objective scores are produced by executing the proposed algorithm on several image datasets. Finally, performance analysis was conducted in terms of various performance parameters using objective and subjective scores.
Images from publicly available databases LIVE database [
The regression analysis provided by the Video Quality Expert Group (VQEG) [
Consider the array
SROCC stands for stochastic rank-based correlation approach and it is used to compare two datasets statistically. SROCC coefficient is determined from the mapped subjective scores that is
The RMSE is a mathematical grading rule that determines the error's average magnitude. The MAE is a statistic that quantifies the average magnitude of errors in a group of predictions without taking into account their orientation. The MAE and the RMSE can be combined to detect variance in prediction errors. The RMSE will always be greater than or equal to the MAE; the bigger the gap between them, the more variation in the sample's individual errors. The MAE and the RMSE can both be anywhere between 0 to ∞. Root mean square error and mean absolute error can also be determined from the subjective and objective scores according to the mathematical expression given below:
The performance of the blur metric can also be assessed by the outliers ratio. The ratio of false scores obtained by the objective metric to the total scores is known as outliers ratio. The false scores are the outside interval scores dependent on the standard deviation given by [
Many researchers accepted the non-linear regression analysis for the performance analysis, because it uses the monotonic curve to map the predicted
The mathematical model function given in
No-reference blur detection algorithm | Technique used | No of images | Performance parameters | ||||
---|---|---|---|---|---|---|---|
PLCC (%) | SROCC (%) | MAE | RMSE | OR | |||
JNB Algorithm [ |
Probability summation model based JNB concept | 174 | 0.8498 | 0.8345 | 6.9317 | 9.0117 | 0.1322 |
CPBD Algorithm [ |
JNB based cumulative probability of blur detection | 174 | 0.9257 | 0.9444 | 4.7098 | 6.4662 | 0.0632 |
EMBM Algorithm [ |
Edge modelling | 174 | 0.9230 | 0.9294 | 5.0592 | 6.5787 | 0.0632 |
Proposed algorithm | Calculation of blur variation after re-blurring of blurred image | 174 | 0.9384 | 0.9474 | 4.4109 | 5.9081 | 0.0402 |
Similarly, for TID2013 database, the
No-reference blur detection algorithm | Technique used | No of images | Performance parameters | ||||
---|---|---|---|---|---|---|---|
PLCC (%) | SROCC (%) | MAE | RMSE | OR | |||
JNB algorithm [ |
Probability summation model based JNB concept | 125 | 0.7113 | 0.6902 | 0.6688 | 0.8771 | 0.7440 |
CPBD algorithm [ |
JNB based cumulative probability of blur detection | 125 | 0.8522 | 0.8718 | 0.5241 | 0.6469 | 0.7040 |
EMBM algorithm [ |
Edge modelling | 125 | 0.8750 | 0.8654 | 0.4810 | 0.6042 | 0.6720 |
Proposed algorithm | Calculation of blur variation after re-blurring of blurred image | 125 | 0.8773 | 0.8664 | 0.4862 | 0.5989 | 0.6880 |
Using the re-blurring concept, we were able to deal with the problem of no-reference image blur assessment. The goal of this re-blurring concept is to get beyond the limitations of earlier no-reference blur assessment algorithms. Unlike earlier research, the proposed technique does not rely on edge detection or block processing to calculate blur in a no-reference image. Instead, it relies on the variation of pixel intensities after re-blurring of image. The computation of the blur score has been done using a two-step process. The Laplace filter has been used to create a Markov-based special mask in the first step. In the next step, the blurred image was re-blurred using this special mask and pixel intensity variations were calculated. The final blur score was calculated using the highest value of blur score from both the horizontal and vertical blur scores. To demonstrate the usefulness of the proposed algorithm, Gaussian blurred images from the LIVE and TID 2013 databases were utilized. The results illustrate that the proposed method was successful in predicting high blur scores with high accuracy as compared to existing no-reference blur assessment algorithms such as JNB, CPBD and EMBM algorithms. The future scope of research included that this method can also be improved by using optimization techniques in the selection of special mask. The work can be further extended by using genetic algorithms and machine learning concepts.