Image blur detection

• Main Author: Chong, Jun Yuan.
• Other Authors: Kot Chichung, Alex
• Format: Final Year Project
• Language: English
• Published: 2013
• Subjects: DRNTU::Engineering::Electrical and electronic engineering::Computer hardware, software and systems
• Online Access: http://hdl.handle.net/10356/54675
• Institution: Nanyang Technological University

In this project, we aim to categorize and quantify blurriness of out-of-focus blurred images. Many existing databases utilize subjective image quality assessment to correlate with objective quality measures and the ground truth of the blurriness of the image is often dependent on a number of factors. The results obtained from subjective image quality assessment depend on the human vision system (HVS). The way how human perceive sharpness varies by age, eyesight, mood, motivation, fatigue and other environmental conditions. Furthermore, subjective testing are relatively time consuming and cumbersome when it comes to designing the experiment trials and may cost a huge amount for all the set-ups. The ground truth of the subjective databases may vary and thus may not be reliable and consistent when correlating with the objective image quality assessment. We have decided to overcome this by using a non-subjective image quality assessment through establishing the ground truth based on the natural blur caused by images. Our proposed database will consist of 51 out-of-focus sets of images which will be blurred naturally. (One sharp reference with four different blur degrees). The four natural blurred images will be correlated with the artificial blurred image generated by blurring the sharp reference image. The maximum correlation obtained will be the scores for the blurred images. The comparative advantage of our database is that the ground truth for the blurred image will be more consistent and much more reliable as it is based on correlation between the natural blur with the artificial blur rather than subjects where the results may vary a lot. Secondly, our database consists of the most common type of image degradation which is out of focus blur. Thirdly, the scores of the database images will be more precise. The subjective scorings are based on a non-numerical scale: Bad, Poor, Fair, Good, and Excellent and after which these scores will be linearly mapped to a chosen scale. The mapping of the 5 categories of scoring to linear scale may not be precise enough and lack detailed analysis. However, by doing correlation, we take into consideration the pixels and also the image content which is definitely more accurate and precise. Last but not least, weare able to categorize the database based on the environment setting which is one crucial factor in deciding the overall sharpness of the image. The lower the lighting, the more blur the image may appear. We explored 3 main categories, mainly: Outdoors (Cloudy), Outdoors (Sunny) and Indoors. The 3 main categories each contain 2 domains: Zoom in and Zoom out. After setting up the database, we will run existing algorithms to test the correlation between the results obtained from the metrics and also the evaluated results in the databases. We use 4 main performance metrics: SROCC (Spearman rank order correlation coefficient), CC (Pearson’s linear correlation coefficient), RMSE (Root mean square error), MAE (Mean absolute error) to assess the quality measure performance.