Exploration of Image Blur Detection Methods on Globally Blur Images

Abstract

Digital image has played an essential role in computer vision. However, it often suffers from blurriness. Thus, computer vision systems require an automated blur detection process to exclude blurred images. Furthermore, blur detection could prevent misinterpretation by a subsequent operation, such as image classification. This study explores some existing blur detection methods, including focus measure thresholding, support vector machine (SVM) classifiers, and convolutional neural networks (CNN) on globally blurred images. Those methods are robust for abroad applications such as blur detection, region segmentation, multi-class blur classification, autofocus, and shape-from-focus (SFF). Nevertheless, those methods theoretically could also be used for blur detection on globally blur images. We evaluate the techniques on a public and private dataset. The result indicates that the proposed CNN model performs the best blur detection among other methods. This model achieves 0.900 in accuracy, 0.857 in specificity, 0.942 in sensitivity, 0.868 in precision, 0.937 in negative predictive value (NPV), 0.904 in f1-score, and has an average execution time of 0.233 seconds. This model could be helpful in other computer vision system that requires blur image detection, such as identity card recognition, optical character recognition, and image data cleaning.