Automatic Clarity Threshold Determination of Blurred Images for Computer Vision Tasks’ Security

Blurred images may degrade the performance of Deep Neural Networks (DNN), and then cause security problems of computer vision tasks. However, it is challenging to determine the threshold judging whether an image is blurred or not in most image sharpness/blur detection techniques. In this paper, a method of blurred image threshold determination is proposed. Firstly, blurred image datasets with different degrees of blur are obtained based on frequency domain filtering. Then, the performance of the DNN model and the value of blur detection technique on the blurred image datasets are calculated. The minimum blur degree dataset, which can maintain the model’s acceptable performance, is determined to obtain the clarity threshold of the blur detection technique. This method for blurred image judgment is applicable to DNN and enhances computer vision tasks’ security. Lastly, the thresholds of two blur detection techniques, Frequency Domain Image Blur Measure (FM) and Non-reference structural sharpness (NRSS) are obtained by our method to perform image segmentation tasks based on UNet, which demonstrated its effectiveness.