This work deals with the analysis of motion blur in biometric face images. In today’s society, the verification of one’s own person is increasingly automated. For this purpose, a reference image, such as the passport photo on an ID card, is compared with a photo taken on site. However, if the subject moves during the photo capture process, motion blurring occurs in the image created. Under certain circumstances, this can lead to an incorrect result of the verification procedure. This motion blur is to be analyzed and estimated. For this purpose, different machine learning and handcrafted approaches are compared with each other. The machine learning approaches are divided into three classification models and a regression model. Furthermore, handcrafted approaches, which are based on the Laplace and Sobel filter and the Fast Fourier Transformation are examined. In addition, an image sharpness approach based on the Open Source Face Image Quality project was used for comparison. It was also investigated whether there is a difference in detection performance between real and synthetic motion blur. For this purpose, the approaches were tested on data sets with synthetically created motion blur and real motion blur and the CNN models were trained. Furthermore, it was investigated whether it is possible to distinguish sharp images from images with motion blur and images with other types of blur. Finally, an attempt was made to determine the direction vector of the motion blur.
Various experiments were carried out. The quality estimation approaches were evaluated using Error versus Discard Characteristic curves and the Mean Absolute Error. The model for estimating different types of blur was tested and evaluated on the basis of synthetically created blur, images with synthetic and real motion blur and sharp images. The proposed algorithm for calculating the angle of motion blur was empirically tested for accuracy and reliability.
It was found that the estimation of motion blur is successful with both the trained and the handcrafted approaches. Likewise, synthetic as well as real motion blur could be used to train and estimate the other. It was also shown that it is possible to differentiate between types of blur. However, no algorithm for determining the direction vector could be found.
This work can be used as a basis for further research. On the one hand, the results of this work can be used as a comparison for further research and, on the other hand, to improve the approaches described here.