Imager assessment by classification of geometric primitives

A large variety of image quality metrics has been proposed within the last decades. The majority of these metrics has been investigated only for single image degradations like noise, blur and compression on limited sets of domain-specific images. For assessing imager performance, however, a task-specific evaluation of captured imagers with user-defined content seems, in general, more appropriate than using such metrics. This paper presents an approach to image quality assessment of camera data by comparison of classification rates of models individually trained to solve simple classification tasks on images containing single geometric primitives. Examples of considered tasks are triangle orientation discrimination or the determination of number of line pairs for bar targets. In order to make models more robust against image degradations typically occurring in real cameras, data augmentation is applied on pristine imagery of geometric primitives in the training phase. Pristine imagery is impaired by a variety of simulated image degradations, e.g. Gaussian noise, salt and pepper noise for defective pixels, Gaussian and motion blur, perspective image distortion. The trained models are then applied to real camera images and classification rates are calculated for geometric primitives of different sizes, contrasts and center positions. For task-related performance ranking, these classification rates could be compared for multiple cameras or camera settings. An advantage of this approach is that the amount of training data is practically inexhaustible due to artificial imagery and applied image degradations, which makes it easy to counteract model overfitting by increasing the number of considered realizations of image degradations applied to the imagery and hence increasing the variability of training data.