Impact of motion blur on recognition rates of CNN-based TOD classifier models

This work investigates the impact of various types of motion blur on the recognition rate of triangle orientation discrimination (TOD) models. Models based on convolutional neural networks (CNNs) have been proposed as an automated and faster alternative to observer experiments for range performance assessment. They may also give insights into the impact of system degradations on the performance of automated target recognition algorithms. However, the effects of many image distortions on the recognition rate of such models are relatively unknown. The recognition rate of CNN-based TOD models is examined in terms of different forms of motion blur, such as jitter, linear and sinusoidal motion. For model training and validation, simulated images are used. Triangles with four directions and different sizes, positions are used as targets, which are superposed on natural images as background taken from the image database "Open Images V7". Motion blur of varying strength is applied to both the triangle and the entire image to simulate movements of the target and imager. Additionally, common degradation effects of imagers are applied, such as white sensor noise and blur due to diffraction and detector footprint. The recognition rates of the models are compared for target motion and global motion as well as for the different motion types. Furthermore, dependencies of the recognition rate on blur strength, triangle size and noise level are shown. The study shows interrelationships and differences between target motion and global motion regarding TOD classifications. The inclusion of motion blur in training can also increase model accuracy in validation. These findings are crucial for range performance assessment of thermal imagers for fast-moving targets.