Can Exposure, Noise and Compression Affect Image Recognition? An Assessment of the Impacts on State-of-the-Art ConvNets

摘要

Convolutional Neural Networks stand the current state-of-the-art in image recognition, as well as many computer vision tasks. Nevertheless, these architectures have been shown to be vulnerable to image manipulations, which may undermine the reliability and safety of CNN-based models in autonomous and robotic applications. We present a rigorous evaluation of the robustness of several high-level image recognition models and investigate their performance under distinct image distortions. We propose a testing framework which emulates ill exposure conditions, low-range image sensors, lossy compression, as well as commonly observed noise types. One one side results measured in terms of accuracy, precision, and F1-Score, indicate that most CNN models are marginally affected by mild miss-exposure, heavy compression, and Poisson noise. Severe miss-exposure, impulse noise, or signal-dependent noise, on the other side, show a substantial drop in accuracy and precision. A careful evaluation of some typical image distortions, commonly observed in computer vision and machine vision pipelines, provides insights and directions for further developments in the field.