Factors Influencing CNN Performance
Ieuan M. Israel, Steven A. Israel, John M. Irvine
Abstract
The current drive toward incorporating machine learning (ML) models within automated closed-loop workflows renews existing issues for image quality prediction; specifically, maintaining the operator-in-the-loop's trust in system operation. This paper reviews methods to assess information content in the reconstruction process to maintain trust in closed-loop workflows. We build ML models from commonly used architectures. For previous manual exploitation, the interpretability of an image indicates the potential intelligence value of the data. Historically, the National Imagery Interpretability Rating Scale (NIIRS) has been the standard for quantifying the intelligence potential based on image analysis by human observers. Empirical studies have demonstrated that spatial resolution is the dominant predictor of the NIIRS level of an image and that compression to 1 bit per pixel can maintain that NIIRS level. However, with modern ML that digests images rather than extracted features, what is the corresponding size of the latent space required to maintain the NIIRS levels? To those ends, we operate on moderate size images, 480×480 pixels, to provide realistic generalizable estimates over those experiments against similar sprite size (<100×100 pixels) images.