Image Quality Dependent Degradation for AI Systems

In safety-critical applications like automated driving, deep learning models often struggle when input data deviates from training distributions due to sensor degradation or adverse environmental conditions. The authors address the challenge of maintaining system reliability under these conditions by proposing a fail-degraded architecture that adapts to poor-quality input data without requiring a complete system fallback (such as switching to a secondary sensor suite).

The researchers introduce a two-part strategy to manage perception uncertainty. First, they implement a monitoring system that evaluates image quality using a set of handcrafted features (e.g., brightness, sharpness, noise levels, and Fourier domain energy). These features are fed into a Normalizing Flow (NF) model, which learns the distribution of 'good' quality images from the training set. The NF provides a log-likelihood score that quantifies how much an incoming image deviates from the expected training data distribution. Second, they use this likelihood score to dynamically adjust the confidence threshold of an existing object detector. When the system detects poor-quality input, it lowers the confidence threshold, effectively making the detector more sensitive to potential objects to ensure higher recall.

The study demonstrates that Normalizing Flows trained on handcrafted image features can effectively detect various types of image corruption, including darkness, noise, dirt, blur, lens flares, and rolling shutter artifacts. Unlike latent features extracted from deep backbones like ResNet, which failed to reliably identify these specific augmentations, the handcrafted feature approach provided clear, measurable signals of quality degradation. By coupling these signals with a tanh-based threshold adjustment function, the system can transition from a high-accuracy mode (for clear images) to a high-recall mode (for degraded images), thereby enhancing the safety profile of the perception system.