Image Likelihood

Image likelihood, the probability of observing a specific image given a model, is crucial for various image processing tasks, including inverse problems and out-of-distribution detection. Current research focuses on efficiently computing image likelihoods using generative models, particularly those based on normalizing flows and transformers, often incorporating techniques like vector quantization for high-resolution images. These advancements enable improved performance in applications such as image restoration, robotic control (via robust visual feedback), and ensuring the safety and reliability of clinical image analysis by identifying unreliable predictions. The accurate estimation of image likelihoods is thus vital for bridging the gap between image statistics and human perception, leading to more robust and reliable image processing systems.