CoRAS: Adaptive sensing slashes measurements while ensuring image quality

High-resolution imaging systems often struggle with a fundamental question: when have enough measurements been collected to reconstruct an image accurately? Traditional fixed-rate sensing methods over-collect on easy images and under-collect on hard ones, leading to inefficiency or quality loss. Jiawei Yang and Yao Zhang address this with Conformalized Rate-Adaptive Sensing (CoRAS), a technique that leverages conformal prediction to adapt the acquisition rate per image. As measurements are gathered, a reconstruction model produces a path of improving image quality. CoRAS analyzes this path at an early decision time to estimate the optimal stopping point—when the reconstruction error first falls below a user-defined target. It then calibrates this estimate using a reference set of images with similar early reconstruction behavior, yielding a statistically rigorous upper bound on the stopping time with both marginal and approximate conditional coverage guarantees.

In experiments on standard image datasets, CoRAS achieved target error coverage while reducing the average number of measurements by up to 20% compared to fixed-rate baselines. The method naturally allocated more measurements to complex or noisy images, ensuring consistent reconstruction quality across varying difficulty levels. This adaptive approach has broad implications for fields like medical imaging, satellite remote sensing, and IoT camera systems, where minimizing acquisition time or data transfer costs is critical. CoRAS provides a principled, uncertainty-aware framework that balances efficiency and accuracy, potentially replacing one-size-fits-all sensing strategies with per-sample optimization.