InverseNet: Benchmarking Operator Mismatch and Calibration Across Compressive Imaging Modalities

Researchers Chengshuai Yang and Xin Yuan have published a critical new benchmark, InverseNet, that exposes a fundamental weakness in AI-powered compressive imaging systems. The study reveals that state-of-the-art models like EfficientSCI can catastrophically fail in real-world deployment, losing up to 20.58 dB in performance when their mathematically assumed 'forward operator'—the model of how light is captured—deviates from the physical hardware. This 'operator mismatch' is the default condition in deployed systems like CASSI, CACTI, and single-pixel cameras used in medical and scientific imaging. Until now, no benchmark quantified this problem, leaving a dangerous performance gap between lab results and field applications.

The InverseNet benchmark evaluates 12 methods across four scenarios (ideal, mismatched, oracle-corrected, blind calibration) and makes several stark findings. First, the performance advantage of deep learning methods over classical baselines completely disappears under mismatch. Second, architectures that ignore the physical mask ('mask-oblivious') recover 0% of the loss, while 'operator-conditioned' methods can recover 41-90%. Most promisingly, the research shows that blind calibration via grid-search can recover 85-100% of the ideal 'oracle' performance without needing ground truth data. These results, confirmed on real hardware, provide a crucial roadmap for building robust, deployable AI imaging systems that work outside the controlled lab environment.