InverseNet：压缩成像模式下操作不匹配与校准的基准测试

The study reveals that state-of-the-art EfficientSCI suffers a loss of 20.58 dB when its assumed forward operator deviates from physical reality across just eight parameters. Despite this significant impact, no existing benchmarks quantify operator mismatch, which is a default condition in deployed compressive imaging systems. InverseNet is introduced as the first cross-modality benchmark for operator mismatch, covering CASSI, CACTI, and single-pixel cameras. Evaluating 12 methods under a four-scenario protocol (ideal, mismatched, oracle-corrected, blind calibration) across 27 simulated scenes and 9 real hardware captures, findings indicate that deep learning methods lose 10-21 dB under mismatch, negating their advantages over classical baselines. Additionally, performance and robustness are found to be inversely correlated across modalities.