Bounded Hidden-Screen Reconstruction from Diffuse Wall Reflections via Survivability-Guided Representation Learning

Reconstructing a hidden-screen image from one ordinary Red-Green-Blue (RGB) photograph of a diffuse wall when the screen is outside the camera’s direct line of sight is a challenging problem in passive indirect imaging that highlights potential visual-information leakage. This task presents two main challenges: diffuse reflection unevenly attenuates spatial cues, and different screen content may require observation-dependent rather than fixed feature aggregation. To address these issues, we introduce Survivability-Guided Representation Learning (SGRL). Our method features Survivability-Band Representation (SBR), which uses two Gaussian-initialized trainable depthwise operators to form three nominal complementary construction paths, LOW, MIDDLE, and HIGH. Content-Conditioned Cross-Band Fusion (CCF) then predicts observation-conditioned computational coefficients to combine the branch features and generate the hidden-screen reconstruction. Experiments under controlled passive imaging conditions show that SGRL achieves the strongest recorded reconstruction performance among the evaluated controls across multiple metrics, supporting the structure-level recovery of broad brightness distribution, region layout, and major visual structure from diffuse wall observations.