Reliability-Gated Uncertainty-Aware Objective for Image Denoising

Abstract

Deep image denoisers achieve strong performance on synthetic additive white Gaussian noise (AWGN), but under heavy noise they often oversmooth edges and destabilize fine textures. Existing gradient-and frequency-based regularization can improve detail preservation, yet uniformly enforcing such constraints may over-regularize ambiguous regions and introduce ringing artifacts. This paper proposes a reliability-gated uncertainty-aware objective for image denoising. A denoising network predicts both the restored image and a pixel-wise uncertainty map using a heteroscedastic Gaussian likelihood. The predicted uncertainty is further transformed into a reliability gate that adaptively modulates edge-gradient and spectral constraints during training. Consequently, reliable regions receive stronger supervision, whereas uncertain regions are regularized more conservatively. Unlike conventional uncertainty modeling used only for residual weighting, the proposed method uses uncertainty as a control signal for selective regularization. The proposed reliability gate and the edge/frequency regularization are used only during training and are not evaluated at inference time. In the current implementation, the network still predicts the log-variance map together with the restored image, so retaining uncertainty estimation introduces only a lightweight final-layer overhead. For restoration-only deployment, the log-variance output channels can be pruned from the final projection, in which case the deployed model produces only the restored image and incurs no additional gating or regularization cost.Experiments on standard AWGN benchmarks at noise levels σ = 15/25/50 show a favorable fidelity-stability trade-off and informative uncertainty estimates under increasing noise.