| --- |
| license: mit |
| --- |
| Bayesian Enhancement Model (BEM) |
|
|
| AAAI 2026 — Bayesian Neural Networks for One-to-Many Mapping in Image Enhancement |
|
|
| Paper: [arxiv:2501.14265](https://arxiv.org/abs/2501.14265) |
|
|
| Github: https://github.com/BinCVER/BEM |
|
|
| 🌟 Overview |
|
|
| Image enhancement is inherently one-to-many — a single degraded image often corresponds to multiple plausible enhanced outputs, especially in low-light and underwater environments. |
|
|
| However, existing deterministic models can only produce one prediction, failing to capture this ambiguity. |
|
|
| BEM (Bayesian Enhancement Model) solves this by: |
|
|
| Modeling uncertainty using a Bayesian Neural Network (BNN) |
|
|
| Generating multiple valid enhancement candidates |
|
|
| Selecting or aggregating candidates via Ranking or Monte Carlo |
|
|
| Refining details using a second-stage DNN |
|
|
| Achieving DNN-level inference speed with BNN-level diversity |
|
|
| This makes BEM the first practical Bayesian solution for large-scale image enhancement tasks. |
|
|
| 🔍 Key Features |
| ✓ One-to-Many Enhancement via Bayesian Inference |
|
|
| BEM samples weights from a variational posterior learned by a Bayesian UNet. |
| Each sample produces a distinct enhancement, capturing the ambiguity of the task |
|
|
