Research map / Image generation / Image restoration
Logarithmic Lenses: Exploring Log RGB Data for Image Classification
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Shows that training standard CNNs on log-of-linear RGB data from RAW images (and pseudo-log data recovered from 8-bit sRGB JPGs) improves accuracy and invariance to intensity and color-balance changes, and introduces the RAW10 dataset.
Problem
Image classification pipelines feed networks sRGB images processed for human viewing, and little is known about whether a different input data space (linear or log RGB) that follows physical reflection rules would improve deep network performance and robustness to illumination changes.
Approach
Use log of linear RGB as network input, derived from RAW images, to exploit the fact that log transforms turn illumination-dependent body reflection cylinders into a nearly fixed structure; also show that inverting the sRGB transform on 8-bit JPGs to get pseudo-log data recovers most of the gain.
Key results
- ResNet18 on RAW10 original test set: log RGB 90.6% vs linear 88.9% vs sRGB 86.9%; under combined intensity and color-balance modification log reaches 83.9% vs 67.3% for sRGB.
- DenseNet121 original test: log RGB 91.9% vs sRGB 91.4%; under combined modification log 86.9% vs sRGB 81.0%.
- Pseudo-log data (inverse sRGB from JPGs) trained networks reach nearly the same accuracy as true log data (e.g., ResNet18 pseudo-log 90.29% vs true log 90.64% on original test set).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use log RGB instead of the strongest sRGB or linear RGB input | ResNet18 on RAW10 · Original, random color balance, random intensity, and both modifications | accuracy: 88.9 -> 90.6; 75.7 -> 86.9; 76.3 -> 87.2; 71.9 -> 83.9 | Helped |
| use log RGB instead of the strongest sRGB or linear RGB input | DenseNet121 on RAW10 · Original, random color balance, random intensity, and both modifications | accuracy: 91.6 -> 91.9; 84.7 -> 88.0; 91.4 -> 91.8; 82.0 -> 86.9 | Helped |
| use log RGB instead of sRGB | Custom 11-layer CNN on RAW10 with original training data · Original test set | accuracy: 72.05 -> 64.32 | Hurt |
| use log RGB instead of linear RGB | Custom 11-layer CNN on RAW10 with fixed augmentation · Original test set | accuracy: 73.52 -> 76.92 | Helped |
| use log RGB instead of linear RGB | Custom 11-layer CNN on RAW10 with dynamic augmentation · Original test set | accuracy: 73.87 -> 75.40 | Helped |
| train on pseudo-log rather than sRGB | ResNet18 and DenseNet121 on RAW10 · Original and modified test sets | accuracy | Helped |
| train on pseudo-log rather than true-log | ResNet18 and DenseNet121 on RAW10 · True-log and pseudo-log test sets | accuracy: 90.64 vs 90.29; 91.94 vs 91.74 | No clear effect |
| scale log RGB to [0, 1] (ablation) | Log RGB training pipeline · RAW10 image classification | accuracy | Hurt |
| scale input to [0, 256] before taking the log (ablation) | Log RGB training pipeline · RAW10 image classification | accuracy | Hurt |
| subtract 5.05 from log RGB to center it on zero (ablation) | Log RGB training pipeline · RAW10 image classification | accuracy | Hurt |
| resize images after converting them to log RGB (ablation) | Log RGB preprocessing · RAW10 image classification | accuracy | Hurt |
| use different optimized meta-parameters for log input (ablation) | ResNet18 and DenseNet121 on RAW10 · Validation set | accuracy: 87.60 -> 89.08; 88.30 -> 90.17 | Helped |
Surprising findings
- Log-trained networks required different meta-parameters (e.g., Adam and a different learning rate) than sRGB or linear networks, and simply substituting log RGB into an existing pipeline did not work well.
- Interpolating or normalizing log RGB data (scaling to [0,1] or centering) hurt training, so resizing had to be done in linear space.
Limitations
- Only three architectures, all trained from scratch on a single 10-class dataset of about 10k images.
- Log RGB needs different hyperparameters and careful handling, so it is not a drop-in replacement.
- Pseudo-log data is quantized and compressed, so it only partly recovers the gains of true log data.
Experimental setup
Datasets: RAW10
Benchmarks: RAW10
Metrics: accuracy loss
Baselines: sRGB input linear RGB input Funt and Zhu Laplacian of log RGB Cotogni and Cusano equivariant networks
Builds on: ResNet DenseNet dichromatic reflection model
Closest papers by meaning
- Model-Based Image Signal Processors via Learnable Dictionaries (AAAI 2022 · similarity 0.85)
- ReRAW: RGB-to-RAW Image Reconstruction via Stratified Sampling for Efficient Object Detection on the Edge (CVPR 2025 · similarity 0.84)
- RGB-to-Polarization Estimation: A New Task and Benchmark Study (NeurIPS 2025 · similarity 0.84)
- Alignment-Free RGB-T Salient Object Detection: A Large-Scale Dataset and Progressive Correlation Network (AAAI 2025 · similarity 0.84)
- Discriminating image representations with principal distortions (ICLR 2025 · similarity 0.83)
- Towards RAW Object Detection in Diverse Conditions (CVPR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.