Research map / Architectures and efficiency / CNN design
Feature Quantization for Defending Against Distortion of Images
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add data-dependent power exponents (HPOW) (ablation) | ResNet-50 · ILSVRC-12 motion blur | Top-5 accuracy: 63.29 -> 70.24 | Helped |
| Add data-dependent power exponents (HPOW) (ablation) | ResNet-50 · ILSVRC-12 salt and pepper | Top-5 accuracy: 69.94 -> 75.20 | Helped |
| Add data-dependent power exponents (HPOW) (ablation) | ResNet-50 · ILSVRC-12 mixed distortions | Top-5 accuracy: 65.35 -> 70.96 | Helped |
| Add data-dependent power exponents (HPOW) (ablation) | ResNet-50 · ILSVRC-12 JPEG | Top-5 accuracy: 79.80 -> 78.26 | Hurt |
| Add data-dependent power exponents (HPOW) (ablation) | ResNet-50 · ILSVRC-12 clean | Top-5 accuracy: 93.40 -> 93.70 | Helped |
| Add learnable-exponent power function to the last three convolution layers | ZF detector · Pascal VOC 2007 validation, Mix.Light | mAP: 50.0 -> 52.2 | Helped |
| Add learnable-exponent power function to the last three convolution layers | ZF detector · Pascal VOC 2007 validation, Mix.Heavy | mAP: 14.3 -> 17.6 | Helped |
| Add learnable-exponent power function to the last three convolution layers | ZF detector · Pascal VOC 2007 validation, original | mAP: 58.7 -> 58.7 | No clear effect |
| Use feature quantization nonlinearity rather than the base network (ablation) | ResNet-18 · ILSVRC-12 JPEG | Top-5 accuracy: 59.47 -> 62.92 | Helped |
| Use learnable-exponent power function (POW-1) rather than the base network (ablation) | ResNet-18 · ILSVRC-12 chroma Gaussian noise (CC+N) | Top-5 accuracy: 59.96 -> 67.70 | Helped |
| Combine floor and power functions (SF-POW) (ablation) | ResNet-50 · ILSVRC-12 JPEG | Top-5 accuracy: 79.80 -> 79.14 | Hurt |
| Use DoReFa quantization | ResNet-50 · ILSVRC-12 salt and pepper | Top-5 accuracy: 69.94 -> 41.40 | Hurt |
Closest papers by meaning
- Robust Quantization: One Model to Rule Them All (NeurIPS 2020 · similarity 0.87)
- SeerNet: Predicting Convolutional Neural Network Feature-Map Sparsity Through Low-Bit Quantization (CVPR 2019 · similarity 0.87)
- Vector Quantization With Self-Attention for Quality-Independent Representation Learning (CVPR 2023 · similarity 0.86)
- Generalisation in humans and deep neural networks (NeurIPS 2018 · similarity 0.86)
- Gradient $\ell_1$ Regularization for Quantization Robustness (ICLR 2020 · similarity 0.86)
- Implicit Feature Decoupling With Depthwise Quantization (CVPR 2022 · similarity 0.86)
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.