Research map / Architectures and efficiency / CNN design
Fast Image Processing With Fully-Convolutional Networks
Qifeng Chen, Jia Xu, Vladlen Koltun
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 adaptive normalization instead of no normalization (ablation) | CAN32 · MIT-Adobe 5K test (avg over 10 operators) | PSNR: 35.52 -> 37.59 | Improved |
| use adaptive normalization instead of plain batch normalization (ablation) | CAN32 · MIT-Adobe 5K test (avg over 10 operators) | PSNR: 28.64 -> 37.59 | Improved |
| add batch normalization alone (no identity branch) (ablation) | CAN32 · MIT-Adobe 5K test (avg over 10 operators) | PSNR: 35.52 -> 28.64 | Hurt |
| increase depth and width (d=9,w=24 to d=10,w=32), receptive field 257 to 513 (ablation) | CAN+AN · MIT-Adobe 5K test (avg over 10 operators) | PSNR: 36.04 -> 37.59 | Improved |
| increase depth and width (d=9,w=24 to d=10,w=32) (ablation) | CAN+AN · MIT-Adobe 5K test (avg over 10 operators) | runtime (ms): 190 -> 277 | Hurt |
| train separate networks per operator instead of one shared network (ablation) | CAN32+AN · MIT-Adobe 5K test (avg over 10 operators) | PSNR: 29.86 -> 37.59 | Improved |
| use CAN32 instead of Plain network (ablation) | fully-convolutional · MIT-Adobe 5K test (avg over 10 operators) | PSNR: 32.05 -> 35.52 | Improved |
| use dilated context aggregation CAN32+AN instead of encoder-decoder | fully-convolutional · MIT-Adobe 5K test (avg over 10 operators) | PSNR: 34.90 -> 37.59 | Improved |
| use CAN24+AN instead of FCN-8s | fully-convolutional · MIT-Adobe 5K test (avg over 10 operators) | PSNR: 26.36 -> 36.04 | Improved |
| use perceptual or adversarial loss instead of MSE (ablation) | CAN · MIT-Adobe 5K test (avg over 10 operators) | approximation accuracy | No clear effect |
| use CAN approximator instead of BGU-opt | image operators · MIT-Adobe 5K test (avg over 10 operators) | PSNR: 25 -> 36 | Improved |
Closest papers by meaning
- Involution: Inverting the Inherence of Convolution for Visual Recognition (CVPR 2021 · similarity 0.89)
- Accelerating Very Deep Convolutional Networks for Classification and Detection (TPAMI 2016 · similarity 0.88)
- Approximation and Learning with Deep Convolutional Models: a Kernel Perspective (ICLR 2022 · similarity 0.88)
- Bilinear Convolutional Neural Networks for Fine-Grained Visual Recognition (TPAMI 2017 · similarity 0.88)
- An Empirical Evaluation of Current Convolutional Architectures' Ability to Manage Nuisance Location and Scale Variability (CVPR 2016 · similarity 0.88)
- KernelWarehouse: Rethinking the Design of Dynamic Convolution (ICML 2024 · similarity 0.87)
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.
