PyTorch implementation of paper "StarEnhancer: Learning Real-Time and Style-Aware Image Enhancement" (ICCV 2021 Oral)

Overview

StarEnhancer

StarEnhancer: Learning Real-Time and Style-Aware Image Enhancement (ICCV 2021 Oral)

Abstract: Image enhancement is a subjective process whose targets vary with user preferences. In this paper, we propose a deep learning-based image enhancement method covering multiple tonal styles using only a single model dubbed StarEnhancer. It can transform an image from one tonal style to another, even if that style is unseen. With a simple one-time setting, users can customize the model to make the enhanced images more in line with their aesthetics. To make the method more practical, we propose a well-designed enhancer that can process a 4K-resolution image over 200 FPS but surpasses the contemporaneous single style image enhancement methods in terms of PSNR, SSIM, and LPIPS. Finally, our proposed enhancement method has good interactability, which allows the user to fine-tune the enhanced image using intuitive options.

StarEnhancer

Getting started

Install

We test the code on PyTorch 1.8.1 + CUDA 11.1 + cuDNN 8.0.5, and close versions also work fine.

pip install -r requirements.txt

We mainly train the model on RTX 2080Ti * 4, but a smaller mini batch size can also work.

Prepare

You can generate your own dataset, or download the one we generate.

The final file path should be the same as the following:

┬─ save_model
│   ├─ stylish.pth.tar
│   └─ ... (model & embedding)
└─ data
    ├─ train
    │   ├─ 01-Experts-A
    │   │   ├─ a0001.jpg
    │   │   └─ ... (id.jpg)
    │   └─ ... (style folder)
    ├─ valid
    │   └─ ... (style folder)
    └─ test
        └─ ... (style folder)

Download

Data and pretrained models are available on GoogleDrive.

Generate

  1. Download raw data from MIT-Adobe FiveK Dataset.
  2. Download the modified Lightroom database fivek.lrcat, and replace the original database with it.
  3. Generate dataset in JPEG format with quality 100, which can refer to this issue.
  4. Run generate_dataset.py in data folder to generate dataset.

Train

Firstly, train the style encoder:

python train_stylish.py

Secondly, fetch the style embedding for each sample in the train set:

python fetch_embedding.py

Lastly, train the curve encoder and mapping network:

python train_enhancer.py

Test

Just run:

python test.py

Testing LPIPS requires about 10 GB GPU memory, and if an OOM occurs, replace the following lines

lpips_val = loss_fn_alex(output * 2 - 1, target_img * 2 - 1).item()

with

lpips_val = 0

Notes

Due to agreements, we are unable to release part of the source code. This repository provides a pure python implementation for research use. There are some differences between the repository and the paper as follows:

  1. The repository uses a ResNet-18 w/o BN as the curve encoder's backbone, and the paper uses a more lightweight model.
  2. The paper uses CUDA to implement the color transform function, and the repository uses torch.gather to implement it.
  3. The repository removes some tricks used in training lightweight models.

Overall, this repository can achieve higher performance, but will be slightly slower.

Comments
  • Multi-style, unpaired setting

    Multi-style, unpaired setting

    您好,在多风格非配对图场景,能否交换source和target的位置,并将得到的output_A和output_B进一步经过enhancer,得到recover_A和recover_B。最后计算l1_loss(source, recover_A)和l1_loss(target, recover_B)及Triplet_loss(output_A,target, source) 和 Triplet_loss(output_B,source,target)

    def train(train_loader, mapping, enhancer, criterion, optimizer):
        losses = AverageMeter()
        criterionTriplet = torch.nn.TripletMarginLoss(margin=1.0, p=2)
        FEModel = Feature_Extract_Model().cuda()
    
        mapping.train()
        enhancer.train()
    
        for (source_img, source_center, target_img, target_center) in train_loader:
            source_img = source_img.cuda(non_blocking=True)
            source_center = source_center.cuda(non_blocking=True)
            target_img = target_img.cuda(non_blocking=True)
            target_center = target_center.cuda(non_blocking=True)
    
            style_A = mapping(source_center)
            style_B = mapping(target_center)
    
            output_A = enhancer(source_img, style_A, style_B)
            output_B = enhancer(target_img, style_B, style_A)
            recoverA = enhancer(output_A, style_B, style_A)
            recoverB = enhancer(output_B, style_A, style_B)
    
            source_img_feature = FEModel(source_img)
            target_img_feature = FEModel(target_img)
            output_A_feature = FEModel(output_A)
            output_B_feature = FEModel(output_B)
    
            loss_l1 = criterion(recoverA, source_img) + criterion(recoverB, target_img)
            loss_triplet = criterionTriplet(output_B_feature, source_img_feature, target_img_feature) + \
                           criterionTriplet(output_A_feature, target_img_feature, source_img_feature)
            loss = loss_l1 + loss_triplet
    
            losses.update(loss.item(), args.t_batch_size)
    
            optimizer.zero_grad()
            loss.backward()
            optimizer.step()
    
        return losses.avg
    
    opened by jxust01 4
  • Questions about dataset preparation

    Questions about dataset preparation

    您好,我想用您的工程跑一下自己的数据,现在有输入,输出一组数据对,训练数据里面A-E剩下的4种效果是怎样生成的呢,这些目标效果数据能否是非成对的呢?如果只有一种风格,能否A-E目标效果都拷贝成一样的数据呢,在train_enhancer.py所训练的单风格脚本是需要embeddings.npy文件,这个文件在单风格训练时是必须的吗

    opened by zener90818 4
  • Dataset processing

    Dataset processing

    你好,我在您提供的fivek.lrcat没找到 DeepUPE issue里的"(default) input with ExpertC"。请问单风格实验的输入是下图中的“InputAsShotZeroed”还是“(Q)InputZeroed with ExpertC WhiteBalance” image

    opened by madfff 2
  • Configure Renovate

    Configure Renovate

    WhiteSource Renovate

    Welcome to Renovate! This is an onboarding PR to help you understand and configure settings before regular Pull Requests begin.

    🚦 To activate Renovate, merge this Pull Request. To disable Renovate, simply close this Pull Request unmerged.


    Detected Package Files

    • requirements.txt (pip_requirements)

    Configuration Summary

    Based on the default config's presets, Renovate will:

    • Start dependency updates only once this onboarding PR is merged
    • Enable Renovate Dependency Dashboard creation
    • If semantic commits detected, use semantic commit type fix for dependencies and chore for all others
    • Ignore node_modules, bower_components, vendor and various test/tests directories
    • Autodetect whether to pin dependencies or maintain ranges
    • Rate limit PR creation to a maximum of two per hour
    • Limit to maximum 20 open PRs at any time
    • Group known monorepo packages together
    • Use curated list of recommended non-monorepo package groupings
    • Fix some problems with very old Maven commons versions
    • Ignore spring cloud 1.x releases
    • Ignore http4s digest-based 1.x milestones
    • Use node versioning for @types/node
    • Limit concurrent requests to reduce load on Repology servers until we can fix this properly, see issue 10133

    🔡 Would you like to change the way Renovate is upgrading your dependencies? Simply edit the renovate.json in this branch with your custom config and the list of Pull Requests in the "What to Expect" section below will be updated the next time Renovate runs.


    What to Expect

    With your current configuration, Renovate will create 1 Pull Request:

    Pin dependency torch to ==1.10.0
    • Schedule: ["at any time"]
    • Branch name: renovate/pin-dependencies
    • Merge into: main
    • Pin torch to ==1.10.0

    ❓ Got questions? Check out Renovate's Docs, particularly the Getting Started section. If you need any further assistance then you can also request help here.


    This PR has been generated by WhiteSource Renovate. View repository job log here.

    opened by renovate[bot] 1
  • The results are not the same as the paper

    The results are not the same as the paper

    I am the author.

    Some peers have emailed me asking about the performance of the open source model that does not agree with the results in the paper. As stated in the README, the model is not the model of the paper, but the performance is similar. The exact result should be: PSNR: 25.41, SSIM: 0.942, LPIPS: 0.085

    If you find that your result is not this, then it may be that the JPEG codec is different, which is related to the version of opencv and how it is installed.

    You can uninstall your opencv (either with pip or conda) and reinstall it using pip (it must be pip, because conda installs a different JPEG codec):

    pip install opencv-python==4.5.5.62​
    
    opened by IDKiro 0
Owner
IDKiro
Stroll in the abyss
IDKiro
Tgbox-bench - Simple TGBOX upload speed benchmark

TGBOX Benchmark This script will benchmark upload speed to TGBOX storage. Build

Non 1 Jan 09, 2022
Fully Convlutional Neural Networks for state-of-the-art time series classification

Deep Learning for Time Series Classification As the simplest type of time series data, univariate time series provides a reasonably good starting poin

Stephen 572 Dec 23, 2022
Implementation of SETR model, Original paper: Rethinking Semantic Segmentation from a Sequence-to-Sequence Perspective with Transformers.

SETR - Pytorch Since the original paper (Rethinking Semantic Segmentation from a Sequence-to-Sequence Perspective with Transformers.) has no official

zhaohu xing 112 Dec 16, 2022
A BaSiC Tool for Background and Shading Correction of Optical Microscopy Images

BaSiC Matlab code accompanying A BaSiC Tool for Background and Shading Correction of Optical Microscopy Images by Tingying Peng, Kurt Thorn, Timm Schr

Marr Lab 34 Dec 18, 2022
Crowd-Kit is a powerful Python library that implements commonly-used aggregation methods for crowdsourced annotation and offers the relevant metrics and datasets

Crowd-Kit: Computational Quality Control for Crowdsourcing Documentation Crowd-Kit is a powerful Python library that implements commonly-used aggregat

Toloka 125 Dec 30, 2022
Bulk2Space is a spatial deconvolution method based on deep learning frameworks

Bulk2Space Spatially resolved single-cell deconvolution of bulk transcriptomes using Bulk2Space Bulk2Space is a spatial deconvolution method based on

Dr. FAN, Xiaohui 60 Dec 27, 2022
Official implementation for NIPS'17 paper: PredRNN: Recurrent Neural Networks for Predictive Learning Using Spatiotemporal LSTMs.

PredRNN: A Recurrent Neural Network for Spatiotemporal Predictive Learning The predictive learning of spatiotemporal sequences aims to generate future

THUML: Machine Learning Group @ THSS 243 Dec 26, 2022
Compare neural networks by their feature similarity

PyTorch Model Compare A tiny package to compare two neural networks in PyTorch. There are many ways to compare two neural networks, but one robust and

Anand Krishnamoorthy 181 Jan 04, 2023
All public open-source implementations of convnets benchmarks

convnet-benchmarks Easy benchmarking of all public open-source implementations of convnets. A summary is provided in the section below. Machine: 6-cor

Soumith Chintala 2.7k Dec 30, 2022
Source code for the paper "PLOME: Pre-training with Misspelled Knowledge for Chinese Spelling Correction" in ACL2021

PLOME:Pre-training with Misspelled Knowledge for Chinese Spelling Correction (ACL2021) This repository provides the code and data of the work in ACL20

197 Nov 26, 2022
A Repository of Community-Driven Natural Instructions

A Repository of Community-Driven Natural Instructions TLDR; this repository maintains a community effort to create a large collection of tasks and the

AI2 244 Jan 04, 2023
Python Actor concurrency library

Thespian Actor Library This library provides the framework of an Actor model for use by applications implementing Actors. Thespian Site with Documenta

Kevin Quick 177 Dec 11, 2022
SoK: Vehicle Orientation Representations for Deep Rotation Estimation

SoK: Vehicle Orientation Representations for Deep Rotation Estimation Raymond H. Tu, Siyuan Peng, Valdimir Leung, Richard Gao, Jerry Lan This is the o

FIRE Capital One Machine Learning of the University of Maryland 12 Oct 07, 2022
Connecting Java/ImgLib2 + Python/NumPy

imglyb imglyb aims at connecting two worlds that have been seperated for too long: Python with numpy Java with ImgLib2 imglyb uses jpype to access num

ImgLib2 29 Dec 21, 2022
PyTorch implementation of UNet++ (Nested U-Net).

PyTorch implementation of UNet++ (Nested U-Net) This repository contains code for a image segmentation model based on UNet++: A Nested U-Net Architect

4ui_iurz1 642 Jan 04, 2023
MicRank is a Learning to Rank neural channel selection framework where a DNN is trained to rank microphone channels.

MicRank: Learning to Rank Microphones for Distant Speech Recognition Application Scenario Many applications nowadays envision the presence of multiple

Samuele Cornell 20 Nov 10, 2022
PyTorch 1.5 implementation for paper DECOR-GAN: 3D Shape Detailization by Conditional Refinement.

DECOR-GAN PyTorch 1.5 implementation for paper DECOR-GAN: 3D Shape Detailization by Conditional Refinement, Zhiqin Chen, Vladimir G. Kim, Matthew Fish

Zhiqin Chen 72 Dec 31, 2022
A trusty face recognition research platform developed by Tencent Youtu Lab

Introduction TFace: A trusty face recognition research platform developed by Tencent Youtu Lab. It provides a high-performance distributed training fr

Tencent 956 Jan 01, 2023
A framework for Quantification written in Python

QuaPy QuaPy is an open source framework for quantification (a.k.a. supervised prevalence estimation, or learning to quantify) written in Python. QuaPy

41 Dec 14, 2022
RSC-Net: 3D Human Pose, Shape and Texture from Low-Resolution Images and Videos

RSC-Net: 3D Human Pose, Shape and Texture from Low-Resolution Images and Videos Implementation for "3D Human Pose, Shape and Texture from Low-Resoluti

XiangyuXu 42 Nov 10, 2022