Zero-DCE and Zero-DCE++(Lite architechture for Mobile and edge Devices)
Papers Abstract
The paper presents a novel method, Zero-Reference Deep Curve Estimation (Zero-DCE), which formulates
light enhancement as a task of image-specific curve estimation with a deep network.
Our method trains a lightweight deep network, DCE-Net, to estimate pixel-wise and
high-order curves for dynamic range adjustment of a given image. The curve estimation
is specially designed, considering pixel value range, monotonicity, and differentiability.
Zero-DCE is appealing in its relaxed assumption on reference images, i.e., it does not
require any paired or unpaired data during training. This is achieved through a set of
carefully formulated non-reference loss functions, which implicitly measure the
enhancement quality and drive the learning of the network. Our method is efficient
as image enhancement can be achieved by an intuitive and simple nonlinear curve mapping.
Despite its simplicity, we show that it generalizes well to diverse lighting conditions.
Extensive experiments on various benchmarks demonstrate the advantages of our method over
state-of-the-art methods qualitatively and quantitatively. Furthermore, the potential benefits
of our Zero-DCE to face detection in the dark are discussed. We further present an
accelerated and light version of Zero-DCE, called (Zero-DCE++), that takes advantage
of a tiny network with just 10K parameters. Zero-DCE++ has a fast inference speed
(1000/11 FPS on single GPU/CPU for an image with a size of 1200*900*3) while keeping
the enhancement performance of Zero-DCE.
Check out the original Pytorch Implementation of Zero-DCE here and the original Pytorch implementation of Zero-DCE++ here
Proposed Zero-DCE Framework
The paper proposed a Zero Reference(without a label/reference image) Deep Curve Estimation network which estimates the best-fitting Light-Enhancement curve (LE-Curve) for a given image. Further the framework then maps all the pixels of the image's RGB channels by applying the best-fit curve iteratively and get the final enhanced output image.
DCE-net and DCE-net++
The paper proposes a simple CNN bases Deep neural network called DCE-net, which learns to map the input low-light image to its best-fit curve parameters maps. The network consist of 7 convolution layers with symmetrical skip concatenation. First 6 convolution layers consist of 32 filters each with kernel size of 3x3 with stride of 1 followed by RelU activation. The last convolution layer has interation
x 3 number of filters (if we set iteration to 8 it will produce 24 curve parameters maps for 8 iteration, where each iteration generates three curve parameter maps for the three RGB channels) followed by tanh activation. The proposed DCE-net architechture does not contains any max-pooling, downsampling or batch-normalization layers as it can break the relations between neighboring pixels.
DCE-net++ is the lite version of DCE-net. DCE-net is already a very light model with just 79k parameters. The main changes in DCE-net++ are:
- Instead of traditional convolutional layers, we use Depthwise separable convolutional layers which significantly reduces the total number of parameters, uses less memory and computational power. The DCE-net++ architecture has a total of 10k parameters with same architecture design as DCE-net.
- The last convolution layers has only 3 filters instead of
interation
x 3 number of filters which can be used to iteratively enhance the images.
Zero-Reference Loss Functions
The paper proposes set of zero-reference loss functions that differntiable which allows to assess the quality of enhanced image.
-
Spatial Consistency Loss The spatial consistency loss $L_{spa}$ encourages spatial coherence of the enhanced image through preserving the difference of neighboring regions between the input image and its enhanced version
-
Exposure controll loss To restrain the exposure of the enhanced image, the exposure control loss $L_{exp}$ is designed to control the exposure of the enhanced image. The exposure control loss measures the distance between the average intensity value of a local region to the well-exposedness level $E$.
-
Color Constancy loss By Following the Gray-world hypothesis that color in each sensor channel(RGB) averages to gray over the entire image, the paper proposes a color constancy loss $L_{col}$ to correct the potential diviation of color in the enhanced image.
-
Illumination Smoothness Loss To preserve the monotonicity relations between neighboring pixels, we add an illumination smoothness loss to each curve parameter map A.
Training and Testing Model
Zero-DCE and Zero-DCE++ model was created using Tensorflow 2.7.0 and Keras and trained on google colab's Tesla K80 GPU (12GB VRAM)
Dataset pipeline and Dataset used
I used Tensorflow's tf.data api to create a dataset input pipeline. Input data pipeline
dataset structure:
lol_datasetv2
βββ 100.png
βββ 101.png
βββ 102.png
βββ 103.png
βββ 109.png
βββ 10.png
βββ 95.png
βββ 96.png
βββ 97.png
βββ 98.png
βββ 99.png
βββ 9.png
0 directories, 500 files
LoL-dataset
Dataset link:Usage
- Clone this github repo
- Run
$pip install -r requirements.txt
to install required python packgages.
For training the model, run following
$ python train_model.py --help
usage: train_model.py [-h] --dataset_dir DATASET_DIR [--checkpoint_dir CHECKPOINT_DIR] [--model_type MODEL_TYPE] [--IMG_H IMG_H]
[--IMG_W IMG_W] [--IMG_C IMG_C] [--batch_size BATCH_SIZE] [--epoch EPOCH] [--learning_rate LEARNING_RATE]
[--dataset_split DATASET_SPLIT] [--logdir LOGDIR] [--iteration ITERATION]
Model training scipt for Zero-DCE models
optional arguments:
-h, --help show this help message and exit
--dataset_dir DATASET_DIR
Dataset directory
--checkpoint_dir CHECKPOINT_DIR
Checkpoint directory
--model_type MODEL_TYPE
Type of Model.should be any of: ['zero_dce', 'zero_dce_lite']
--IMG_H IMG_H Image height
--IMG_W IMG_W Image width
--IMG_C IMG_C Image channels
--batch_size BATCH_SIZE
Batch size
--epoch EPOCH Epochs
--learning_rate LEARNING_RATE
Learning rate
--dataset_split DATASET_SPLIT
Dataset split
--logdir LOGDIR Log directory
--iteration ITERATION
Post enhancing iteration
Example
!python train_model.py --dataset_dir lol_datasetv2/ \
--model_type zero_dce_lite \
--checkpoint_dir Trained_model/ \
--IMG_H 512 \
--IMG_W 512 \
--epoch 60 \
--batch_size 4 \
--iteration 6 \
Testing the model on the test dataset
$ python test_model.py --help
usage: test_model.py [-h] --model_path MODEL_PATH [--dataset_path DATASET_PATH] [--img_h IMG_H] [--img_w IMG_W] [--save_plot SAVE_PLOT]
[--load_random_data LOAD_RANDOM_DATA]
Test model on test dataset
optional arguments:
-h, --help show this help message and exit
--model_path MODEL_PATH
path to the saved model folder
--dataset_path DATASET_PATH
path to the dataset
--img_h IMG_H image height
--img_w IMG_W Image width
--save_plot SAVE_PLOT
save plot of original vs enhanced image. 0: no, 1: yes
--load_random_data LOAD_RANDOM_DATA
load random data. 0: no, 1: yes
Example
!python test_model.py --model_path Trained_model/zero_dce_lite_iter8/zero_dce_lite_200x300_iter8_60/ \
--datset_path lol_datasetv2/ \
--img_h 200 \
--img_w 300 \
--save_plot 1 \
--load_random_data 0
Inferencing on single image for enhancement
$ python single_image_enhance.py --help
usage: single_image_enhance.py [-h] --model_path MODEL_PATH --image_path IMAGE_PATH [--img_h IMG_H] [--img_w IMG_W] [--plot PLOT] [--save_result SAVE_RESULT] [--iteration ITERATION]
Single Image Enhancement
optional arguments:
-h, --help show this help message and exit
--model_path MODEL_PATH
path to tf model
--image_path IMAGE_PATH
path to image file
--img_h IMG_H image height
--img_w IMG_W image width
--plot PLOT plot enhanced image
--save_result SAVE_RESULT
save enhanced image
--iteration ITERATION
number of Post Ehnancing iterations
Example
$ python single_image_enhance.py --model_path Trained_model/zero_dce_iter6/zero_dce_200x300_iter6_30 \
--img_h 200 \
--img_w 300 \
--image_path sample_images/ low_light_outdoor.jpg \
--plot 0 \
--save_result 1 \
--iteration 6 \
Visual Results
Testset Results
1.Model: Zero-DCE, Epoch:30 , Input size:200x300, Iteration:4, Average Time: CPU-170.0 ms
2.Model: Zero-DCE, Epoch:30, Input size: 200x300, Iteration:6, Average Time: CPU-170.0 ms
3.Model: Zero-DCE, Epoch:30, Inout size: 200x300, Iteration:8, Average Time: CPU-170.0 ms
4.Model: Zero-DCE Lite, Epoch:60, Input size: 512x512, Iteration:6, Average Time: CPU-450 ms
5.Model: Zero-DCE Lite, Epoch:60, Input size: 200x300, Iteration:8, Average Time: CPU-90 ms
Enhance Image with its Alpha Maps.(Curve Parameter Maps)
Test Results on out of dataset images
low light image | Enhanced Image(Zero-DCE, epoch:60, interation:4) |
low light image | Enhanced Image(Zero-DCE, epoch:60, interation:6) |
low light image | Enhanced Image(Zero-DCE, epoch:30, interation:8) |
low light image | Enhanced Image(Zero-DCE, epoch:30, interation:6) |
low light image | Enhanced Image(Zero-DCE lite, epoch:60, interation:8) |
low light image | Enhanced Image(Zero-DCE, epoch:30, interation:8) |
low light image | Enhanced Image(Zero-DCE lite, epoch:60, interation:8) |
low light image | Enhanced Image(Zero-DCE lite, epoch:60, interation:6) |
Best SavedModel for Zero-DCE and Zero-DCE Lite
Releasing soon
Demo Apllication
Mobile Demo application of our trained model is comming soon
References
- Zero-reference deep curve estimation for low-light image enhancement
- Learning to Enhance Low-Light Image via Zero-Reference Deep Curve Estimation (Zero-DCE++)
- A Basic Introduction to Separable Convolutions
Citation
Paper: Zero-DCE
@Article{Zero-DCE,
author = {Guo, Chunle and Li, Chongyi and Guo, Jichang and Loy, Chen Change and Hou,
Junhui and Kwong, Sam and Cong Runmin},
title = {Zero-reference deep curve estimation for low-light image enhancement},
journal = {CVPR},
pape={1780-1789},
year = {2020}
}
Paper: Zero-DCE++
@Article{Zero-DCE++,
author ={Li, Chongyi and Guo, Chunle and Loy, Chen Change},
title = {Learning to Enhance Low-Light Image via Zero-Reference Deep Curve Estimation},
journal = {IEEE Transactions on Pattern Analysis and Machine Intelligence},
pape={},
year = {2021},
doi={10.1109/TPAMI.2021.3063604}
}
Dataset
@inproceedings{Chen2018Retinex,
title={Deep Retinex Decomposition for Low-Light Enhancement},
author={Chen Wei, Wenjing Wang, Wenhan Yang, Jiaying Liu},
booktitle={British Machine Vision Conference},
year={2018},
}