Review of low-illuminance image enhancement algorithm based on deep learning
-
摘要:
在弱光条件下拍摄的图像往往存在亮度和对比度较低、颜色失真和噪声较大等特点，严重影响人眼的主观效果，极大地限制了高阶视觉任务的性能。低照度图像增强(low illuminance image enhancement, LIIE)旨在改善这类图像的视觉效果，为后续处理提供有利条件。在诸多低照度图像增强算法中，基于深度学习的低照度图像增强成为最新的解决方案。首先梳理了基于深度学习的低照度图像增强的代表性方法；其次介绍了现有低照度图像数据集、损失函数和评价指标；再次通过基准测试与实验分析，进一步对现有基于深度学习的低照度图像增强算法进行全面评估；最后对目前研究进行总结，并对低照度图像增强的发展方向进行讨论和展望。
Abstract:Images captured under low-light conditions are often characterized by low brightness and contrast, color distortion, and high noise, which seriously affect the subjective vision of human eyes and greatly limit the performance of higher-order vision tasks. Low illuminance image enhancement (LIIE) aims to improve the visual effect of such images and provide favorable conditions for subsequent processing. Among many low-illuminance image enhancement algorithms, the LIIE based on deep learning has become the latest solution. Firstly, the representative methods for LIIE based on deep learning were reviewed. Secondly, the existing low-illuminance image datasets, loss functions, and evaluation indicators were introduced. Thirdly, the existing LIIE algorithms based on deep learning were comprehensively evaluated through benchmark testing and experimental analysis. Finally, a summary of current research was provided, and the development direction of LIIE was discussed and prospected.
-
Keywords:
- low-illuminance images /
- image enhancement /
- deep learning /
- loss function /
- benchmark testing
-
引言
英伟达TensorRT是一种高性能神经网络推理（Inference）引擎，是一个标准C++库。TensorRT只能用来做Inference（推理），不能用来进行训练，用于在生产环境中部署深度学习应用程序[1]。应用领域包括图像分类、分割和目标检测等，可提供最大的推理吞吐量和效率。TensorRT需要CUDA（compute unified device architecture）的支持，包含一个为优化生产环境中部署的深度学习模型而创建的库，可获取经过训练的神经网络（通常使用32位或16位数据），并针对降低精度的INT8运算来优化这些网络。借助CUDA的可编程性，TensorRT将能够应对深度神经网络日益多样化、复杂化的趋势。TensorRT在不断的改进过程中，在保证软件精度的同时，自动优化训练过的神经网络，不断提高速度。TensorRT能够支持Caffe等主流深度学习框架[2-3]。
本文实现了一个利用TensorRT执行智能视频分析的典型应用，演示了使用片上解码器进行解码，使用片上转换器进行视频缩放，利用TensorRT执行对象标识，利用OpenGL2和XWindow-11进行渲染，并在标识的对象周围生成包围框。此外，还使用视频转换器函数进行各种格式转换。使用EGLImage来演示缓冲区共享和图像显示。图1展示了使用TensorRT的流程细节[4-5]。
1 定义视频处理结构体
本文例程将本地存储的H.264视频文件进行解码、格式转换和渲染，为了使流程清晰，便于管控，定义context_t结构管理视频处理全部资源。如表1所示成员主要包括一个解码器类、一个转换器类、一个渲染器类、一个数据指针。解码器类NvVideoDecoder封装了用于视频解码的多媒体API函数，用于从H.264视频文件解码压缩的视频。转换器类NvVideoConverter封装了视频转换的相关函数，包括色彩空间变换、尺度变换以及软硬件缓存空间的变换。渲染器类NvEglRenderer类封装了图像渲染的相关函数以及XWindow-11以及OpenGL2的部分函数，使用EGL和OpenGL ES 2.0进行呈现。渲染器需要缓冲区的文件描述符（FD）作为输入，创建自己的X窗口。渲染速率（以帧/s为单位），窗口的宽度、高度、水平偏移量和垂直偏移量都是可配置的。所有EGL调用只能通过一个线程进行。该类在内部创建一个线程，该线程执行所有EGL/GL初始化，从FD获取EGLImage对象，然后反初始化所有EGL/GL结构[6-9]。
表 1 Context_t结构主要成员Table 1. Main members of Context_t structure成员 描述 NvVideoDecoder 包含视频解码相关的成员和函数 NvVideoConverter 包含视频格式转换相关的成员和函数 NvEglRenderer 包含EGL显示渲染相关函数 EGLImageKHR EGLImage图像数据指针，用于CUDA处理，这个类型来源于EGL开源库 2 利用TRT_Context类进行加速推理
英伟达提供的TRT_Context类包含一系列接口来加载Caffe模型并执行推理。表2描述了本示例中使用的关键TRT_Context成员。本文中使用buildTrtContext实现Caffe模型到TensorRT模型转换，它的输入参数包括caffe网络结构文件和模型参数文件。实际上也可借助转换工具实现Caffe模型到gie模型的转换。TRT_Context:: getNumTrtInstances用于获取加速上下文的实例。TRT_Context::doInference用转换好的模型利用TensorRT进行加速推理。此外，TRT_Context还实现了一些模型控制和剪裁的一些函数接口[10-13]。
表 2 TRT_Context类主要成员Table 2. Main members of TRT_Context classTRT_Context类成员 描述 TRT_Context::buildTrtContext 构建Tensorrt上下文 TRT_Context::getNumTrtInstances 获取TRT_context 实例. TRT_Context::doInference TensorRT 推理接口 3 主进程
主进程调用以上定义的类和结构实现整个处理流程，主要代码如下：
TRT_Context g_trt_context;
main(int argc, char *argv[])
{
//程序入口参数处理
context_t ctx[CHANNEL_NUM];
global_cfg cfg;
char **argp;
set_globalcfg_default(&cfg);
argp = argv;
parse_global(&cfg, argc, &argp);
parse_csv_args(&ctx[0], &g_trt_context, argc-cfg.channel_num-1, argp)；
//设置g_trt_context参数
g_trt_context.setModelIndex(TRT_MODEL);
g_trt_context.buildTrtContext(cfg.deployfile, cfg.modelfile, true);
pthread_create(&TRT_Thread_handle, NULL, trt_thread, NULL);
// 获取EGL默认值
egl_display = eglGetDisplay(EGL_DEFAULT_DISPLAY);
// EGL 初始化
eglInitialize(egl_display, NULL, NULL)
for (iterator = 0; iterator < cfg.channel_num; iterator++)
{
int ret = 0;
sem_init(&(ctx[iterator].dec_run_sem), 0, 0);
set_defaults(&ctx[iterator]);
char decname[512];
sprintf(decname, "dec%d", iterator);
ctx[iterator].channel = iterator;
ctx[iterator].in_file_path = cfg.in_file_path[iterator];
ctx[iterator].nvosd_context = nvosd_create_context();
//创建解码器
ctx[iterator].dec = NvVideoDecoder::createVideoDecoder(decname);
//设置输出面板格式
ctx[iterator].dec->setOutputPlaneFormat(ctx[iterator].decoder_pixfmt, CHUNK_SIZE);
//映射输出面板缓存
ctx[iterator].dec->output_plane.setupPlane(V4L2_MEMORY_MMAP, 10, true, false);
//创建渲染线程
pthread_create(&ctx[iterator].render_feed_handle, NULL, render_thread, &ctx[iterator]);
char convname[512];
// 创建BL到 PL转换器
ctx[iterator].conv = NvVideoConverter::createVideoConverter(convname);
ctx[iterator].conv->output_plane.setDQThreadCallback(conv_output_dqbuf_thread_callback);
ctx[iterator].conv->capture_plane.setDQThreadCallback(conv_capture_dqbuf_thread_callback);
if (ctx[iterator].cpu_occupation_option!= PARSER)
pthread_create(&ctx[iterator].dec_capture_loop, NULL, dec_capture_loop_fcn, &ctx[iterator]);
pthread_create(&ctx[iterator].dec_feed_handle, NULL, dec_feed_loop_fcn, &ctx[iterator]);
//等待解码器获取EOS
sem_wait(&(ctx[iterator].dec_run_sem));
//向渲染器发送命令
ctx[iterator].stop_render = 1;
pthread_cond_broadcast(&ctx[iterator].render_cond);
pthread_join(ctx[iterator].render_feed_handle, NULL);
}
}
4 测试与分析
在这个示例中，对象检测仅限于在960×540分辨率的视频流中识别汽车。该网络基于GoogleNet。推理是在逐帧的基础上进行的，不涉及任何对象跟踪，展示了如何使用TensorRT快速构建计算管道。示例使用训练过的GoogleNet网络，它是用NVIDIA深度学习GPU训练系统（DIGITS）训练的。训练是大约3 000帧从1.5 m（5英尺）~3 m（10英尺）的高度拍摄的。根据输入的视频样本，预计会有不同程度的检测精度。运行程序对H.264本地视频进行测试，TensorRT能够成功运行，实时识别目标图像[14-15]，测试效果如图2所示。
其运行性能如下：
FP32 run:400 batches of size 100 starting at 100
........................................
Top1: 0.9904, Top5: 1
Processing 40000 images averaged 0.00157702 ms/image and 0.157702 ms/batch.
FP16 run:400 batches of size 100 starting at 100
Engine could not be created at this precision
INT8 run:400 batches of size 100 starting at 100
........................................
Top1: 0.9908, Top5: 1
Processing 40000 images averaged 0.00122583 ms/image and 0.122583 ms/batch.
可以看到这个例子中采用int8量化时，提速可以达到20%以上，对于大计算量的应用，提速效果更好。推理（Inference）可以使用低精度的技术，训练的时候因为要保证前后向传播，每次梯度的更新是很微小的，这个时候需要相对较高的精度，一般来说需要float型，如FP32，32位的浮点型来处理数据。但是在推理的时候，对精度的要求没有那么高，很多研究表明可以用低精度，如半长（16字节）的FP16，也可以用8位的整型INT8来做推理，结果没有特别大的精度损失。低精度计算的好处是一方面可以减少计算量，原来计算32位的单元处理FP16的时候，理论上可以达到2倍的速度，处理INT8的时候理论上可以达到4倍的速度。另一方面是模型需要的空间减少，不管是权值的存储还是中间值的存储，应用更低的精度，模型大小会相应减小。
TensorRT的运行效果与GPU的硬件性能和采用的网络结构直接相关，量化标准仅仅是其中一个影响因素，不同的硬件和网络结构也会带来不同程度的速度提升。为了对比上述例子的加速性能，图3展示了采用GPU卡V100对ResNet网络进行TensorRT加速的实际效果。
![]()
图 3 V100卡+ResNet网络使用TensorRT推理的效果Figure 3. Effect of TensorRT reasoning in V100 card+ResNet network这是一个比较极端的例子，该例中使用的是先进的GPU卡V100，V100添加了专门针对深度学习优化的TensorCore，TensorCore可以完成4×4矩阵的半精度乘法，也就是可以完成一个4×4的FP16矩阵和另外一个4×4的FP16矩阵相乘，当然可以再加一个矩阵（FP16 或FP32），得到一个FP32或者FP16的矩阵的过程。TensorCore在V100上理论峰值可以达到120 Tflops。如果只是用CPU来做推理，首先它的吞吐只能达到140，也就是说每秒只能处理140张图片，同时整个处理过程需要有14 ms的延迟，也就是说用户提交请求后，推理阶段最快需要14 ms才能返回结果；如果使用V100，在TensorFlow中去做推理，大概是6.67 ms的延时，但是吞吐只能达到305；如果使用V100加TensorRT，在保证延迟不变的情况下，吞吐可以提高15倍，高达5 700张图片帧/s。可以看到随着GPU性能的提升，以及网络结构的复杂化，TensorRT对推理速度的提升非常明显，对于大数据应用是一个很好的选择。目标英伟达公司已经将TensorRT项目部分开源，这势必会使TensorRT得到更好的推广应用。
-
![]()
图 1 配对低光数据集的低照度图像示例(第1行为弱光图像，第2行为参考图像)
Figure 1. Examples of low-illuminance images in some paired low-light datasets (row 1 refers to low-light images, and row 2 refers to reference images)
![]()
图 2 无配对低光数据集的低照度图像示例
Figure 2. Examples of low-illuminance images in some unpaired low-light datasets
![]()
图 3 不同方法在MIT-Adobe FiveK-test数据集中的低照度图像增强效果对比
Figure 3. Contrast of low-illuminance images enhancement effect in MIT-Adobe FiveK-test dataset with different methods
![]()
图 4 不同方法在LOL-test数据集中的低照度图像增强效果对比
Figure 4. Contrast of low-illuminance images enhancement effect in LOL-test dataset with different methods
表 1 基于深度学习的代表性低照度图像增强方法
Table 1 Representative low-illuminance image enhancement methods based on deep learning
Method Learinng Network structure(model) Loss function Format Code Publication Project address LLNet[14] SL Stacked sparse denoising autoencoder SRR loss RGB Theano PR(2017) https://github.com/kglore/llnet_color LightenNet[27] SL Four layers $ {L}_{2} $ loss RGB Caffe PR(2018) https://github.com/Li-Chongyi/low-light-codes MBLLEN[19] SL Feature extraction module;
Enhancement module;
Fusion moduleSSIM loss; Region loss;
Perceptual lossRGB TensorFlow BNVC(2018) https://github.com/Lvfeifan/MBLLEN RetinexNet[26] SL Multi-scale network $ {L}_{1} $ loss; Smoothness loss;
Invariable reflectance lossRGB TensorFlow BMVC(2018) https://github.com/weichen582/RetinexNet SICE[40] SL Low frequency luminance;
High frequency detail$ {L}_{1} $ loss; $ {L}_{2} $ loss;
SSIM lossRGB Caffe TIP(2018) https://github.com/csjcai/SICE SID[20] SL Amplification ratio $ {L}_{1} $ loss RAW TensorFlow CVPR(2018) https://github.com/cchen156/Learning-to-See-in-the-Dark SMD[41] SL Filtered results;
Siamese networkRecovery loss; Self-Consistency loss RAW TensorFlow ICCV(2019) https://github.com/cchen156/Seeing-Motion-in-the-Dark SMOID[42] SL 3D U-Net $ {L}_{1} $ loss RAW TensorFlow ICCV(2019) https://github.com/MichaelHYJiang/Learning-to-See-Moving-Objects-in-the-Dark DeepUPE[23] SL Illumination map $ {L}_{1} $ loss; Color loss;
Smoothness lossRGB TensorFlow ACM(2019) https://github.com/dvlab-research/DeepUPE EnlightenGAN[30] UL Attention map;
Self-regularzationAdversarial loss;
Self feature preserving lossRGB PyTorch arXiv(2019) https://github.com/VITA-Group/EnlightenGAN KinD[29] SL Reflectance layers $ {L}_{1} $ loss; SSIM loss; Reflectance similarity loss; $ {L}_{2} $ loss; smoothness loss; RGB TensorFlow ACMMM(2019) https://github.com/zhangyhuaee/KinD ExCNet[37] ZSL Fully connected layers Energy minimization loss RGB PyTorch ACMMM(2019) https://cslinzhang.github.io/ExCNet/ DSLR[22] SL Laplacian pyramid;
U-Net like network$ {L}_{2} $ loss; Color loss;
Laplacian lossRGB PyTorch IEEE(2020) https://github.com/SeokjaeLIM/DSLR-release TBEFN[44] SL Three stages;
U-Net like networkSSIM loss; Perceptual loss
Smoothness loss;RGB PyTorch IEEE(2020) https://github.com/lukun199/TBEFN Zero-DCE[39] ZSL Fully connected network Illumination smoothness loss; Spatial consistency loss; Color constancy loss RGB PyTorch CVPR(2020) https://github.com/soumik12345/Zero-DCE DRBN[35] SSL Recursive network SSIM loss; Perceptual loss;
Adversarial lossRGB PyTorch CVPR(2021) https://github.com/flyywh/CVPR-2020-Semi-Low-Light Retinex DIP[45] ZSL Encoder-decoder network Reflectance loss; Smoothness loss RGB PyTorch IEEE(2021) https://github.com/zhaozunjin/RetinexDIP RUAS[46] ZSL Neural architecture search Cooperative loss; Similar
Loss; Total variation lossRGB PyTorch CVPR (2021) https://github.com/KarelZhang/RUAS SCI[34] UL Self-Calibrated Illumination Fidelity loss;
Smoothness lossRGB PyTorch CVPR(2022) https://github.com/tengyu1998/SCI SNR-aware[43] SL SNR-guided attention Charbonnier loss; Perceptual loss RGB PyTorch CVPR(2022) https://github.com/dvlab-research/SNR-Aware-Low-Light-Enhance Dimma[36] SSL Mixture density network;
U-Net like networkMean squared error loss;
Perceptual lossRGB PyTorch arXiv(2023) https://github.com/WojciechKoz/Dimma PairLIE[15] UL Encoder-decoder networks Projection Loss; Retinex Loss; Reflectance consistency loss RGB PyTorch CVPR(2023) https://github.com/zhenqifu/PairLIE CUE[62] SL Masked autoencoder;
Customized learnable priorsIllumination smoothness loss; Noise prior loss RGB PyTorch ICCV(2023) https://github.com/zheng980629/CUE 
下载: 导出CSV
表 2 低照度图像增强数据集总结
Table 2 Summary of low-illuminance image enhancement dataset
Dataset Date Number Resolution/pixel Format Real/
SyntheticPaired Download link LIME[47] 2017 10 326×326～2 000× 15000 RGB Real No https://github.com/estija/LIME NPE[12] 2013 85 267×304～749×492 RGB Real No https://github.com/Spirals-Team/npe-dataset DICM[48] 2013 64 481×321 RGB Real No https://github.com/JoshuaEbenezer/LDR ExDark[49] 2019 7363 500×332～ 1600 ×1066 RGB Real No https://github.com/cs-chan/Exclusively-Dark-Image-Dataset VE-LOL-H[50] 2021 10940 1080 ×720RGB Real No https://flyywh.github.io/IJCV2021LowLight_VELOL/ SID[20] 2018 5094 4240 ×2832 or6000 ×4000 RAW Real Yes https://github.com/cchen156/Learning-to-See-in-the-Dark LOL[26] 2018 789 400×600 RGB Real Yes https://daooshee.github.io/BMVC2018website/ SICE[40] 2018 4413 3000 ×2 000 or6000 ×4 000RGB Real Yes https://github.com/csjcai/SICE MIT-Adobe
Fivek[51]2011 5000 $ 1\;440\times 2\;160 $～ 6048 ×4032 RAW Real+
SyntheticYes https://data.csail.mit.edu/graphics/fivek/ DRV[41] 2019 202 3672 ×5496 RAW Real Yes https://github.com/cchen156/Seeing-Motion-in-the-Dark VE-LOL-L[50] 2021 2500 400×600 RGB Real+
SyntheticYes https://flyywh.github.io/IJCV2021LowLight_VELOL/ UHD-LOL[52] 2023 11065 4000 $ \times 4\;000 $ or8000 ×8000 RGB Real Yes https://github.com/TaoWangzj/LLFormer
下载: 导出CSV
表 3 在MIT-Adobe FiveK-test数据集上的性能比较
Table 3 Performance comparison on MIT-Adobe FiveK-test dataset
Metnod MSE↓ PSNR↑ SSIM↑ NIQE↓ BRISQUE↓ input[51] 1.723 14.825 0.764 6.124 31.183 LLNet[14] 4.241 9.698 0.473 6.688 38.794 lightenNet[27] 4.172 15.139 0.635 6.971 29.159 MBLLEN[19] 1.267 14.965 0.863 6.947 32.318 RetinexNet[26] 3.943 11.333 0.492 4.036 26.355 KinD[29] 1.609 14.054 0.558 4.217 35.661 TBEFN[44] 3.690 9.760 0.461 5.308 30.181 EnlightenGAN[30] 3.837 14.598 0.793 3.915 31.345 SCI[34] 3.608 14.820 0.723 3.801 29.395 ExCNet[37] 2.927 12.698 0.473 7.648 38.794 Zero-DCE[39] 3.360 12.416 0.736 4.037 31.637 RRDNet[38] 6.199 9.966 0.416 5.173 37.278 RUAS[46] 3.376 9.588 0.460 4.194 27.971 DRBN[35] 3.410 13.639 0.754 4.361 38.904
下载: 导出CSV
表 4 在LOL-test数据集上的性能比较
Table 4 Performance comparison on LOL-test dataset
Metnod MSE↓ PSNR↑ SSIM↑ NIQE↓ BRISQUE↓ input[26] 9.712 7.951 0.131 7.948 42.153 LLNet[14] 1.080 16.583 0.715 5.453 38.549 lightenNet[27] 7.993 11.876 0.428 5.394 13.440 MBLLEN[19] 1.172 18.257 0.723 5.318 12.238 RetinexNet[26] 1.714 16.141 0.464 5.702 38.681 KinD[29] 1.133 16.240 0.704 4.764 27.438 TBEFN[44] 1.445 16.002 0.730 3.983 10.956 EnlightenGAN[30] 1.824 15.997 0.715 3.720 11.070 SCI[34] 1.503 15.654 0.491 3.308 14.863 ExCNet[37] 2.752 15.370 0.538 4.648 19.134 Zero-DCE[39] 2.851 15.278 0.535 5.596 15.408 RRDNet[38] 5.993 13.982 0.461 4.909 14.071 RUAS[46] 3.144 14.667 0.416 5.625 12.287 DRBN[35] 2.622 15.047 0.432 5.121 22.781
下载: 导出CSV
-
[1] YIN D, TANG W, CHEN P, et al. An improved algorithm for target detection in low light conditions[J]. Journal of Physics: Conference Series, 2022, 2203(1): 012045. doi: 10.1088/1742-6596/2203/1/012045
[2] WANG W, YUAN X, WU X, et al. Fast image dehazing method based on linear transformation[J]. IEEE Transactions on Multimedia, 2017, 19(6): 1142-1155. doi: 10.1109/TMM.2017.2652069
[3] KIM W, SUH S, HWANG W, et al. SVD face: illumination-invariant face representation[J]. IEEE Signal Processing Letters, 2014, 21(11): 1336-1340. doi: 10.1109/LSP.2014.2334656
[4] WANG W, WU X, YUAN X, et al. An experiment-based review of low-light image enhancement methods[J]. IEEE Access, 2020, 8: 87884-87917. doi: 10.1109/ACCESS.2020.2992749
[5] GONZALEZ R C. Digital image processing[M]. New York: Pearson Education India, 2009.
[6] FU X, ZENG D, HUANG Y, et al. A fusion-based enhancing method for weakly illuminated images[J]. Signal Processing, 2016, 129: 82-96. doi: 10.1016/j.sigpro.2016.05.031
[7] LAND E H, MCCANN J J. Lightness and retinex theory[J]. JOSA, 1971, 61(1): 1-11. doi: 10.1364/JOSA.61.000001
[8] CHENG H D, SHI X J. A simple and effective histogram equalization approach to image enhancement[J]. Digital signal processing, 2004, 14(2): 158-170. doi: 10.1016/j.dsp.2003.07.002
[9] JUNG C, YANG Q, SUN T, et al. Low light image enhancement with dual-tree complex wavelet transform[J]. Journal of Visual Communication and Image Representation, 2017, 42: 28-36. doi: 10.1016/j.jvcir.2016.11.001
[10] WANG C, WU H, JIN Z. FourLLIE: boosting low-light image enhancement by fourier frequency information[EB/OL]. [2023-10-10]. https://arxiv.org/abs/2308.03033?context=cs.CV.
[11] YANG X, JIANG X, DU J. Low illumination image enhancement algorithm based on gamma transformation and fractional order[J]. Computational Engineering Design, 2021, 42: 762-769.
[12] WANG S, ZHENG J, HU H M, et al. Naturalness preserved enhancement algorithm for non-uniform illumination images[J]. IEEE Transactions on Image Processing, 2013, 22(9): 3538-3548. doi: 10.1109/TIP.2013.2261309
[13] LI C, GUO C, HAN L, et al. Low-light image and video enhancement using deep learning: a survey[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021, 44(12): 9396-9416.
[14] LORE K G, AKINTAYO A, SARKAR S. LLNet: a deep autoencoder approach to natural low-light image enhancement[J]. Pattern Recognition, 2017, 61: 650-662. doi: 10.1016/j.patcog.2016.06.008
[15] FU Z, YANG Y, TU X, et al. Learning a simple low-light image enhancer from paired low-light instances[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE, 2023: 22252-22261.
[16] LV F, LIU B, LU F. Fast enhancement for non-uniform illumination images using light-weight CNNs[C]//Proceedings of the 28th ACM International Conference on Multimedia. New York: IEEE, 2020: 1450-1458.
[17] LI C, GUO C, LOY C C. Learning to enhance low-light image via zero-reference deep curve estimation[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021, 44(8): 4225-4238.
[18] JAIN V, SEUNG S. Natural image denoising with convolutional networks[J]. Advances in Neural Information Processing Systems, 2008, 21: 1-8.
[19] LV F, LU F, WU J, et al. MBLLEN: low-light image/video enhancement using CNNs[J]. BMVC, 2018, 220(1): 1-4.
[20] CHEN C, CHEN Q, XU J, et al. Learning to see in the dark[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. New York: IEEE, 2018: 3291-3300.
[21] LI J, LI J, FANG F, et al. Luminance-aware pyramid network for low-light image enhancement[J]. IEEE Transactions on Multimedia, 2020, 23: 3153-3165.
[22] LIM S, KIM W. DSLR: deep stacked Laplacian restorer for low-light image enhancement[J]. IEEE Transactions on Multimedia, 2020, 23: 4272-4284.
[23] WANG R, ZHANG Q, FU C W, et al. Underexposed photo enhancement using deep illumination estimation[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE, 2019: 6849-6857.
[24] WANG L W, LIU Z S, SIU W C, et al. Lightening network for low-light image enhancement[J]. IEEE Transactions on Image Processing, 2020, 29: 7984-7996. doi: 10.1109/TIP.2020.3008396
[25] ZOTIN A. Fast algorithm of image enhancement based on multi-scale retinex[J]. Procedia Computer Science, 2018, 131: 6-14. doi: 10.1016/j.procs.2018.04.179
[26] WEI C, WANG W, YANG W, et al. Deep retinex decomposition for low-light enhancement[EB/OL]. [2023-10-10]. https://arxiv.org/abs/1808.04560v1.
[27] LI C, GUO J, PORIKLI F, et al. LightenNet: a convolutional neural network for weakly illuminated image enhancement[J]. Pattern Recognition Letters, 2018, 104: 15-22. doi: 10.1016/j.patrec.2018.01.010
[28] ZHANG Y, ZHANG J, GUO X. Kindling the darkness: a practical low-light image enhancer[C]//Proceedings of the 27th ACM international conference on multimedia. New York: IEEE, 2019: 1632-1640.
[29] ZHANG Y, GUO X, MA J, et al. Beyond brightening low-light images[J]. International Journal of Computer Vision, 2021, 129: 1013-1037. doi: 10.1007/s11263-020-01407-x
[30] JIANG Y, GONG X, LIU D, et al. EnlightenGAN: deep light enhancement without paired supervision[J]. IEEE Transactions on Image Processing, 2021, 30: 2340-2349. doi: 10.1109/TIP.2021.3051462
[31] RONNEBERGER O, FISCHER P, BROX T. U-net: convolutional networks for biomedical image segmentation[C]//Medical Image Computing and Computer-Assisted Intervention–MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. Berlin: Springer International Publishing, 2015: 234-241.
[32] XIONG W, LIU D, SHEN X, et al. Unsupervised low-light image enhancement with decoupled networks[C]//2022 26th International Conference on Pattern Recognition (ICPR). New York: IEEE, 2022: 457-463.
[33] HU J, GUO X, CHEN J, et al. A two-stage unsupervised approach for low light image enhancement[J]. IEEE Robotics and Automation Letters, 2021, 6(4): 8363-8370. doi: 10.1109/LRA.2020.3048667
[34] MA L, MA T, LIU R, et al. Toward fast, flexible, and robust low-light image enhancement[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE, 2022: 5637-5646.
[35] YANG W, WANG S, FANG Y, et al. From fidelity to perceptual quality: A semi-supervised approach for low-light image enhancement[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. New York: IEEE, 2020: 3063-3072.
[36] KOZLOWSKI W, SZACHNIEWICZ M, STYPULKOWSKI M, et al. Dimma: semi-supervised low light image enhancement with adaptive dimming[EB/OL]. [2023-10-10]. http://www.paperreading.club/page?id=188848.
[37] ZHANG L, ZHANG L, LIU X, et al. Zero-shot restoration of back-lit images using deep internal learning[C]//Proceedings of the 27th ACM international conference on multimedia. New York: IEEE, 2019: 1623-1631.
[38] ZHU A, ZHANG L, SHEN Y, et al. Zero-shot restoration of underexposed images via robust retinex decomposition[C]//2020 IEEE International Conference on Multimedia and Expo (ICME). New York: IEEE, 2020: 1-6.
[39] GUO C, LI C, GUO J, et al. Zero-reference deep curve estimation for low-light image enhancement[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. New York: IEEE, 2020: 1780-1789.
[40] CAI J, GU S, ZHANG L. Learning a deep single image contrast enhancer from multi-exposure images[J]. IEEE Transactions on Image Processing, 2018, 27(4): 2049-2062. doi: 10.1109/TIP.2018.2794218
[41] CHEN C, CHEN Q, DO M N, et al. Seeing motion in the dark[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision. New York: IEEE, 2019: 3185-3194.
[42] JIANG H, ZHENG Y. Learning to see moving objects in the dark[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision. New York: IEEE, 2019: 7324-7333.
[43] XU X, WANG R, FU C W, et al. SNR-aware low-light image enhancement[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE, 2022: 17714-17724.
[44] LU K, ZHANG L. TBEFN: a two-branch exposure-fusion network for low-light image enhancement[J]. IEEE Transactions on Multimedia, 2020, 23: 4093-4105.
[45] ZHAO Z, XIONG B, WANG L, et al. RetinexDIP: a unified deep framework for low-light image enhancement[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2021, 32(3): 1076-1088.
[46] LIU R, MA L, ZHANG J, et al. Retinex-inspired unrolling with cooperative prior architecture search for low-light image enhancement[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE, 2021: 10561-10570.
[47] GUO X, LI Y, LING H. LIME: Low-light image enhancement via illumination map estimation[J]. IEEE Transactions on Image Processing, 2016, 26(2): 982-993.
[48] LEE C, LEE C, KIM C S. Contrast enhancement based on layered difference representation of 2D histograms[J]. IEEE Transactions on Image Processing, 2013, 22(12): 5372-5384. doi: 10.1109/TIP.2013.2284059
[49] LOH Y P, CHAN C S. Getting to know low-light images with the exclusively dark dataset[J]. Computer Vision and Image Understanding, 2019, 178: 30-42. doi: 10.1016/j.cviu.2018.10.010
[50] LIU J, XU D, YANG W, et al. Benchmarking low-light image enhancement and beyond[J]. International Journal of Computer Vision, 2021, 129: 1153-1184. doi: 10.1007/s11263-020-01418-8
[51] BYCHKOVSKY V, PARIS S, CHAN E, et al. Learning photographic global tonal adjustment with a database of input/output image pairs[C]//CVPR 2011. New York: IEEE, 2011: 97-104.
[52] WANG T, ZHANG K, SHEN T, et al. Ultra-high-definition low-light image enhancement: a benchmark and transformer-based method[J]. Proceedings of the AAAI Conference on Artificial Intelligence. 2023, 37(3): 2654-2662.
[53] ZHAO H, GALLO O, FROSIO I, et al. Loss functions for image restoration with neural networks[J]. IEEE Transactions on Computational Imaging, 2016, 3(1): 47-57.
[54] LEDIG C, THEIS L, HUSZAR F, et al. Photo-realistic single image super-resolution using a generative adversarial network[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. New York: IEEE, 2017: 4681-4690.
[55] WANG X, YU K, WU S, et al. Esrgan: enhanced super-resolution generative adversarial networks[C]//Proceedings of the European Conference on Computer Vision (ECCV) Workshops. New York: IEEE, 2018: 63-79.
[56] BONTONOU M, LASSANCE K R E C, HACENE B G, et al. Introducing graph smoothness loss for training deep learning architectures[EB/OL]. [2023-10-10]. https://arxiv.org/abs/1905.00301v1.
[57] BING L, HAINA Q, WEIHUA X, et al. Ranking-based color constancy with limited training samples[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023, 45(10): 12304-12320. doi: 10.1109/TPAMI.2023.3278832
[58] 鄢杰斌, 方玉明, 刘学林. 图像质量评价研究综述——从失真的角度[J]. 中国图象图形学报, 2022, 27(5): 1430-1466. doi: 10.11834/jig.210790 YAN Jiebin, FANG Yuming, LIU Xuelin. The review of distortion related image quality assessment[J]. Journal of Image and Graphics, 2022, 27(5): 1430-1466. doi: 10.11834/jig.210790
[59] ANTKOWIAK J, BAINA T J, BARONCINI F V, et al. Final report from the video quality experts group on the validation of objective models of video quality assessment march 2000[EB/OL]. [2023-10-10]. https://xueshu.baidu.com/usercenter/paper/show?paperid=b46770167ca1a4c317e76cb50da54121&site=xueshu_se.
[60] MITTAL A, MOORTHY A K, BOVIK A C. No-reference image quality assessment in the spatial domain[J]. IEEE Transactions on Image Processing, 2012, 21(12): 4695-4708. doi: 10.1109/TIP.2012.2214050
[61] MITTAL A, SOUNDARARAJAN R, BOVIK A C. Making a “completely blind” image quality analyzer[J]. IEEE Signal Processing Letters, 2012, 20(3): 209-212.
[62] ZHENG N, ZHOU M, DONG Y, et al. Empowering low-light image enhancer through customized learnable priors[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision. New York: IEEE, 2023: 12559-12569.
-
期刊类型引用(24)
1. 陈俊杰，洪心皓. 面向嵌入式平台的智能巡视点位校正系统研究. 福建电脑. 2025(02): 88-92 . 
百度学术
2. 孙晨，邓宽. 基于嵌入式AI设备的光伏电池片缺陷智能检测系统. 电子设计工程. 2024(04): 129-134 . 百度学术
3. 杨海斌. 基于改进YOLOv5S算法的铝型材表面缺陷检测. 电脑知识与技术. 2024(06): 47-51 . 百度学术
4. 陈炀，周雁，王庆娟，张馨元，谌业恒. 融合图像处理技术的红树林鸟类鸣声识别算法. 现代计算机. 2024(21): 38-42+48 . 百度学术
5. 程强，张友兵，周奎. 基于改进YOLOX的动态视觉SLAM方法. 电子测量技术. 2024(23): 123-133 . 百度学术
6. 马泉，张欣怡，李洪波，石广洋，郝斌，张飞. 胜利煤矿智能环境监测系统研究与设计. 中国煤炭. 2023(01): 77-82 . 百度学术
7. 郭奕裕，周箩鱼. 安全帽佩戴检测网络模型的轻量化设计. 计算机工程. 2023(04): 312-320 . 百度学术
8. 胡天鑫，邓超，马俊杰，刘旺. 方程式赛车ROS平台下基于TensorRT的YOLOv5算法改进. 农业装备与车辆工程. 2023(05): 14-19 . 百度学术
9. 程绳，葛雄，肖非，朱传刚，吴军，肖海涛，李嗣，楚江平，袁雨薇. 基于多任务学习的输电线路小金具缺失推理加速算法. 计算机测量与控制. 2023(07): 251-257 . 百度学术
10. 顾成伟，丁勇，李登华. 基于计算机视觉的工业厂区人员安全警戒系统. 计算机与现代化. 2023(09): 20-26 . 百度学术
11. 张宇昂，李琦. 基于Jetson TX2的路面病害检测应用. 信息技术与信息化. 2023(09): 112-115 . 百度学术
12. 刘一呈，赵建敏，赵宇飞. 基于机器视觉和TX2的牛生长参数测量系统设计. 信息技术与信息化. 2023(10): 13-18 . 百度学术
13. 范亚龙，李琦，于令君. 基于深度学习的冶炼工人安全着装监测系统. 科学技术与工程. 2023(31): 13626-13631 . 百度学术
14. 张宇昂，李琦，薛芳芳，于令君. 基于Jetson TX2的路面裂缝检测系统设计. 公路. 2023(12): 337-344 . 百度学术
15. 郭智超，徐君明，刘爱东. 基于嵌入式平台与优化YOLOv3的航拍目标检测方法. 兵工自动化. 2022(03): 10-15+20 . 百度学术
16. 徐丁天，李海峰，徐良. 基于人眼跟踪的360°悬浮显示系统及其畸变校正方法. 光学学报. 2022(09): 232-240 . 百度学术
17. 贾云飞，郑红木，刘闪亮. 基于YOLOv5s的金属制品表面缺陷的轻量化算法研究. 郑州大学学报(工学版). 2022(05): 31-38 . 百度学术
18. 何曦，李良福，王娇颖，王洁，卢晓燕，钱钧，杨一洲，刘培桢. 基于NPU的实时深度学习跟踪算法实现. 应用光学. 2022(04): 682-692 . 本站查看
19. 王昀，刘泓，叶珺，刘亚璇，吴小莉. 一种基于YoloV4-tiny算法的智能电子秤设计. 长江信息通信. 2022(10): 37-41 . 百度学术
20. 刘之禹，李述，王英鹤. 基于ZYNQ的深度学习卷积神经网络加速平台设计. 计算机测量与控制. 2022(12): 264-269 . 百度学术
21. 刘智文，刘全，宋玮，赵天成. 基于共享主干网络的人物属性识别推理加速算法. 智能物联技术. 2022(06): 4-10 . 百度学术
22. 钟昊，陈博，李占文，杨永成. 面向铁路安全的智能视觉PaaS平台研究. 中国铁路. 2021(08): 128-132 . 百度学术
23. 李良熹，荣进国. 基于深度学习的智能烘培类商品识别系统研究. 信息与电脑(理论版). 2021(13): 156-158 . 百度学术
24. 柴荣轩，赵津发，吴航，张文昌，张广，陈炜. 基于IPPG的非接触式血氧饱和度检测技术研究. 医疗卫生装备. 2021(12): 23-28 . 百度学术
其他类型引用(79)
计量
- 文章访问数: 259
- HTML全文浏览量: 36
- PDF下载量: 115
- 被引次数: 103
陕公网安备 61011302001501号 