WO2025256045A1 - 基于多级收敛块和综合损失函数系统的极低光图像增强方法 - Google Patents

基于多级收敛块和综合损失函数系统的极低光图像增强方法

Info

Publication number
WO2025256045A1
WO2025256045A1 PCT/CN2024/131289 CN2024131289W WO2025256045A1 WO 2025256045 A1 WO2025256045 A1 WO 2025256045A1 CN 2024131289 W CN2024131289 W CN 2024131289W WO 2025256045 A1 WO2025256045 A1 WO 2025256045A1
Authority
WO
WIPO (PCT)
Prior art keywords
block
loss function
image
extremely low
function system
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
PCT/CN2024/131289
Other languages
English (en)
French (fr)
Inventor
王康康
刘宁
周红
陆昱翀
洪展鹏
许吉
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Nanjing University of Posts and Telecommunications
Original Assignee
Nanjing University of Posts and Telecommunications
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Nanjing University of Posts and Telecommunications filed Critical Nanjing University of Posts and Telecommunications
Publication of WO2025256045A1 publication Critical patent/WO2025256045A1/zh
Pending legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/60—Image enhancement or restoration using machine learning, e.g. neural networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/70—Denoising; Smoothing
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/77—Retouching; Inpainting; Scratch removal
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/90—Dynamic range modification of images or parts thereof
    • G06T5/92—Dynamic range modification of images or parts thereof based on global image properties
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10024—Color image
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20081—Training; Learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20084—Artificial neural networks [ANN]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20172—Image enhancement details
    • G06T2207/20192—Edge enhancement; Edge preservation

Definitions

  • This invention belongs to the field of image processing, specifically relating to an ultra-low light image enhancement method based on a multi-level convergence block and a comprehensive loss function system.
  • ultra-low light image enhancement technology aims to improve image sharpness, enhance details, reduce noise, and simultaneously improve brightness and contrast. This technology has wide applications in various fields, including night shooting, surveillance, military reconnaissance, and medical imaging.
  • LLNet deep learning-based methods for extremely low-light image enhancement
  • supervised learning-based methods such as MBLLEN, Retinex-Net, LPNet, and DSLR
  • reinforcement learning-based methods such as DeepExposure
  • unsupervised learning-based methods such as EnlightenGAN, ExCNet, and ZERO-DCE
  • semi-supervised learning-based methods such as DRBN.
  • a dedicated dataset for extremely low-light images and a multi-level convergence block for extremely low-light image enhancement are provided.
  • This invention relates to a multi-level convergence (MSC) block and a comprehensive loss function system for ultra-low light image enhancement.
  • the MSC block includes a selection block and a denoising block. By incorporating an attention mechanism, this block can effectively extract and recover fine-grained image details in ultra-low lighting environments, restoring true colors while reducing noise caused by insufficient sensitivity.
  • the ultra-low light image enhancement system described in this method employs a U-Net network structure, renowned for its efficiency in extracting and processing multi-scale features.
  • a feature extraction block is used at each layer to enhance details through an internal attention mechanism and suppress noise through an external attention mechanism.
  • a downsampling or upsampling operation is introduced after each feature extraction block to capture deeper feature information.
  • this invention provides a dedicated dataset for extremely low-light images and proposes MRASN (Multi-Residual Attention Shrinkage Network) as an advanced method for enhancing extremely low-light images.
  • MRASN Multi-Residual Attention Shrinkage Network
  • the structure of the multi-level convergence block is shown in Figure 2-3, including a selection block and a denoising block.
  • the first n (n 3) layers and output layer of the U-shaped network...
  • the feature maps are fed into the selection block.
  • channel weights are computed in space using global pooling and a multilayer perceptron (MLP).
  • MLP multilayer perceptron
  • the output is obtained through multiplication and convolution of the three data points.
  • Pools and Poolc represent global pooling in the space and channel, respectively.
  • the input is fed into the denoising block, the absolute value of all features in the feature map is calculated, and after global average pooling, a one-dimensional vector is obtained. This vector is then passed to a two-layer fully connected (FC) network to calculate the threshold. Finally, subtract the threshold and activate to obtain the output.
  • FC fully connected
  • z and c are the features at neuron cth, and i, j, and c are the width, height, and channel of the feature map, respectively.
  • the loss function primarily employs the mean absolute error loss ( L1 ) to enable rapid network convergence and achieve overall image enhancement. Secondly, to enrich color information and reduce the error between the image and the true color, this invention designs a color restoration loss function to calculate the average color error of each of the three channels.
  • the color restoration loss function is as follows:
  • Gj and Yj represent the average color values of the p channel in the real image and the output image, respectively.
  • this invention incorporates an average structure loss (L ⁇ sub>ssim ⁇ /sub> ), which segments the image into smaller parts, calculates the structure loss within each local region, and then averages the results across the entire image.
  • the total loss function is shown below:
  • ⁇ and ⁇ are the weights of the loss function.
  • a novel method combining global attention mechanism and CNN model is proposed to enhance extremely low-light images. This method introduces a multi-level convergence block and a comprehensive loss function system, which can improve the brightness, detail recovery and color recovery capabilities of the image. It can effectively restore fine-grained image details and true colors, while reducing noise, and the effect surpasses that of previous models.
  • the model adopted in this invention exhibits a significantly faster processing speed and can smoothly process extremely low-light videos.
  • Figure 1 is a flowchart of the ultra-low light image enhancement method of the present invention.
  • Figure 2 is a structural diagram of the selected block in the multi-level convergence block of the present invention.
  • Figure 3 is a structural diagram of the noise reduction block in the multi-level convergence block of the present invention.
  • FIG. 1 is a flowchart of the method described in this invention, employing a U-Net network structure.
  • U-Net structure given an RGB image, feature extraction blocks are used at each layer. Internal attention mechanisms enhance details, while external attention mechanisms suppress noise. Furthermore, downsampling or upsampling operations are introduced after each feature extraction block to capture deeper feature information.
  • this invention develops a novel loss function system to enhance the ability of blocks to extract complex details and accurately recover colors, thereby avoiding color shift phenomena.
  • Figure 1 shows the proposed network model, and the loss function is used during model training.
  • Training and testing were performed on a Dell small form factor workstation with the following specifications: one Intel(R) Xeon(R) Gold 5218R CPU and four NVIDIA GeForce 3090 GPUs.
  • the framework was implemented using PyTorch and Python 3.7.
  • images in the training set were randomly cropped into 512 ⁇ 512 patches.
  • Data was further augmented using random flipping, rotation, and scaling techniques.
  • the batch size was set to 8, and the network was initialized using default parameters.
  • the network was trained for a total of 70 epochs using the Adam optimizer. For the first 20 epochs, the initial learning rate was set to 1e-3, then reduced to 1e-4 for the remaining epochs.
  • the loss function hyperparameters ⁇ and ⁇ were set to 0.5 and 0.2, respectively, to achieve the desired balance between the different components of the loss function.
  • the method described in this invention is extensively compared with state-of-the-art algorithms for low-light and ultra-low-light image enhancement proposed in recent years.
  • the method described in this invention is evaluated against CNN-based algorithms MBLLEN, DSLR, Zero-DCE, and MRASN, and Transformer-based algorithms IAT and LLFormer.
  • SID and LOL were widely used datasets: SID and LOL, as well as the proposed MRASN. These datasets can test the processing effect, generalization ability, noise suppression ability, and detail representation ability of each method. Calculations show that the LOL dataset has moderate noise, the SID dataset has a dark environment, while MRASN exhibits low brightness and high noise levels.
  • PSNR Peak Signal-to-Noise Ratio
  • SSIM Structural Similarity
  • the method described in this invention excels in both SSIM and PSNR, achieving the highest values in both metrics.
  • the higher PSNR value ensures reduced distortion and improved fidelity in the reconstructed image.
  • the higher SSIM value signifies significant improvements in brightness, contrast, color, and detail features.
  • the LOL dataset contains images of size 600 ⁇ 400, while the SID dataset images were converted to PNG format with a size of 1920 ⁇ 1080.
  • the dataset used in this invention also consists of images of size 1920 ⁇ 1080.
  • Each method was thoroughly tested on each image in each dataset, and the average runtime results were recorded, as shown in Table 2.
  • the method described in this invention has superior computational efficiency and significant stability in terms of efficiency when applied to various datasets.
  • This invention focuses on developing a novel method to enhance extremely low-light images by combining a global attention mechanism with a CNN model.
  • This method introduces a multi-level convergence block and a comprehensive loss function system, which improves the image's brightness, detail recovery, and color restoration capabilities. It effectively restores fine-grained image details and true colors while reducing noise.
  • the model employed in this invention exhibits significantly faster processing speed and can smoothly handle videos in extremely low light conditions.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Molecular Biology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Image Processing (AREA)

Abstract

基于多级收敛块和综合损失函数系统的极低光图像增强方法，包括一个多级收敛(MSC)块和一个综合损失函数系统。MSC块包括一个选择块和一个去噪块。通过结合注意力机制，该MSC块可以在极低照明环境中有效地提取和恢复细粒度的图像细节，在还原真实色彩的同时减少因感光度不足而引起的噪声。极低光图像增强能够在高度黑暗的环境中提供亮度和清晰度与日光相似的视觉效果，在民用和军事应用中都发挥着关键作用。

Description

基于多级收敛块和综合损失函数系统的极低光图像增强方法 技术领域
本发明属于图像处理领域，具体涉及基于多级收敛块和综合损失函数系统的极低光图像增强方法。
背景技术
在极低的光照条件下，由于光通量受限，图像通常会出现低亮度、低对比度和大量噪声。这导致图像细节不清晰，并使人眼难以辨别有意义的信息。为了解决这个问题，极低光图像增强技术旨在提高图像清晰度，增强细节，减少噪声，同时提高亮度和对比度。这项技术在夜间射击、监视、军事侦察和医学成像等各个领域都有广泛的应用。
自2017年以来，以LLNet为代表的基于深度学习的极低光图像增强方法越来越受欢迎。这些解决方案在RAW格式或RGB格式的数据集上使用不同的学习策略，试图解决极低光图像增强的问题。如基于监督学习的MBLLEN、Retinex-Net、LPNet、DSLR、基于强化学习的DeepExposure、基于无监督学习的EnlightenGAN、ExCNet、ZERO-DCE和基于半监督学习的DRBN等。
然而，这些模型是使用LOL数据集训练的，并且它们的图像照明仍然与真实的暗光环境不同。此外，Transformer模型缺乏对空间信息进行有效建模的能力，并且具有大量的参数，这使得它们不太适合处理高分辨率的图像。
发明内容
针对上述数据集不理想、网络模型不合适等问题，本发明的目的是
提供一个针对极低光图像的专用数据集以及一种用于极低光图像增强的多级收敛块。
技术方案：本发明是一种用于极低光图像增强的多级收敛(MSC)块和一个综合损失函数系统。所述MSC块包括一个选择块和一个去噪块。通过结合注意力机制，该块可以在极低照明环境中有效地提取和恢复细粒度的图像细节，在还原真实色彩的同时减少因感光度不足而引起的噪声。
所述方法的极低光图像增强系统如图1所示，采用了U-Net网络结构，该结构以其在提取和处理多尺度特征方面的效率而闻名。在所述U-Net结构中，给定一幅RGB图像，在每一层使用特征提取块，通过内部注意力机制增强细节，并通过外部注意力机制抑制噪声。此外，在每个特征提取块之后引入下采样或上采样操作，以捕获更深层次的特征信息。为了进一步解决细节丢失的问题，引入了一个多级收敛块。该块从前n个(所述方法中设置n＝3)输出层和网络的最终输出层获取特征图，并执行收敛操作。结果是一个三通道的RGB图像作为最终输出。
为了更好地模拟极低光环境，并利用Transformer中自注意力机制的优势，本发明提供了一个针对极低光图像的专用数据集，并提出了MRASN(Multi-Residual Attention Shrinkage Network)作为增强极低光图像的高级方法。
所述多级收敛块的结构如图2-3所示，其中包括一个选择块和一个去噪块。所述U形网络的前n(n＝3)层和输出层的特征图被送到所述选择块中。首先使用全局池化和多层感知机(MLP)在空间中计算通道权重以捕获包含丰富而复杂信息的通道，然后通过在所述通道中使用全局池化来获得空间权重以寻找所述特征图中的边缘信息。通过三者相乘和卷积运算获得输出
M＝Conv(I+I·Fc·Fs)
＝Conv(I+I·ReLU(MLP(Pools(I)))
·ReLU(Conv(Poolc(I)))
其中Pools和Poolc分别表示空间和通道中的全局池化。
由于补充信息也增加了噪声，需要将所述输入到所述去噪块中，计算所述特征图中所有特征的绝对值，经过全局平均池化后获得一维向量，然后将其传递到两层FC网络中计算阈值 最后减去所述阈值并激活以获得输出公式如下：
O＝ReLU(M-τ)
其中，zc是cth神经元处的特征，i、j和c分别是特征图的宽度、高度和通道。
所述损失函数主体选择平均绝对误差损失(L1)，使网络快速收敛，实现图像的整体增强。其次，为了丰富颜色信息，减少图像与真实颜色之间的误差，本发明设计了一个颜色恢复损失函数来计算三个通道中每个通道的颜色平均误差。颜色恢复损失函数如下：
其中Gj和Yj分别表示真实图像和输出图像中p通道的平均颜色值。
为了消除色彩偏移和解决恢复图像的边缘模糊问题，本发明添加了一个平均结构损失(Lssim)，它将图像分割成更小的部分，计算局部范围内的结构损失，然后取整体的平均值。总损失函数如下所示：
L＝αL1+βLc+(1-α-β)Lssim
其中α和β是损失函数的权重。
有益效果：(1)提出一种将全局注意力机制和CNN模型相结合的新方法来增强极低光图像。该方法引入了多级收敛块和综合损失函数系统，可以提高图像的亮度、细节恢复和色彩恢复能力。可以有效地还原细粒度的图像细节和真实的色彩，同时减少噪声，效果超越了以前的模型。(2)此外，与其他模型相比，本发明所采用的模型展现出明显更快的处理速度，能够顺利处理极低光线的视频。
附图说明
图1是本发明所述极低光图像增强方法的流程图。
图2是本发明多级收敛块中选择块的结构图。
图3是本发明多级收敛块中去噪块的结构图。
图中附图标记如下：1-ReLU激活函数；2-矩阵相乘；3-元素相加。
具体实施方式
下面结合说明书附图对本发明的技术方案做进一步的详细说明。
图1是本发明所述方法的流程图，采用了U-Net网络结构。在U-Net结构中，给定一幅RGB图像，在每一层使用特征提取块，通过内部注意力机制增强细节，并通过外部注意力机制抑制噪声。此外，在每个特征提取块之后引入下采样或上采样操作，以捕获更深层次的特征信息。为了进一步解决细节丢失的问题，引入了一个多级收敛块。该块从前n个(本实施例中设置n＝3)输出层和网络的最终输出层获取特征图，并执行收敛操作。结果是一个三通道的RGB图像作为最终输出。此外，本发明还开发了一种新的损失函数系统，以增强块提取复杂细节和准确恢复颜色的能力，从而避免色彩偏移现象，图1是提出的网络模型，损失函数使用在训练模型的时候。
训练和测试程序是在戴尔小型计算工作站上进行的，该工作站具有以下规格：一个lntel(R)Xeon(R)Gold 5218R CPU和四个NVIDIA GeForce 3090GPU。使用PyTorch和Python 3.7实现该框架。在训练过程中，训练集中的图像被随机裁剪成大小为512×512的小块。通过随机翻转、旋转和缩放技术进一步增强数据。批大小设置为8，使用默认参数进行网络初始化。在训练过程中，使用Adam优化器对网络进行总共70个epoch的训练。对于前20个epoch，初始学习率设置为1e-3，然后对剩余epoch降低为1e-4。将损失函数超参数α和β分别设置为0.5和0.2，以实现损失函数不同分量之间所需的平衡。
基于上述内容，将本发明所述方法与近年来提出的最先进的用于低光图像增强和极低光图像增强的算法进行广泛的比较。具体来说，就是针对基于CNN的MBLLEN、DSLR、Zero-DCE、MRASN和基于Transformer的IAT和LLFormer评估本发明所述的方法。为了评估不同方法在不同数据集上的性能，选择两个广泛使用的数据集：SID、LOL，以及所提出的MRASN。这些数据集能够检验每种方法的处理效果、泛化能力、噪声抑制能力和细节表征能力。经过计算发现LOL数据集具有中等的噪声，SID数据集具有较暗的环境，而MRASN表现出较低的亮度和较高的噪声水平。
常用于图像评估的关键参数是峰值信噪比(PSNR)和结构相似度(SSIM)。PSNR是处理后的图像质量的可靠指标，PSNR值越高，图像噪声越低。由于图像细节和噪声都是高频分量，因此SSIM提供了细节保留程度的补充度量，包括亮度、对比度、颜色和细节特征等方面。在对每个数据集进行处理后，计算上述方法以及本发明所述方法的PSNR和SSIM。
根据表1中的结果可知，本发明所述方法在SSIM和PSNR方面都很出色，在这两个指标上都达到了最高值。较高的PSNR值保证了重建图像失真的减少和保真度的提高，同时，与其他方法相比，较高的SSIM值意味着亮度、对比度、颜色和细节特征方面都有显著改善。
表1
除此之外，还在各种数据集上进行了运行时的测试，将本发明所述方法与其他方法的处理时间进行比较，同时确保平台配置的一致性。 LOL数据集包含大小为600×400的图像，而SID数据集图像被转换为尺寸为1920×1080的PNG格式。同样，本发明所采用的数据集也由尺寸为1920×1080的图像组成。在每个数据集中的每个图像上对每种方法进行彻底的测试，并记录平均运行时间结果，如表2所示。综合来看，与其他方法相比，本发明所述的方法具有优越的计算效率，并且在应用于各种数据集时，在效率方面具有显著的稳定性。
表2
为了验证本发明所述方法中每个组件的有效性，还进行了单独的测试来评估仅使用选择块和仅使用去噪块时的效果，结果与合并整个MSC块相比表现较差。此外，为了确定损失函数的系数，还进行了各种测试，包括单独使用L1损失、L1和Lc的组合应用以及L1和Lssim的组合应用。这些测试证明了每个损失函数的有效性。最后，还测试了层数变化对于补偿缺失细节的影响。这些具体数据如表3所示。
表3
在本发明中，致力于研究一种将全局注意力机制和CNN模型相结合的新方法来增强极低光图像。该方法引入了多级收敛块和综合损失函数系统，可以提高图像的亮度、细节恢复和色彩恢复能力。可以有效地还原细粒度的图像细节和真实的色彩，同时减少噪声。与其他模型相比，本发明所采用的模型展现出明显更快的处理速度，能够顺利处理极低光线的视频。
以上所述仅为本发明的较佳实施方式，本发明的保护范围并不以上述实施方式为限，但凡本领域普通技术人员根据本发明所揭示内容所作的等效修饰或变化，皆应纳入权利要求书中记载的保护范围内。

Claims (5)

  1. 基于多级收敛块和综合损失函数系统的极低光图像增强方法，其特征在于：
    所述方法包括采用了多级收敛MSC块的U-Net网络结构和综合损失函数系统；
    所述MSC块包括选择块和去噪块，通过结合注意力机制，MSC块可以在极低照明环境中提取和恢复细粒度的图像细节，在还原真实色彩的同时减少因感光度不足而引起的噪声；
    给定一幅RGB图像，在U-Net网络结构的每一层使用特征提取块，通过内部注意力机制增强细节，并通过外部注意力机制抑制噪声；此外，在每个特征提取块之后引入下采样或上采样操作，以捕获深层次的特征信息；
    多级收敛块从U-Net网络结构的前n个输出层和最终输出层获取特征图，n＝3，并执行收敛操作，获得三通道的RGB图像作为最终输出。
  2. 根据权利要求1所述的基于多级收敛块和综合损失函数系统的极低光图像增强方法，其特征在于：U-Net网络结构的前n层和输出层的特征图被送到所述选择块中，首先使用全局池化和多层感知机MLP在空间中计算通道权重以捕获包含信息的通道，然后通过在所述通道中使用全局池化来获得空间权重以寻找所述特征图中的边缘信息；通过三者相乘和卷积运算获得输出其中H、W、C分别表示高度、宽度和通道数；
    M＝Conv(I+I·Fc·Fs)
    ＝Conv(I+I·ReLU(MLP(Pools(I)))·ReLU(Conv(Poolc(I)))
    其中Pools和Poolc分别表示空间和通道中的全局池化。
  3. 根据权利要求2所述的基于多级收敛块和综合损失函数系统的极低光图像增强方法，其特征在于：将所述输入到所述去噪块中，计算所述特征图中所有特征的绝对值，经过全局平均池化后获得一维向量，然后将其传递到两层FC网络中计算阈值 最后减去所述阈值并激活以获得输出公式如下：

    O＝ReLU(M-τ)
    其中，zc是cth神经元处的特征，i、j和c分别是特征图的宽度、高度和通道。
  4. 根据权利要求1所述的基于多级收敛块和综合损失函数系统的极低光图像增强方法，其特征在于：所述损失函数主体选择平均绝对误差损失L1，使网络快速收敛，实现图像的整体增强；设计一个颜色恢复损失函数来计算三个通道中每个通道的颜色平均误差，颜色恢复损失函数如下：
    其中，Gj和Yj分别表示真实图像和输出图像中p通道的平均颜色值。
  5. 根据权利要求1所述的基于多级收敛块和综合损失函数系统 的极低光图像增强方法，其特征在于：所述综合损失函数系统含有一个平均结构损失Lssim，用于消除色彩偏移和解决恢复图像的边缘模糊问题；它将图像分割成更小的部分，计算局部范围内的结构损失，然后取整体的平均值；总损失函数如下所示：
    L＝αL1+βLc+(1-α-β)Lssim
    其中α和β是损失函数的权重。
PCT/CN2024/131289 2024-06-13 2024-11-11 基于多级收敛块和综合损失函数系统的极低光图像增强方法 Pending WO2025256045A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202410761487.9A CN118628377A (zh) 2024-06-13 2024-06-13 基于多级收敛块和综合损失函数系统的极低光图像增强方法
CN202410761487.9 2024-06-13

Publications (1)

Publication Number Publication Date
WO2025256045A1 true WO2025256045A1 (zh) 2025-12-18

Family

ID=92609197

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2024/131289 Pending WO2025256045A1 (zh) 2024-06-13 2024-11-11 基于多级收敛块和综合损失函数系统的极低光图像增强方法

Country Status (2)

Country Link
CN (1) CN118628377A (zh)
WO (1) WO2025256045A1 (zh)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN118628377A (zh) * 2024-06-13 2024-09-10 南京邮电大学 基于多级收敛块和综合损失函数系统的极低光图像增强方法

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20210272246A1 (en) * 2018-11-26 2021-09-02 Guangdong Oppo Mobile Telecommunications Corp., Ltd. Method, system, and computer-readable medium for improving quality of low-light images
CN115393225A (zh) * 2022-09-07 2022-11-25 南京邮电大学 一种基于多层次特征提取融合的低光照图像增强方法
CN118628377A (zh) * 2024-06-13 2024-09-10 南京邮电大学 基于多级收敛块和综合损失函数系统的极低光图像增强方法

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20210272246A1 (en) * 2018-11-26 2021-09-02 Guangdong Oppo Mobile Telecommunications Corp., Ltd. Method, system, and computer-readable medium for improving quality of low-light images
CN115393225A (zh) * 2022-09-07 2022-11-25 南京邮电大学 一种基于多层次特征提取融合的低光照图像增强方法
CN118628377A (zh) * 2024-06-13 2024-09-10 南京邮电大学 基于多级收敛块和综合损失函数系统的极低光图像增强方法

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
蔡闻超 (CAI, WENCHAO): "基于多残差注意力深度收缩神经网络的超微光图像增强技术研究 (non-official translation: Research on Ultra-low Light Image Enhancement Technology Based on Multi-residual Attention Deep Shrinkage Neural Network)", 中国优秀硕士学位论文全文数据库 (CHINA MASTER'S THESES FULL-TEXT DATABASE), 15 May 2024 (2024-05-15), XP009566107 *

Also Published As

Publication number Publication date
CN118628377A (zh) 2024-09-10

Similar Documents

Publication Publication Date Title
Chen et al. Progressive semantic-aware style transformation for blind face restoration
Zhang et al. Multi-branch and progressive network for low-light image enhancement
Huang et al. Deep hyperspectral image fusion network with iterative spatio-spectral regularization
CN116152120B (zh) 一种融合高低频特征信息的低光图像增强方法及装置
Liu et al. PD-GAN: perceptual-details gan for extremely noisy low light image enhancement
Guo et al. Underwater image restoration via polymorphic large kernel cnns
CN108921786A (zh) 基于残差卷积神经网络的图像超分辨率重构方法
CN107610049B (zh) 基于稀疏正则化技术和加权引导滤波的图像超分辨率方法
CN115170915A (zh) 一种基于端到端注意力网络的红外与可见光图像融合方法
CN115035011B (zh) 一种融合策略下自适应RetinexNet的低照度图像增强方法
CN110796622B (zh) 一种基于串联神经网络多层特征的图像比特增强方法
CN113284061B (zh) 一种基于梯度网络的水下图像增强方法
CN115760640A (zh) 基于含噪Retinex模型的煤矿低光照图像增强方法
Xu et al. Bilateral interaction for local-global collaborative perception in low-light image enhancement
Chen et al. Attention-based broad self-guided network for low-light image enhancement
Ye et al. CodedBGT: Code bank-guided transformer for low-light image enhancement
Zhang et al. A cross-scale framework for low-light image enhancement using spatial–spectral information
Jin et al. Deep low light image enhancement via multi-scale recursive feature enhancement and curve adjustment
Tian et al. Deformable convolutional network constrained by contrastive learning for underwater image enhancement
Zhou et al. A Novel Lightweight Underwater Image Enhancement Framework for Resource-Limited IoT Devices
CN118628377A (zh) 基于多级收敛块和综合损失函数系统的极低光图像增强方法
Wang et al. Channel self-attention based low-light image enhancement network
CN119205539B (zh) 一种基于低光特征先验的低照度图像增强方法
CN114926359A (zh) 联合双颜色空间恢复和多级解码结构的水下图像增强方法
CN119579432B (zh) 一种基于耦合退化处理的水下图像质量增强方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 24943199

Country of ref document: EP

Kind code of ref document: A1