CN109785263B - Retinex-based inverse tone mapping image conversion method - Google Patents
Retinex-based inverse tone mapping image conversion method Download PDFInfo
- Publication number
- CN109785263B CN109785263B CN201910030455.0A CN201910030455A CN109785263B CN 109785263 B CN109785263 B CN 109785263B CN 201910030455 A CN201910030455 A CN 201910030455A CN 109785263 B CN109785263 B CN 109785263B
- Authority
- CN
- China
- Prior art keywords
- image
- component
- illumination
- dynamic range
- reflection component
- 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.)
- Active
Links
Images
Landscapes
- Image Processing (AREA)
Abstract
The invention discloses a reverse tone mapping image conversion method based on Retinex, which is based on a multi-scale Retinex model and divides reverse tone mapping into a dynamic range expansion subtask and a detailed texture recovery subtask; different inverse tone mapping networks are designed according to the subtasks of inverse tone mapping, the recovered illumination component and reflection component are obtained, and then merging is carried out based on a multi-scale Retinex model, so that a high dynamic range image is obtained. Compared with the prior art, the method has the advantages of reducing color cast, recovering information of an overexposed area and the like; the conversion of the low dynamic range image into the high dynamic range image can be done robustly. And the dynamic range is expanded, and meanwhile, the texture of an overexposed area can be recovered, and the color is ensured not to be distorted.
Description
Technical Field
The invention relates to the technical field of digital image processing, in particular to an inverse tone mapping image conversion method based on Retinex theory.
Background
Currently, 4K television technology and related applications are rapidly developing. In the 4K standard, high dynamic range is an important component. Since most media resources are still stored in a low dynamic range, the conversion of media resources from a low dynamic range to a high dynamic range, i.e. the inverse tone mapping technique, is a key technique in 4K tv technology.
Since inverse tone mapping (inverse tone mapping for short) is a ill-conditioned problem, it is necessary to recover the information lost in the quantization and compression processes of low dynamic range images. The conventional method usually proposes a parametric model, and the conversion from the low dynamic range image to the high dynamic range image is completed through the model. Such a method has two drawbacks, one is that the information lost in the low dynamic range image cannot be completely recovered; and the complicated parameter setting is difficult for common users. The problem with the currently proposed deep learning method is mainly that the recovery of different missing information, such as over-exposed areas, under-exposed areas and color information, cannot be well taken into account.
Disclosure of Invention
The invention aims to provide an inverse tone mapping network based on a Retinex theory, which can robustly complete the conversion from a low dynamic range image to a high dynamic range image by carrying out image conversion based on the inverse tone mapping of the Retinex theory.
The principle of the invention is as follows: based on Retinex theory (multi-scale Retinex model), inverse tone mapping is divided into two subtasks, including expansion of dynamic range and restoration of detail texture. Rtex is a theory and method widely used in digital image processing, and it is believed that a digital image can be decomposed into an illumination component and a reflection component, which are independent of each other. The invention decomposes the low dynamic range image to be processed into an illumination component and a reflection component based on Retinex theory. The illumination component mainly represents global illumination information in the image and is responsible for restoring the dynamic range; the reflection component mainly represents color and detail information in the image and is responsible for recovering over-exposure and under-exposure missing information. Based on this, a reasonable inverse tone mapping network is designed for the two subtasks of inverse tone mapping. For the illumination component, in order to maintain structural integrity and avoid information loss, a residual error network is adopted to complete information recovery by learning a residual error. And the reflection component contains a large amount of detail information such as color texture and the like, the multi-scale features are required to be used for recovery, and in order to extract sufficient features, a U-Net convolution neural network structure is adopted. And after the recovered illumination component and the recovered reflection component are obtained through the two networks, the illumination component and the reflection component are combined based on a Retinex theory to obtain a final high dynamic range image.
The technical scheme of the invention is as follows:
a method for converting reverse tone mapping images based on Retinex is characterized in that based on a multi-scale Retinex model, reverse tone mapping is divided into two subtasks, including expansion of a dynamic range and recovery of detailed textures; designing different inverse tone mapping networks aiming at two subtasks of inverse tone mapping to obtain a recovered illumination component and a recovered reflection component, and then combining based on a multi-scale Retinex model to obtain a high dynamic range image; the method specifically comprises the following steps:
1) decomposing the low dynamic range image into an illumination component and a reflection component based on retinex theory, wherein the illumination component represents the global illumination condition and the dynamic range of the image, and the reflection component represents the color and detail information of the image;
specifically, based on the Retinex theory, the original image can be decomposed by a multi-scale gaussian filtering method, the original image is first subjected to a filtering operation by gaussian filtering, the result of the filtering is used as an illumination component, and the original image and the illumination component are subtracted to form a reflection component, so that the image is decomposed into two parts, namely the illumination component and the reflection component.
2) Respectively designing different inverse tone mapping networks for the illumination component and the reflection component of the image; designing an illumination recovery network for the illumination component of the image, adopting a full convolution network with unchanged feature map size, and introducing a residual error structure;
because the illumination component represents global information, in order to reduce the loss of the information, down-sampling needs to be avoided, so a full convolution network with unchanged feature diagram size is adopted, meanwhile, in order to stabilize and efficiently train, the input and the output are added according to a residual error structure, and the illumination component is recovered by learning a residual error; in specific implementation, the illumination recovery network has 7 layers, each layer comprises a convolution layer and an activation layer, the activation function adopts SELU, the convolution kernel size is 3 × 3, the step length is 1, the number of characteristic channels in the first six layers is 64, the number of characteristic channels in the last layer is 3, and in order to keep the size of the characteristic diagram unchanged, edge filling is performed in a mirror symmetry mode.
Different from the illumination component, the reflection component has a large amount of color texture information, the information of the part is important for recovering an overexposed and underexposed area, and the reflection component needs to be recovered by utilizing multi-scale information, so that the U-Net is adopted as a reflection component recovery network in the invention.
3) The two components are respectively restored through two different inverse tone mapping networks and then combined. Thus, the complexity of the inverse tone mapping task is reduced, and the recovery effect is more robust.
Specifically, the recovered illumination component and the reflected component are added by using a multi-scale Retinex model to obtain a finally recovered high dynamic range image.
Compared with the prior art, the invention has the beneficial effects that:
the invention provides a reverse tone mapping image conversion method based on Retinex, which decomposes an image into an illumination component and a reflection component, respectively designs sub-networks, an illumination component recovery sub-network and a reflection component recovery sub-network are used for conversion,
the existing technology does not carry out decomposition, and the problems of color cast, difficult recovery of an overexposed area and the like caused by the non-decomposition are solved. Compared with the prior art, the method has the advantages of reducing color cast, recovering information of an overexposed area and the like; the conversion of the low dynamic range image into the high dynamic range image can be done robustly. And the dynamic range is expanded, and meanwhile, the texture of an overexposed area can be recovered, and the color is ensured not to be distorted.
Drawings
Fig. 1 is an overall flowchart in the embodiment of the present invention.
Fig. 2 is a diagram illustrating an example of a specific structure of an illumination pattern recovery network according to an embodiment of the present invention.
Fig. 3 is a diagram illustrating an example of a specific structure of a reflection graph recovery network according to an embodiment of the present invention.
Detailed Description
The invention is described in detail below with reference to the figures and examples, without in any way limiting the scope of the invention.
The invention provides a reverse tone mapping image conversion method based on Retinex, which can robustly complete the conversion from a low dynamic range image to a high dynamic range image.
The method of the invention decomposes the low dynamic range image to be processed into illumination component and reflection component based on Retinex theory. The illumination component mainly represents global illumination information in the image and is responsible for restoring the dynamic range; the reflection component mainly represents color and detail information in the image and is responsible for recovering over-exposure and under-exposure missing information. For the illumination component, in order to maintain structural integrity and avoid information loss, a residual error network is adopted to complete information recovery by learning a residual error. And the reflection component contains a large amount of detailed information such as color texture and the like, needs to be recovered by utilizing multi-scale features, and adopts a U-Net network structure in order to extract sufficient features. And after the recovered illumination component and the recovered reflection component are obtained through the two networks, the illumination component and the reflection component are combined based on a Retinex theory to obtain a final high dynamic range image.
FIG. 1 shows the flow of the process of the present invention. Based on Retinex theory, the digital image can be decomposed into an illumination component and a reflection component which are independent and do not influence each other; dividing the inverse tone mapping of the image into two subtasks, including the expansion of a dynamic range and the recovery of detail texture; a reasonable inverse tone mapping network is designed for the two subtasks of inverse tone mapping.
In the following embodiments, as shown in fig. 1, first we decompose the low dynamic range image into two parts, illumination component and reflection component, based on Retinex theory. Specifically, based on Retinex theory, the original image may be decomposed by a multi-scale gaussian filtering method, the original image is first subjected to a filtering operation by gaussian filtering, the result of the filtering is an illumination component, and the original image and the illumination component are subtracted to form a reflection component, thereby decomposing the image into two parts, namely an illumination component and a reflection component. The illumination component primarily represents the global illumination condition of the image and the reflection component represents the color and texture details of the image. The illumination component needs to ensure the structural integrity and reduce the information loss, so the invention does not adopt a downsampling structure, but uses a full convolution network with a constant size to recover. In particular, the size of the characteristic diagram is not changed in each convolution operation, so that the resolution of each layer of convolution is ensured to be the same.
On the other hand, considering that the residual error network can improve the learning efficiency and reduce the learning difficulty, a residual error structure is introduced, specifically, the input of the residual error structure is an unprocessed illumination component, the output is an illumination component processed by the residual error structure network, the input and the output are added, and the restoration of the illumination component is completed by learning the residual error. The illumination recovery network has 7 layers in total, each layer comprises a convolution layer and an activation layer, the activation function adopts SELU, the convolution kernel size is 3 x 3, the step length is 1, the number of characteristic channels of the first six layers is 64, the number of characteristic channels of the last layer is 3, and in order to keep the size of the characteristic graph unchanged, edge filling is carried out in a mirror symmetry mode. Referring to fig. 2, a specific structure is shown, pixel filling is performed on an edge of an input feature map, the value of the pixel filling is the pixel value of an original edge, and the purpose of performing edge filling in a mirror symmetry mode is to avoid edge pixel loss in a convolution process so as to ensure that the size of the feature map is unchanged.
The reflection component contains a large amount of detail information such as color texture and the like, the information has an important effect on recovering information of an overexposed and underexposed area, and in order to recover the reflection component by utilizing multi-scale information, a U-Net structure is adopted. The network has 10 layers, 5 convolution layers and 5 deconvolution layers, the convolution kernel size of the first four convolution layers is 3 x 3, the step length is 2, and the number of the characteristic channels is respectively 64,128,256 and 512; the convolution kernel size of the 5 th convolution layer is 3 x 3, the step size is 1, and the number of characteristic channels is 1024. In order to avoid the chessboard artifact, in the deconvolution layer, line sampling is carried out, the resolution of the characteristic diagram is enlarged, and then convolution operation is carried out. Specifically, the deconvolution kernel size is 3 × 3, the step size is 1, and the number of characteristic channels is 512,256,128,64, and 3, respectively. To speed up convergence, batch normalization operations are added for each layer. The specific structure is shown in fig. 3.
And after the recovered illumination component and the recovered reflection component are obtained, combining the illumination component and the reflection component by utilizing a Retinex theory, specifically, adding the illumination component and the reflection component (after recovery) by utilizing a multi-scale Retinex model to obtain a finally recovered high dynamic range image.
Thereby completing the conversion of the low dynamic range image into the high dynamic range image.
The method reduces the complexity of the inverse tone mapping task, and has high efficiency and good robustness.
It is noted that the disclosed embodiments are intended to aid in further understanding of the invention, but those skilled in the art will appreciate that: various substitutions and modifications are possible without departing from the spirit and scope of the invention and appended claims. Therefore, the invention should not be limited to the embodiments disclosed, but the scope of the invention is defined by the appended claims.
Claims (7)
1. A reverse tone mapping image conversion method based on Retinex is based on a multi-scale Retinex model, and divides reverse tone mapping into a dynamic range expansion subtask and a detailed texture recovery subtask; designing different inverse tone mapping networks aiming at the subtasks of inverse tone mapping to obtain the recovered illumination component and reflection component, and then combining based on a multi-scale Retinex model to obtain a high dynamic range image; the method comprises the following steps:
1) decomposing the low dynamic range image into an illumination component and a reflection component based on a multi-scale Retinex model, wherein the illumination component represents the global illumination condition and the dynamic range of the image, and the reflection component represents the color and the detail information of the image;
decomposing a low dynamic range image into an illumination component and a reflection component based on a multi-scale Retinex model, specifically decomposing an original image in a multi-scale Gaussian filtering mode, and comprising the following operations:
11) filtering the original image by Gaussian filtering, wherein the filtering result is used as an illumination component;
12) subtracting the original image from the illumination component to obtain a reflection component;
thereby decomposing the image into two parts, an illumination component and a reflection component;
2) respectively designing different inverse tone mapping networks for the illumination component and the reflection component of the image; designing an illumination recovery network for the illumination component of the image, adopting a full convolution network with unchanged feature map size, and introducing a residual error structure; edge filling is carried out in a mirror symmetry mode, so that the size of the feature map is unchanged; adopting a U-Net convolution neural network as a reflection component recovery network for the reflection component of the image; the U-Net convolutional neural network adopted by the reflection component recovery network is of a 10-layer structure and comprises 5 convolutional layers and 5 deconvolution layers; in the deconvolution layer, on-line sampling is adopted, the resolution of the characteristic diagram is enlarged, and then convolution operation is carried out to avoid chessboard artifacts; adding batch normalization operation to each layer to accelerate convergence speed;
3) respectively recovering the illumination component and the reflection component through an illumination recovery network and a reflection component recovery network;
4) and combining the recovered illumination component and reflection component by using a multi-scale Retinex model to obtain a finally recovered high dynamic range image.
2. The method according to claim 1, wherein in step 2), the illumination recovery network comprises 7 layers, each layer comprising a convolutional layer and an active layer; the convolution kernel size is 3 × 3, the step size is 1, the number of the characteristic channels of the first six layers is 64, and the number of the characteristic channels of the last layer is 3.
3. The method of claim 2, wherein the SELU is used as the activation function.
4. The method of Retinex-based inverse tone-mapped image conversion of claim 1, wherein edge-filling is performed in a mirror-symmetric manner, specifically, pixel-filling is performed at the edges of the input feature map, and the values are pixel values of the original edges.
5. The method of Retinex-based inverse tone mapping image conversion according to claim 1, wherein said reflection component recovery network employs 5 convolutional layers of a U-Net convolutional neural network, the convolutional kernel size of the first 4 convolutional layers is 3 x 3, the step size is 2, and the number of characteristic channels is 64,128,256,512; the convolution kernel size of the convolution layer 5 is 3 x 3, the step size is 1, and the number of characteristic channels is 1024.
6. The method of Retinex-based inverse tone-mapped image conversion of claim 1, wherein, in particular, said reflected component recovery network employs deconvolution layers of U-Net convolutional neural networks having convolution kernel sizes of 3 x 3 and step sizes of 1; the number of the characteristic channels is 512,256,128,64 and 3 respectively.
7. The method as claimed in claim 1, wherein the combining in step 4) adds the recovered illumination component and reflection component by using a multi-scale Retinex model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910030455.0A CN109785263B (en) | 2019-01-14 | 2019-01-14 | Retinex-based inverse tone mapping image conversion method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910030455.0A CN109785263B (en) | 2019-01-14 | 2019-01-14 | Retinex-based inverse tone mapping image conversion method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109785263A CN109785263A (en) | 2019-05-21 |
| CN109785263B true CN109785263B (en) | 2022-09-16 |
Family
ID=66500464
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910030455.0A Active CN109785263B (en) | 2019-01-14 | 2019-01-14 | Retinex-based inverse tone mapping image conversion method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109785263B (en) |
Families Citing this family (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20220051375A1 (en) * | 2019-03-18 | 2022-02-17 | Peking University Shenzhen Graduate School | An inverse tone mapping method, system, device and computer readable medium |
| CN110223256A (en) * | 2019-06-10 | 2019-09-10 | 北京大学深圳研究生院 | A kind of inverse tone mapping (ITM) method, apparatus and electronic equipment |
| CN110276729B (en) * | 2019-06-10 | 2021-12-17 | 浙江工业大学 | Low-illumination color image enhancement method |
| CN110717868B (en) * | 2019-09-06 | 2022-05-03 | 上海交通大学 | Video high dynamic range inverse tone mapping model construction and mapping method and device |
| CN111683269B (en) * | 2020-06-12 | 2021-08-17 | 腾讯科技（深圳）有限公司 | Video processing method, video processing device, computer equipment and storage medium |
| CN117441186A (en) * | 2021-06-24 | 2024-01-23 | Oppo广东移动通信有限公司 | Image decoding and processing method, device and equipment |
Citations (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104346776A (en) * | 2013-08-02 | 2015-02-11 | 杭州海康威视数字技术股份有限公司 | Retinex-theory-based nonlinear image enhancement method and system |
| CN104574293A (en) * | 2014-11-28 | 2015-04-29 | 中国科学院长春光学精密机械与物理研究所 | Multiscale Retinex image sharpening algorithm based on bounded operation |
| WO2016083229A1 (en) * | 2014-11-27 | 2016-06-02 | Thomson Licensing | Scalable encoding and decoding methods and corresponding devices with combined intraprediction and interlayer prediction |
| CN105825472A (en) * | 2016-05-26 | 2016-08-03 | 重庆邮电大学 | Rapid tone mapping system and method based on multi-scale Gauss filters |
| CN105894456A (en) * | 2014-10-27 | 2016-08-24 | 南京林业大学 | High dynamic range image toning mapping method based on normalized layering |
| CN106709888A (en) * | 2017-01-09 | 2017-05-24 | 电子科技大学 | High-dynamic-range image generation method based on human eye visual model |
| CN106875358A (en) * | 2017-02-09 | 2017-06-20 | 聚龙智瞳科技有限公司 | Image enchancing method and image intensifier device based on Bayer format |
| US9697592B1 (en) * | 2015-12-30 | 2017-07-04 | TCL Research America Inc. | Computational-complexity adaptive method and system for transferring low dynamic range image to high dynamic range image |
| CN107045715A (en) * | 2017-02-22 | 2017-08-15 | 西南科技大学 | A kind of method that single width low dynamic range echograms generates high dynamic range images |
| CN107507138A (en) * | 2017-07-27 | 2017-12-22 | 北京大学深圳研究生院 | A kind of underwater picture Enhancement Method based on Retinex model |
| CN107657594A (en) * | 2017-09-22 | 2018-02-02 | 武汉大学 | The quick tone mapping method and system of a kind of high quality |
| CN108027961A (en) * | 2015-08-31 | 2018-05-11 | 汤姆逊许可公司 | Method and apparatus for inverse tone mapping (ITM) |
| CN108319966A (en) * | 2017-10-13 | 2018-07-24 | 西安科技大学 | The method for identifying and classifying of equipment in a kind of substation's complex background infrared image |
| EP3373586A1 (en) * | 2017-03-09 | 2018-09-12 | Thomson Licensing | Method for inverse tone mapping of an image with visual effects |
| CN108537746A (en) * | 2018-03-21 | 2018-09-14 | 华南理工大学 | A kind of fuzzy variable method for blindly restoring image based on depth convolutional network |
| CN108550124A (en) * | 2018-04-17 | 2018-09-18 | 重庆邮电大学 | A kind of illumination compensation and image enchancing method based on bionical helical |
| CN109034198A (en) * | 2018-06-25 | 2018-12-18 | 中国科学院计算技术研究所 | The Scene Segmentation and system restored based on characteristic pattern |
Family Cites Families (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US7876974B2 (en) * | 2003-08-29 | 2011-01-25 | Vladimir Brajovic | Method for improving digital images and an image sensor for sensing the same |
| WO2010105036A1 (en) * | 2009-03-13 | 2010-09-16 | Dolby Laboratories Licensing Corporation | Layered compression of high dynamic range, visual dynamic range, and wide color gamut video |
| GB2486348B (en) * | 2009-10-08 | 2014-11-12 | Ibm | Method and system for transforming a digital image from a low dynamic range (LDR) image to a high dynamic range (HDR) image |
| US8831340B2 (en) * | 2010-01-27 | 2014-09-09 | Adobe Systems Incorporated | Methods and apparatus for tone mapping high dynamic range images |
| WO2014077827A1 (en) * | 2012-11-16 | 2014-05-22 | Thomson Licensing | Processing high dynamic range images |
| US10475165B2 (en) * | 2017-04-06 | 2019-11-12 | Disney Enterprises, Inc. | Kernel-predicting convolutional neural networks for denoising |
-
2019
- 2019-01-14 CN CN201910030455.0A patent/CN109785263B/en active Active
Patent Citations (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104346776A (en) * | 2013-08-02 | 2015-02-11 | 杭州海康威视数字技术股份有限公司 | Retinex-theory-based nonlinear image enhancement method and system |
| CN105894456A (en) * | 2014-10-27 | 2016-08-24 | 南京林业大学 | High dynamic range image toning mapping method based on normalized layering |
| WO2016083229A1 (en) * | 2014-11-27 | 2016-06-02 | Thomson Licensing | Scalable encoding and decoding methods and corresponding devices with combined intraprediction and interlayer prediction |
| CN104574293A (en) * | 2014-11-28 | 2015-04-29 | 中国科学院长春光学精密机械与物理研究所 | Multiscale Retinex image sharpening algorithm based on bounded operation |
| CN108027961A (en) * | 2015-08-31 | 2018-05-11 | 汤姆逊许可公司 | Method and apparatus for inverse tone mapping (ITM) |
| US9697592B1 (en) * | 2015-12-30 | 2017-07-04 | TCL Research America Inc. | Computational-complexity adaptive method and system for transferring low dynamic range image to high dynamic range image |
| CN105825472A (en) * | 2016-05-26 | 2016-08-03 | 重庆邮电大学 | Rapid tone mapping system and method based on multi-scale Gauss filters |
| CN106709888A (en) * | 2017-01-09 | 2017-05-24 | 电子科技大学 | High-dynamic-range image generation method based on human eye visual model |
| CN106875358A (en) * | 2017-02-09 | 2017-06-20 | 聚龙智瞳科技有限公司 | Image enchancing method and image intensifier device based on Bayer format |
| CN107045715A (en) * | 2017-02-22 | 2017-08-15 | 西南科技大学 | A kind of method that single width low dynamic range echograms generates high dynamic range images |
| EP3373586A1 (en) * | 2017-03-09 | 2018-09-12 | Thomson Licensing | Method for inverse tone mapping of an image with visual effects |
| CN107507138A (en) * | 2017-07-27 | 2017-12-22 | 北京大学深圳研究生院 | A kind of underwater picture Enhancement Method based on Retinex model |
| CN107657594A (en) * | 2017-09-22 | 2018-02-02 | 武汉大学 | The quick tone mapping method and system of a kind of high quality |
| CN108319966A (en) * | 2017-10-13 | 2018-07-24 | 西安科技大学 | The method for identifying and classifying of equipment in a kind of substation's complex background infrared image |
| CN108537746A (en) * | 2018-03-21 | 2018-09-14 | 华南理工大学 | A kind of fuzzy variable method for blindly restoring image based on depth convolutional network |
| CN108550124A (en) * | 2018-04-17 | 2018-09-18 | 重庆邮电大学 | A kind of illumination compensation and image enchancing method based on bionical helical |
| CN109034198A (en) * | 2018-06-25 | 2018-12-18 | 中国科学院计算技术研究所 | The Scene Segmentation and system restored based on characteristic pattern |
Non-Patent Citations (2)
| Title |
|---|
| ExpandNet: A Deep Convolutional Neural Network for High Dynamic Range Expansion from Low Dynamic Range Content;D. Marnerides et al.;《COMPUTER GRAPHICS FORUM》;20180531;第37卷(第2期);第37-49页 * |
| 采用主成分分析与梯度金字塔的高动态范围图像生成方法;张淑芳 等;《西安交通大学学报》;20180430;第52卷(第4期);第150-157页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109785263A (en) | 2019-05-21 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109785263B (en) | Retinex-based inverse tone mapping image conversion method | |
| CN107403415B (en) | Compressed depth map quality enhancement method and device based on full convolution neural network | |
| Chen et al. | Gated context aggregation network for image dehazing and deraining | |
| CN112991203B (en) | Image processing method, device, electronic equipment and storage medium | |
| CN109410123B (en) | Deep learning-based mosaic removing method and device and electronic equipment | |
| CN109087269B (en) | Weak light image enhancement method and device | |
| CN110942424B (en) | Composite network single image super-resolution reconstruction method based on deep learning | |
| CN106485668A (en) | Mthods, systems and devices for overexposure correction | |
| CN113284064B (en) | Cross-scale context low-illumination image enhancement method based on attention mechanism | |
| US11783459B2 (en) | Method and device of inverse tone mapping and electronic device | |
| CN105765607A (en) | Multiple phase method for image deconvolution | |
| CN113781320A (en) | Image processing method and device, terminal equipment and storage medium | |
| JPH0944657A (en) | Method and device for processing image | |
| Zheng et al. | Ultra-high-definition image hdr reconstruction via collaborative bilateral learning | |
| Arulkumar et al. | Super resolution and demosaicing based self learning adaptive dictionary image denoising framework | |
| CN109767389A (en) | Adaptive weighted double blind super-resolution reconstruction methods of norm remote sensing images based on local and non local joint priori | |
| KR102327758B1 (en) | Apparatus and method of removing moire pattern in image using multi-domain deep learning | |
| Zheng et al. | Windowing decomposition convolutional neural network for image enhancement | |
| CN111626943B (en) | Total variation image denoising method based on first-order forward and backward algorithm | |
| WO2024055458A1 (en) | Image noise reduction processing method and apparatus, device, storage medium, and program product | |
| Akhil et al. | Single Image HDR Synthesis Using a Densely Connected Dilated ConvNet. | |
| CN112184587A (en) | Edge data enhancement model, and efficient edge data enhancement method and system based on model | |
| CN116029932A (en) | Ultra-high resolution image exposure error correction method and system | |
| CN115527258A (en) | Face exchange method based on identity information response | |
| US8180169B2 (en) | System and method for multi-scale sigma filtering using quadrature mirror filters |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |