CN103632357A - Image super-resolution enhancing method based on illumination separation - Google Patents
Image super-resolution enhancing method based on illumination separation Download PDFInfo
- Publication number
- CN103632357A CN103632357A CN201310037260.1A CN201310037260A CN103632357A CN 103632357 A CN103632357 A CN 103632357A CN 201310037260 A CN201310037260 A CN 201310037260A CN 103632357 A CN103632357 A CN 103632357A
- Authority
- CN
- China
- Prior art keywords
- image
- resolution
- low
- illumination
- frequency information
- 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.)
- Granted
Links
Images
Landscapes
- Image Processing (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses an image super-resolution enhancing method based on illumination separation. The method comprises that illumination separation is implemented on reference images, the correspondence between low-frequency information and high-frequency information of the images is learnt by utilizing a lot of reference images, and high-frequency information of a to-be-enhanced image is obtained by utilizing the correspondence, thereby obtaining a high-resolution image. According to the method, influence of external light change on image processing is fully considered, and the enhancing effect for the super-resolution images is greatly improved; one scene only requires images under the same light, so that the scale of a learning sample base is greatly reduced; due to reduction of the sample base scale, the scale of a prior knowledge base is also reduced; and thus, time can be saved and computational efficiency can be improved during high-dimensional data retrieval.
Description
Technical field
The invention belongs to technical field of image information processing, specifically, relate to a kind of image super-resolution Enhancement Method based on illumination separation, be i.e. a kind of super-resolution processing method by low-resolution image recovery full resolution pricture based on illumination separation.
Background technology
Along with the fast development of internet and the information processing technology, people are more and more higher to the requirement of image resolution ratio.But owing to being subject to physics image-forming condition and gathering environmental limit, the image resolution ratio collecting is very low sometimes, and the visual effect of image is difficult to meet people's needs.
Super-resolution image Enhancement Method is utilized the character of image itself and the related information of adjacent image piece, by certain algorithm process, in the situation that not increasing hardware input, from low resolution inferior quality input picture, estimate high resolving power high quality graphic, make image resolution ratio surpass the resolution of optical hardware equipment, obtain the potential detailed information of image, strengthen the availability of image.This method, once proposition, has just caused many scholars' extensive attention and concern.Harris[please refer to document J.LHarris.Diffraction and Resolving Power.Journal of Optical Society of American, Vol.54, No.7, pp.931-936,1964] prove theoretically, two different objects can not produce same image, and therefore, any image obtaining under noiseless condition is only corresponding with an object.Like this, according to the low-resolution image of an object, the details of this object is strengthened and has possibility.Through scientific research personnel's follow-up study, super-resolution image Enhancement Method has obtained more result, and its feasibility is undoubted.
Super-resolution image Enhancement Method has a wide range of applications in a plurality of fields such as vision monitoring, public safety, remote sensing, medical imaging and high-definition televisions.Aspect satellite remote sensing, because the resolution of image-forming condition and imaging device has certain limitation, be in general difficult to obtain the image of high definition, therefore, from the low-resolution image having obtained, generate high-definition picture tool and be of great significance; In Video transmission system, by expanded view picture and increase details, can from video sequence, obtain super-resolution still image, ordinary video signal is changed into high-definition TV signal; At public safety field, process the low-quality image causing due to move, defocus, environment etc., obtain high quality graphic; In living things feature recognition field, can obtain high-resolution facial image by Super-resolution Image Restoration method, this high-resolution human face image, for recognition of face, can improve accuracy of identification greatly.
Super-resolution Enhancement Method has become one of the most active research topic nearly ten years of image processing field in the world.Super-resolution Enhancement Method roughly can be divided into both direction: the method based on reconstruct and the method based on study.The main thought of the super-resolution Enhancement Method based on reconstruct is to utilize multiframe repeat the to take pictures redundant information of image, reconstructs super-resolution image, eliminates and reduces mixing effect.Super-resolution Enhancement Method based on reconstruct need to be set up the imaging model tallying with the actual situation and the model that degrades, and need to carry out accurate sub-pix estimation to image sequence, and this is all very difficult in actual treatment.Super-resolution Enhancement Method based on study is in the situation that the method based on reconstruct is met difficulty grows up, although start late, but it seems at present and can make up a lot of deficiencies of the method based on reconstruct, the development of the super-resolution Enhancement Method combined with intelligent technology based on study, image for particular type, effectively utilize priori, can effectively improve spatial resolution, to be worth the further direction of research, the present invention is exactly the super-resolution field belonging to based on study, to solve the key issue in this field.
Super-resolution Enhancement Method based on study [please refer to document S.Baker by the Baker in Ka Naiji-Mei Long laboratory etc. the earliest, and T.Kanade.Limits on Super-resolution and How to Break Them.Proc.of Computer Vision and Pattern Recognition, pp.372-379,2000] propose, they propose a kind of method strengthening based on priori, utilize machine learning algorithm to go training to specify classification image, the priori obtaining is strengthened for super-resolution.The Freeman of Massachusetts Institute of Technology (MIT) etc. [please refer to document W.T.Freeman, T.R.Jones, and E.C.Pasztor.Example-Based Super-Resolution.IEEE Computer Graphics and Applications, Vol.22, No.2, PP.56-65,2002] method based on sample has been proposed, utilize markov network to carry out the relation between low-resolution image and high-definition picture piece in learning training storehouse, with study to relation predict the detailed information of input low-resolution image.Bishop etc. [please refer to document C.M.Bishop, A.Blake, and B.Marthi.Super-Resolution Enhancement of Video.Proc.of the 9th International Conference on Artificial Intelligence and Statistics, pp.410-414,2003] utilize an image block data storehouse to obtain the Mid Frequency of natural image and the relation between high band, and the training set strengthening as present image by the part of the super-resolution image being enhanced, thereby added additional high-frequency information.Joshi etc. [please refer to document M.V.Joshi, S.Chaudhuri, and R.Panuganti.A Learning-Based Method for Image Super-Resolution from Zoomed Observations.IEEE Transactions on Systems, Man and Cybernetics, Vol.35, No.3, pp.527-537,2005] a kind of method that image that merges different size carries out resolution enhancing proposed.Super-resolution Enhancement Method based on study need to be searched for the information of coupling from extensive prior data bank, and the computational resource needing is very large, lacks the research of this respect in research work in the past, causes elapsed time long, has affected the application of the method.
The basic ideas of super-resolution Enhancement Method based on study are, utilize the corresponding relation between a large amount of reference pictures study image low-frequency information and high-frequency information, utilize this corresponding relation to obtain the high-frequency information of image to be strengthened, and obtain high-definition picture.The method is minute two steps conventionally, off-line training process and online enhancing process, wherein, off-line training process is to utilize the method for statistical learning, the image reference storehouse that analysis consists of a large amount of high-definition pictures and low-resolution image, set up low-frequency information in low-resolution image and the corresponding relation between high-frequency information, build knowledge base, its processing procedure as shown in Figure 1; In online enhancing process, to input picture piecemeal, for each image block extracts search characteristics vector, the high-frequency information that search is mated most in knowledge base obtains high-definition picture after merging, and its processing procedure as shown in Figure 2.
In above-mentioned existing image enchancing method, do not consider the impact of image light, but in actual applications, because changing, light cause the image change of the Same Scene obtained various, the Image Priori Knowledge storehouse of developing under a certain light condition and character representation method, under another light condition, can not apply well, this has just limited the range of application of the super-resolution image Enhancement Method based on study greatly.Therefore, in the urgent need to a kind of image enchancing method that is not subject to light variable effect, utilize limited training sample, when strengthening image detail, keep the continuity of the illumination of image.The present invention, on the basis of existing disposal route, has increased the processing links of illumination separation, can effectively solve above-mentioned problems of the prior art.
Summary of the invention
In order to solve above-mentioned problems of the prior art, the present invention proposes a kind of image super-resolution Enhancement Method based on illumination separation.The method comprises the following steps:
Step 1, sets up the training image storehouse of containing several high-definition pictures;
Step 4, the corresponding low-resolution image that described high-definition picture and described step 3 are obtained asks poor, obtains high-frequency information image;
Step 5, the high-frequency information image that the low-resolution image obtaining for described step 3 and described step 4 obtain, described low-resolution image and described high-frequency information image are divided into a plurality of overlapping high-frequency information image blocks and low-resolution image piece, according to the correspondence position of image block, set up a plurality of high-low-resolution image piece pair, and respectively high-frequency information image block and low-resolution image piece being extracted to high-frequency characteristic vector sum characteristics of low-frequency vector, the high-frequency characteristic vector sum characteristics of low-frequency vector that extraction is obtained is to forming priori storehouse;
Step 6, expands the low-resolution image wait strengthening to size after enhancing;
Step 7, utilizes illumination method for separating and processing, and the low-resolution image after expanding is carried out to illumination pretreatment, obtains illumination invariant image and light image;
Step 8, is divided into a plurality of overlapping image blocks by the illumination invariant image obtaining in described step 7, and each image block is extracted to the proper vector identical with the feature of extracting in described step 5;
Step 9, based on described step 8, extract the proper vector obtaining and in described priori storehouse, inquire about a plurality of proper vectors that the proper vector that obtains obtaining with described step 8 extraction is mated most, thereby obtain corresponding high-frequency information image block according to the corresponding relation between described proper vector or described proper vector;
Step 10, the illumination invariant image additive fusion obtaining in the high-frequency information image block that described step 9 is obtained and described step 7, obtains the high-definition picture of illumination invariant;
Step 11, carries out image co-registration to the high-definition picture of the light image obtaining in described step 7 and described illumination invariant, the high-definition picture after being enhanced.
The inventive method application is extensive, particularly in fields such as public security, monitoring, remote sensing, because prior art is subject to the restriction of acquisition condition, there is sometimes the problems such as resolution is not high, fuzzy in the image collecting, therefore greatly limited the actual use value of image, and the inventive method can promote the resolution of image effectively, strengthen the data value of image.
Accompanying drawing explanation
Fig. 1 is the off-line training process flow diagram flow chart of super-resolution Enhancement Method in prior art;
Fig. 2 is the online enhancing process flow diagram flow chart of super-resolution Enhancement Method in prior art;
Fig. 3 has shown the off-line training process of image super-resolution Enhancement Method according to an embodiment of the invention;
Fig. 4 has shown the online enhancing process of image super-resolution Enhancement Method according to an embodiment of the invention.
Embodiment
For making the object, technical solutions and advantages of the present invention clearer, below in conjunction with specific embodiment, and with reference to accompanying drawing, the present invention is described in more detail.
The definition of super-resolution is to rebuild high-definition picture by low-resolution image.Given single width low-resolution image, is divided into overlapping image block according to setting scheme, utilizes priori storehouse, estimates the high-definition picture piece of each image block, then through merging the high-definition picture being enhanced.In low-resolution image, lack and describe details high-frequency information, therefore, the object of enhancing is namely utilized priori storehouse, obtains high-frequency information, thereby high-frequency information and low-resolution image stack are produced to the image strengthening.
The present invention is directed to method of the prior art and be subject to light to affect problem bigger than normal, increased the link of illumination separation in processing procedure, utilizing the information of illumination invariant in image is foundation, obtains high-frequency information, finally increases illumination information, rediscover scene again.The inventive method is divided into two processes: off-line training process and online enhancing process.The general off-line of previous process carries out, and builds effective priori storehouse; A rear process is carried out online, for the image of input, obtains fast the image strengthening.The present invention, in off-line training process and online enhancing process, has all increased the processing procedure of illumination separation, impact super-resolution image being strengthened to reduce illumination.
Fig. 3 has shown the off-line training process of image super-resolution Enhancement Method according to an embodiment of the invention.The object of described off-line training process is to build a priori storehouse, and in the present invention, described priori storehouse is obtained by one group of high-resolution image set training.
As shown in Figure 3, the off-line training process of image super-resolution Enhancement Method of the present invention comprises following step:
Step 1, sets up the training image storehouse of containing several high-definition pictures;
Consider illumination variation on the impact of image normally the overall situation, it has determined the basic keynote of image chroma and brightness.Land[please refer to document E.H.Land.An Alternative Technique for the Computation of the Designator in the Retinex Theory of Color Vision.Proc.of National Academy of Science:National Academy of Science, UAS, 83, pp:3078-3080, 1986] how propose one regulates and perceives the color of object and the model Retinex algorithm of brightness about human visual system, it is in fact a kind of image processing algorithm based on illumination compensation, can be at gray scale dynamic range compression, edge strengthens and color constancy three aspects: reaches balance.By this algorithm, can separating reaction light image and the reaction object of extraneous light reflect essential illumination invariant image.The present invention utilizes this algorithm to carry out illumination separating treatment to high-definition picture exactly, obtains illumination invariant image and light image.
In this step, described in degrade to process and can adopt several different methods of the prior art, such as can first adopting Gaussian smoothing, then carry out down-sampled, but finally utilize interpolation method to obtain the image that unidimensional resolution is low.
Step 4, the corresponding low-resolution image that described high-definition picture and described step 3 are obtained asks poor, obtains high-frequency information image;
Step 5, the high-frequency information image that the low-resolution image obtaining for described step 3 and described step 4 obtain, adopt identical image block strategy, described low-resolution image and described high-frequency information image are divided into a plurality of overlapping high-frequency information image blocks and low-resolution image piece, according to the correspondence position of image block, set up a plurality of high-low-resolution image piece pair, and respectively high-frequency information image block and low-resolution image piece are extracted to high-frequency characteristic vector sum characteristics of low-frequency vector, the high-frequency characteristic vector sum characteristics of low-frequency vector that extraction is obtained is to forming priori storehouse,
In this step, described high-frequency characteristic vector general direct in high-frequency information image block corresponding image information form, and described characteristics of low-frequency vector can consist of gray feature, the Gradient Features of corresponding low-resolution image piece, also can be formed by the statistical nature of image block, as histogram etc., described high-frequency characteristic vector sum characteristics of low-frequency vector has formed the locating vector of respective image piece.
The object of super-resolution Enhancement Method of the present invention is to recover high-frequency information that high-definition picture possesses and that low-resolution image lacks.Therefore, in training image real valuable be that the difference information of low-resolution image and original high resolution image is high-frequency information image, and the corresponding relation between difference information and low-resolution image.Above-mentioned off-line training process has been set up this relation by priori storehouse just.
Fig. 4 has shown the online enhancing process of image super-resolution Enhancement Method according to an embodiment of the invention.As shown in Figure 4, described online enhancing process comprises following step:
Step 6, by low-resolution image to be strengthened, such as by interpolation method, expands the size after enhancing to;
Step 7, utilizes illumination method for separating and processing, and the low-resolution image after expanding is carried out to illumination pretreatment, obtains illumination invariant image and light image;
Step 8, for the illumination invariant image obtaining in described step 7, is divided into a plurality of overlapping image blocks according to method identical in same training process, and each image block is extracted to the proper vector identical with the feature of extracting in described step 5;
Step 9, based on described step 8, extract the proper vector obtaining and in described priori storehouse, inquire about a plurality of proper vectors that the proper vector that obtains obtaining with described step 8 extraction is mated most, thereby obtain corresponding high-frequency information image block according to the corresponding relation between described proper vector or described proper vector;
Proper vector matching process in this step, can adopt traditional vector similarity computing method, such as carrying out vector similarity calculating according to Euclidean distance, weighted euclidean distance, mahalanobis distance etc.
Step 10, the illumination invariant image additive fusion obtaining in the high-frequency information image block that described step 9 is obtained and described step 7, obtains the high-definition picture of illumination invariant;
Image additive fusion method in this step can adopt the fusion method of Pixel-level, such as image interfusion methods such as average of the prior art, weighted sums.
Step 11, carries out image co-registration to the high-definition picture of the light image obtaining in described step 7 and described illumination invariant, the high-definition picture after being enhanced.
The obtained advantage of the inventive method has following several aspect:
1) promote the effect that super-resolution image strengthens
The present invention can take into full account extraneous light and change the impact that image is processed, thereby has greatly promoted the effect that super-resolution image strengthens.
2) reduce the Sample Storehouse of study
The present invention, for the image of same scene, only needs the image under a light, thereby has greatly reduced the scale in learning sample storehouse.
3) improve counting yield
The present invention is due to the reduction of the scale of Sample Storehouse, and the scale in the priori storehouse that training obtains also decreases, and therefore, when carrying out high dimensional data retrieval, can save time, and improves operation efficiency.
Above-described specific embodiment; object of the present invention, technical scheme and beneficial effect are further described; institute is understood that; the foregoing is only specific embodiments of the invention; be not limited to the present invention; within the spirit and principles in the present invention all, any modification of making, be equal to replacement, improvement etc., within all should being included in protection scope of the present invention.
Claims (10)
1. the image super-resolution Enhancement Method based on illumination separation, is characterized in that, the method comprises the following steps:
Step 1, sets up the training image storehouse of containing several high-definition pictures;
Step 2, carries out illumination separating treatment to each the panel height image in different resolution in described training image storehouse, obtains illumination invariant image and light image;
Step 3, the described illumination invariant image that described step 2 is obtained degrades to process and obtains corresponding low-resolution image;
Step 4, the corresponding low-resolution image that described high-definition picture and described step 3 are obtained asks poor, obtains high-frequency information image;
Step 5, the high-frequency information image that the low-resolution image obtaining for described step 3 and described step 4 obtain, described low-resolution image and described high-frequency information image are divided into a plurality of overlapping high-frequency information image blocks and low-resolution image piece, according to the correspondence position of image block, set up a plurality of high-low-resolution image piece pair, and respectively high-frequency information image block and low-resolution image piece being extracted to high-frequency characteristic vector sum characteristics of low-frequency vector, the high-frequency characteristic vector sum characteristics of low-frequency vector that extraction is obtained is to forming priori storehouse;
Step 6, expands the low-resolution image wait strengthening to size after enhancing;
Step 7, utilizes illumination method for separating and processing, and the low-resolution image after expanding is carried out to illumination pretreatment, obtains illumination invariant image and light image;
Step 8, is divided into a plurality of overlapping image blocks by the illumination invariant image obtaining in described step 7, and each image block is extracted to the proper vector identical with the feature of extracting in described step 5;
Step 9, based on described step 8, extract the proper vector obtaining and in described priori storehouse, inquire about a plurality of proper vectors that the proper vector that obtains obtaining with described step 8 extraction is mated most, thereby obtain corresponding high-frequency information image block according to the corresponding relation between described proper vector or described proper vector;
Step 10, the illumination invariant image additive fusion obtaining in the high-frequency information image block that described step 9 is obtained and described step 7, obtains the high-definition picture of illumination invariant;
Step 11, carries out image co-registration to the high-definition picture of the light image obtaining in described step 7 and described illumination invariant, the high-definition picture after being enhanced.
2. method according to claim 1, is characterized in that, in described step 3, adopts successively Gaussian smoothing, down-sampled and interpolation to obtain described low-resolution image.
3. method according to claim 1, is characterized in that, described low-resolution image and described illumination invariant image are unidimensional.
4. method according to claim 1, is characterized in that, in step 5, described high-frequency characteristic vector directly in high-frequency information image block corresponding image information form.
5. method according to claim 1, is characterized in that, in step 5, described characteristics of low-frequency vector consists of gray feature, Gradient Features or the statistical nature of corresponding low-resolution image piece.
6. method according to claim 1, is characterized in that, described statistical nature is histogram.
7. method according to claim 1, is characterized in that, in described step 6, by interpolation method, will expand low-resolution image to be strengthened.
8. method according to claim 1, is characterized in that, in described step 9, adopts vector similarity computing method to mate proper vector.
9. method according to claim 8, is characterized in that, described vector similarity computing method are for to carry out vector similarity calculating according to Euclidean distance, weighted euclidean distance or mahalanobis distance.
10. method according to claim 1, is characterized in that, in described step 10 and described step 11, adopts the fusion method of Pixel-level to merge image.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201310037260.1A CN103632357B (en) | 2013-01-30 | 2013-01-30 | A kind of image super-resolution Enhancement Method separated based on illumination |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201310037260.1A CN103632357B (en) | 2013-01-30 | 2013-01-30 | A kind of image super-resolution Enhancement Method separated based on illumination |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN103632357A true CN103632357A (en) | 2014-03-12 |
| CN103632357B CN103632357B (en) | 2016-08-03 |
Family
ID=50213374
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201310037260.1A Active CN103632357B (en) | 2013-01-30 | 2013-01-30 | A kind of image super-resolution Enhancement Method separated based on illumination |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN103632357B (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106780318A (en) * | 2015-11-25 | 2017-05-31 | 北京大学 | The unreal structure method of face and the unreal construction system of face |
| CN108460722A (en) * | 2018-01-31 | 2018-08-28 | 中国科学院上海技术物理研究所 | A kind of high-resolution wide visual field rate remotely sensed image method and device |
| CN111369475A (en) * | 2020-03-26 | 2020-07-03 | 北京百度网讯科技有限公司 | Method and apparatus for processing video |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101923721A (en) * | 2010-08-31 | 2010-12-22 | 汉王科技股份有限公司 | Non-illumination face image reconstruction method and system |
| CN101976435A (en) * | 2010-10-07 | 2011-02-16 | 西安电子科技大学 | Combination learning super-resolution method based on dual constraint |
-
2013
- 2013-01-30 CN CN201310037260.1A patent/CN103632357B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101923721A (en) * | 2010-08-31 | 2010-12-22 | 汉王科技股份有限公司 | Non-illumination face image reconstruction method and system |
| CN101976435A (en) * | 2010-10-07 | 2011-02-16 | 西安电子科技大学 | Combination learning super-resolution method based on dual constraint |
Non-Patent Citations (2)
| Title |
|---|
| HUANG K. RUIMIN HU ET AL.: "A face super-resolution method based on illumination invariant feature", 《MULTIMEDIA TECHNOLOGY(ICMT),2011 INTERNATIONAL CONFERENCE ON》 * |
| 乔建苹: "基于Log-WT的人脸图像超分辨率重建", 《电子与信息学报》 * |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106780318A (en) * | 2015-11-25 | 2017-05-31 | 北京大学 | The unreal structure method of face and the unreal construction system of face |
| CN106780318B (en) * | 2015-11-25 | 2020-07-14 | 北京大学 | Face magic structure method and face magic structure system |
| CN108460722A (en) * | 2018-01-31 | 2018-08-28 | 中国科学院上海技术物理研究所 | A kind of high-resolution wide visual field rate remotely sensed image method and device |
| CN111369475A (en) * | 2020-03-26 | 2020-07-03 | 北京百度网讯科技有限公司 | Method and apparatus for processing video |
| CN111369475B (en) * | 2020-03-26 | 2023-06-23 | 北京百度网讯科技有限公司 | Method and apparatus for processing video |
Also Published As
| Publication number | Publication date |
|---|---|
| CN103632357B (en) | 2016-08-03 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Song et al. | Monocular depth estimation using laplacian pyramid-based depth residuals | |
| CN110706157B (en) | Face super-resolution reconstruction method for generating confrontation network based on identity prior | |
| Chen et al. | Searching for efficient multi-scale architectures for dense image prediction | |
| CN110689482A (en) | Face super-resolution method based on supervised pixel-by-pixel generation countermeasure network | |
| CN109886159B (en) | Face detection method under non-limited condition | |
| CN110334607B (en) | Video human interaction behavior identification method and system | |
| Gao et al. | A review of dynamic maps for 3D human motion recognition using ConvNets and its improvement | |
| Zhu et al. | A hybrid algorithm for automatic segmentation of slowly moving objects | |
| Du et al. | Srh-net: Stacked recurrent hourglass network for stereo matching | |
| Wang et al. | PACCDU: Pyramid attention cross-convolutional dual UNet for infrared and visible image fusion | |
| Liao et al. | Residual attention unit for action recognition | |
| CN103632357A (en) | Image super-resolution enhancing method based on illumination separation | |
| CN110555406A (en) | Video moving target identification method based on Haar-like characteristics and CNN matching | |
| Zhang et al. | Dyna-depthformer: Multi-frame transformer for self-supervised depth estimation in dynamic scenes | |
| Zhang et al. | Dual feature enhanced video super-resolution network based on low-light scenarios | |
| Feng et al. | Coal mine image dust and fog clearing algorithm based on deep learning network | |
| Wang et al. | Gait Recognition based on lightweight CNNs | |
| Cai et al. | Human action recognition based on improved FCN framework | |
| Liu et al. | Super-resolution reconstruction of thermal imaging of power equipment based on Generative Adversarial Network with Channel Filtering | |
| Zhu et al. | Human Pose Estimation Based on Deep Neural Network | |
| CN113887371B (en) | Data enhancement method for low-resolution face recognition | |
| Fan et al. | A ConvNet structure learning spatiotemporal features for gesture recognition | |
| Zhou et al. | 3D Deformable Kernels for Video super-resolution | |
| Zhang et al. | Temporal Squeeze-and-Excitation Networks for Skeleton-Based Action Recognition | |
| Yu et al. | Fusion of shallow and deep features for gait recognition through multiple walking views |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant |