CN111340695A - Super-resolution reconstruction method of dome screen video - Google Patents
Super-resolution reconstruction method of dome screen video Download PDFInfo
- Publication number
- CN111340695A CN111340695A CN202010083678.6A CN202010083678A CN111340695A CN 111340695 A CN111340695 A CN 111340695A CN 202010083678 A CN202010083678 A CN 202010083678A CN 111340695 A CN111340695 A CN 111340695A
- Authority
- CN
- China
- Prior art keywords
- spherical screen
- resolution reconstruction
- screen video
- image
- neural network
- 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
Links
- 238000000034 method Methods 0.000 title claims abstract description 26
- 238000012937 correction Methods 0.000 claims abstract description 32
- 238000013528 artificial neural network Methods 0.000 claims abstract description 19
- 238000012545 processing Methods 0.000 claims abstract description 13
- 230000004913 activation Effects 0.000 claims abstract description 11
- 238000000605 extraction Methods 0.000 claims description 16
- 239000013598 vector Substances 0.000 claims description 14
- 238000010606 normalization Methods 0.000 claims description 10
- 230000003287 optical effect Effects 0.000 claims description 10
- 238000012821 model calculation Methods 0.000 claims description 9
- 238000013507 mapping Methods 0.000 description 9
- 230000000694 effects Effects 0.000 description 6
- 238000004422 calculation algorithm Methods 0.000 description 5
- 238000012549 training Methods 0.000 description 5
- 238000004364 calculation method Methods 0.000 description 3
- 239000003086 colorant Substances 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000004806 packaging method and process Methods 0.000 description 2
- 230000008569 process Effects 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 208000037170 Delayed Emergence from Anesthesia Diseases 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000006870 function Effects 0.000 description 1
- 238000003384 imaging method Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000012886 linear function Methods 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 238000003909 pattern recognition Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T3/00—Geometric image transformation in the plane of the image
- G06T3/40—Scaling the whole image or part thereof
- G06T3/4053—Super resolution, i.e. output image resolution higher than sensor resolution
-
- G06T5/73—
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10016—Video; Image sequence
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR 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; CALCULATING OR 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]
Abstract
The invention relates to a super-resolution reconstruction method of a spherical screen video, which comprises the following steps: introducing distortion correction parameters into the spherical screen video image to perform 3D correction processing; and (4) extracting the features of the spherical screen video image, and performing super-resolution reconstruction on the extracted features through a wide activation WDSR-B neural network. The jump connection and splicing operation between the shallow layer and the deep layer is introduced, based on a residual SISR network, the simple feature expansion before Relu can be obviously improved on the premise of no additional parameters and computation amount, and the feature expansion before Relu allows more information to pass through, and meanwhile, the high nonlinearity of the deep neural network is still maintained. The low-level SISR characteristics from the shallow layer can be more easily transmitted to the last layer to obtain better dense pixel value prediction, and the technology is applied to the dome cinema, so that the definition of images is improved, and the viewing experience of audiences is better.
Description
Technical Field
The invention relates to a super-resolution reconstruction technology of a dome screen video image, in particular to a super-resolution reconstruction method of a dome screen video.
Background
With the continuous development of pattern recognition and artificial intelligence, the super-resolution reconstruction technology has high application value in many fields such as video monitoring, medical images, satellite remote sensing and the like. The super-resolution reconstruction technique of images refers to restoring a given low-resolution image into a corresponding high-resolution image through a specific algorithm. The spherical screen video image applied to the spherical screen cinema keeps the scenes except the center of the picture unchanged, and other scenes which are supposed to be horizontal or vertical are correspondingly changed, so that the resolutions of different areas of the imaging surface of the spherical screen lens are different, the closer to the center of the picture, the higher the resolution, the more detailed information, the more deviated from the center of the picture, the lower the resolution, the less detailed information and the more serious deformation are caused.
The former image super-resolution network uses relatively shallow convolutional neural networks, which have poor accuracy, and the increase of depth can more effectively improve the resolution, but does not effectively utilize the characteristic information from the shallow layer, thereby increasing the time complexity.
At present, image processing is directly carried out on a distorted spherical screen video image, but because the spherical screen video image data are stored in a nonlinear mode and cannot be directly processed, the processing mode cannot obtain a good image deblurring effect.
In the hyper-resolution reconstruction algorithm, shallow information is not effectively utilized in the deep layer at present, the training network is very complex, and even a plurality of redundant convolutional layers exist, so that the influence on the hyper-resolution reconstruction effect of only considering a single image is not large, but when the hyper-resolution reconstruction algorithm is applied to a video, the speed is very low, the consumed time is long, and the real-time performance is poor under the condition that the network is complex.
Disclosure of Invention
In order to solve the problem of the definition of a spherical screen video image, the invention provides a hyper-resolution reconstruction method of a spherical screen video.
In order to solve the problems, the invention adopts the following technical scheme:
a method for hyper-resolution reconstruction of a dome screen video comprises the following steps:
s1, extracting an image from a spherical screen video by frames, and introducing distortion correction parameters into the image to perform 3D correction processing;
s2, extracting the characteristics of the corrected image;
and S3, carrying out hyper-resolution reconstruction on the extracted features through a wide-activation WDSR-B neural network.
Preferably, in the 3D correction in S1, a projection plane is set, and a mapping relationship is established between a distance r from a spherical screen pixel point to an optical axis center and a projection plane light incident angle θ'; and calculating a projection plane light ray incidence angle according to the distance meter from the spherical screen pixel point to the center of the optical axis through the mapping relation, and performing back projection on the projection plane light ray incidence angle to the plane to obtain a corrected point.
Preferably, the 3D correction is calculated according to different equations for different angles of incidence:
when the light incidence angle is less than or equal to a first threshold value, an orthogonal model calculation mode is adopted:
r＝fθ′；
when the first threshold value is larger than the light ray incidence angle and smaller than or equal to the second threshold value, an equiangular model calculation mode is adopted:
r＝2·f sinθ′/2；
when the light incidence angle is larger than a second threshold value, adopting an equidistant model calculation mode:
r＝f sinθ′；
f is the distance from the center of the optical axis to the plane of the spherical screen.
Preferably, the S2 performing feature extraction on the image after the rectification processing includes performing feature extraction on brightness, edges, textures and/or colors.
Preferably, the wide-activation WDSR-B neural network comprises a feature extraction module, a residual module, and an output module; the characteristic extraction module adopts the image characteristics extracted by the convolution layer; extracting the characteristics of the potential region by a residual error module; and the output module performs convolution on the extracted features and the residual error output by the residual error module through the convolution layer and then outputs the convolution result.
Preferably, the residual module includes sequentially arranged 1 × 1 convolutional layers for expanding the number of the same tracks, convolutional layers for feature extraction, a Relu activation link for adding a nonlinear feature, 1 × 1 convolutional layer for shallow feature extraction, and two 3 × 3 convolutional layers.
Preferably, the S3 further includes weight normalizing the WDSR-B neural network, the weight normalized parameterized weight vector being:
v is a k-dimensional vector, k being the dimension of the extracted feature; g is a scalar, w | | | is g; i v i is the euclidean norm of v, which will result in g, independent of the parameter v.
The technical scheme of the invention has the following beneficial technical effects:
(1) the jump connection and splicing operation between the shallow layer and the deep layer is introduced, based on a residual SISR network, the simple feature expansion before Relu can be obviously improved on the premise of no additional parameters and computation amount, and the feature expansion before Relu allows more information to pass through, and meanwhile, the high nonlinearity of the deep neural network is still maintained. The low-level SISR characteristics from the shallow layer can be more easily transmitted to the last layer to obtain better dense pixel value prediction, and the technology is applied to the dome cinema, so that the definition of images is improved, and the viewing experience of audiences is better.
(2) And 3D correction processing is carried out on the fisheye image of the spherical screen video, distortion correction parameters are introduced, and then characteristic extraction is carried out to obtain a better image deblurring effect. Reflecting each point of the spherical curtain to a plane, so that pixel holes or discontinuity cannot be formed; the correction method is introduced into the spherical screen video, and the problem that the spherical screen video and the common video are different in spatial special distribution is solved.
Drawings
FIG. 1 is a schematic diagram of a hyper-resolution reconstruction method for a dome screen video;
FIG. 2 is a schematic diagram of a spherical screen video image 3D correction model of a spherical screen video hyper-resolution reconstruction method;
fig. 3 is a flowchart of the hyper-resolution reconstruction of the dome video.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention will be described in further detail with reference to the accompanying drawings in conjunction with the following detailed description. It should be understood that the description is intended to be exemplary only, and is not intended to limit the scope of the present invention. Moreover, in the following description, descriptions of well-known structures and techniques are omitted so as to not unnecessarily obscure the concepts of the present invention.
Referring to fig. 1, a method for super-resolution reconstruction of a dome screen video includes: 3D correction processing of a dome video image and wide activation WDSR-B processing.
Referring to fig. 2, the dome video image 3D correction process:
the spherical screen video image is from a plane video image, and each frame of video image in the video is obtained. Images of certain frames may also be selected. And 3D correction processing is carried out on the image to obtain a corresponding projection mode from the plane to the spherical screen.
Introducing distortion correction parameters into the spherical screen video image to perform 3D correction processing, establishing a 3D correction model, establishing a mapping relation between the distance from a spherical screen pixel point P to an optical axis center 0 and a space object light incident angle, then giving a projection plane 0', and reversely projecting the projection plane along the solved incident angle to obtain a corrected point. And further establishing a mapping relation between each point on the spherical screen and each point on the plane, and calculating projection points on the spherical screen corresponding to each point of the image according to the mapping relation to finish 3D correction. And performing feature extraction on brightness, edges, textures, colors and other data with features of the corrected image.
The 3D correction model is used for establishing a mapping relation between the distance from the spherical screen video image point to the center of the optical axis and the incident angle of the light rays of the space object.
D is a correction model;
r is the distance from the spherical screen pixel point to the center of the optical axis;
theta is an included angle between the incident ray and the plumb line; theta' is an included angle between the incident ray of the projection plane and the plumb line;
psi, psi' are the azimuth angles (radians) before and after correction of the incident ray;
p is a pixel point of the space object light ray on the spherical screen;
p' is the point where the incident ray is back projected onto the correction plane.
In one embodiment, in conjunction with FIG. 2, the 3D correction is calculated using different equations based on different angles of incidence.
When the light incidence angle is less than or equal to a first threshold value, an orthogonal model calculation mode is adopted:
r＝fθ′；
when the first threshold value is larger than the light ray incidence angle and smaller than or equal to the second threshold value, an equiangular model calculation mode is adopted:
r＝2·f sinθ′/2；
when the light incidence angle is larger than a second threshold value, adopting an equidistant model calculation mode:
r＝f sinθ′；
r is the distance from the spherical screen pixel point to the center of the optical axis; theta' is an included angle between the incident ray and the main shaft; f is the focal length of the spherical screen; the 3D correction is performed using different calculation methods by different incident angles. The first threshold value is, for example, 60 °, and the second threshold value is, for example, 65 °. And adjusting according to the angle of the actual plane video shooting lens.
The azimuth angle ψ' of the incident ray is substantially not distorted in the actual screen video image or is distorted to a negligible extent, i.e., the equation: since ψ 'is always true, the model d can be expressed as θ'd (r). The formula of theta ═ D (r) only contains one parameter of the distance r between the point on the spherical screen video image and the optical axis, and the point of the mapping relation between r and theta' after 3D correction through the formula is located on the projection plane to form a plane image.
For the wide-activation WDSR-B neural network, features are expanded before a ReLU activation layer, and the wide activation effect in a residual block is considered, the most basic mode is to directly increase the number of channels in all the features, namely, the number of channels is increased by using convolution of 3X3, then an activation function is used, and then the number of channels is reduced by using convolution of 3X3, wherein the structure mainly aims at the condition that the expansion multiple is small (the value of the structure is between 2 and 4). When the value is large, the effect is rapidly reduced, and the method is not applicable.
As shown in fig. 1, in order to solve the above limitation, in A WDSR-B neural network, A large convolution kernel in an original WDSR- A neural network is split into two low-rank convolution kernels while keeping the number of equivalent mapping path channels unchanged, so as to save parameters. The WDSR-B neural network comprises a feature extraction module, a residual error module and an output module; the characteristic extraction module adopts the image characteristics extracted by the convolution layer; extracting the characteristics of the potential region by a residual error module; and the output module performs convolution on the extracted features and the residual error output by the residual error module through the convolution layer and then outputs the convolution result. The residual module includes 1 × 1 convolutional layer, nonlinear (Relu) activation, 1 × 1 convolutional layer, two 3 × 3 convolutional layers.
First the number of channels is enlarged with 1 x 1 convolutional layers (Conv), then a non-linear function (Relu) is applied after the convolutional layers, further proposing an efficient linear low-rank convolution, which decomposes a large convolution kernel into two low-rank convolution kernels. The method is a combination of 1 × 1Conv for reducing the number of channels and 3 × 3Conv for extracting spatial features, further increases the number of feature map channels before Relu under the condition of the same parameters, and can improve the resolution ratio without increasing the time complexity so as to make the dome screen video image clearer.
On the premise of not increasing the calculation overhead, the number of filters of the convolution kernel before the Relu activation layer is increased so as to increase the width of the feature mapping.
The residual network of the wide-active WDSR-B effectively reduces the time complexity. And a plurality of redundant convolution layers are removed, the time complexity is reduced, and the real-time performance of the dome screen video is met.
Weight normalization is the reparameterization of weight vectors in neural networks, and the length of those weight vectors and their direction can be decoupled. It has no dependency on the samples in the mini-batch, and has the same formula when training and testing.
When the model reaches a certain depth (180 layers), the model is difficult to train, and therefore weight normalization operation is provided. The method aims to recalculate the distribution of the network in the network, introduce weight normalization to allow training with higher learning rate, and improve the accuracy of training and testing.
The weight normalization algorithm is as follows:
y＝w.x+b (1)
where w is a k-dimensional vector matrix; x is a vector of input features in k dimensions; y is a vector of output features; b is a bias, a parameter for adjusting balance; the k is a dimension;
weight normalization parameterized weight vector algorithm:
where v is a k-dimensional vector, k being the dimension of the extracted feature; g is a scalar quantity, has no direction and only represents the size; if v is the euclidean norm of v, then we will get w, which is independent of the parameter v; w ranges between 0 and 1.
W in the formula (2) affects y in the formula (1), and y represents the output feature vector, so that weight normalization w affects the y output feature vector, and the accuracy of the output feature vector is improved by adding a weight normalization method.
And training the neural network model by using a sample, packaging and packaging after meeting the precision requirement, and performing over-resolution reconstruction after correcting the image.
Referring to fig. 3, the invention provides a method for hyper-resolution reconstruction of a dome video, which is mainly implemented by the following steps:
s1, extracting an image from a spherical screen video by frames, and introducing distortion correction parameters into the image to perform 3D correction processing;
s2, extracting the characteristics of the corrected image;
and S3, carrying out hyper-resolution reconstruction on the extracted features through a wide-activation WDSR-B neural network.
And in the process of wide activation of the WDSR-B neural network, weight normalization is adopted to improve the accuracy of the spherical screen video image.
Through the steps, various jump connection and splicing operations between the shallow layer and the deep layer are realized, based on the residual SISR network, the simple feature expansion before Relu can be obviously improved on the premise of no additional parameters and calculation amount, and the feature expansion before Relu allows more information to pass through, and meanwhile, the high nonlinearity of the deep neural network is still maintained. Low-level SISR features from shallow layers can be more easily passed to the last layer for better dense pixel value prediction.
In summary, the present invention relates to a method for super-resolution reconstruction of a dome screen video, including: introducing distortion correction parameters into the spherical screen video image to perform 3D correction processing; and (4) extracting the features of the spherical screen video image, and performing super-resolution reconstruction on the extracted features through a wide activation WDSR-B neural network. The jump connection and splicing operation between the shallow layer and the deep layer is introduced, based on a residual SISR network, the simple feature expansion before Relu can be obviously improved on the premise of no additional parameters and computation amount, and the feature expansion before Relu allows more information to pass through, and meanwhile, the high nonlinearity of the deep neural network is still maintained. The low-level SISR characteristics from the shallow layer can be more easily transmitted to the last layer to obtain better dense pixel value prediction, and the technology is applied to the dome cinema, so that the definition of images is improved, and the viewing experience of audiences is better.
It is to be understood that the above-described embodiments of the present invention are merely illustrative of or explaining the principles of the invention and are not to be construed as limiting the invention. Therefore, any modification, equivalent replacement, improvement and the like made without departing from the spirit and scope of the present invention should be included in the protection scope of the present invention. Further, it is intended that the appended claims cover all such variations and modifications as fall within the scope and boundaries of the appended claims or the equivalents of such scope and boundaries.
Claims (7)
1. A method for hyper-resolution reconstruction of a dome screen video is characterized by comprising the following steps:
s1, extracting an image from a spherical screen video by frames, and introducing distortion correction parameters into the image to perform 3D correction processing;
s2, extracting the characteristics of the corrected image;
and S3, carrying out hyper-resolution reconstruction on the extracted features through a wide-activation WDSR-B neural network.
2. The method for super-resolution reconstruction of dome video according to claim 1, wherein the 3D correction in S1 includes: setting a projection plane, and calculating the distance r from a projection point on the spherical screen to the center of an optical axis through a light incident angle theta' passing through the point on the projection screen; the azimuth angles of the projection points on the spherical screen and the projection points on the projection screen are the same, and then the corresponding projection points on the spherical screen are obtained.
3. The method of claim 2, wherein the 3D correction is calculated according to different formula based on different light incidence angles θ':
when the light incidence angle is less than or equal to a first threshold value, an orthogonal model calculation mode is adopted:
r＝fθ′；
when the first threshold value is larger than the light ray incidence angle and smaller than or equal to the second threshold value, an equiangular model calculation mode is adopted:
r＝2·f sinθ′/2；
when the light incidence angle is larger than a second threshold value, adopting an equidistant model calculation mode:
r＝f sinθ′；
f is the distance from the center of the optical axis to the plane of the spherical screen.
4. The method for reconstructing the dome screen video according to claim 1 or 2, wherein the S2 performs feature extraction on the rectified image, including performing feature extraction on brightness, edge, texture and/or color.
5. The method of claim 1 or 2, wherein the wide-activation WDSR-B neural network comprises a feature extraction module, a residual error module and an output module; the characteristic extraction module adopts the image characteristics extracted by the convolution layer; extracting the characteristics of the potential region by a residual error module; and the output module performs convolution on the extracted features and the residual error output by the residual error module through the convolution layer and then outputs the convolution result.
6. The method according to claim 5, wherein the residual module comprises sequentially arranged 1 x 1 convolutional layers for increasing the number of the same channel, convolutional layers for feature extraction, Relu activation links for adding nonlinear features, 1 x 1 convolutional layers for shallow feature extraction, and two 3x3 convolutional layers.
7. The method of super-resolution reconstruction of dome video according to claim 1 or 2, wherein said S3 further comprises performing weight normalization on WDSR-B neural network, the weight normalization parameterized weight vector is:
v is a k-dimensional vector, k being the dimension of the extracted feature; g is a scalar, w | | | is g; i v i is the euclidean norm of v, which will result in g, independent of the parameter v.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010083678.6A CN111340695A (en) | 2020-02-10 | 2020-02-10 | Super-resolution reconstruction method of dome screen video |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010083678.6A CN111340695A (en) | 2020-02-10 | 2020-02-10 | Super-resolution reconstruction method of dome screen video |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN111340695A true CN111340695A (en) | 2020-06-26 |
Family
ID=71183852
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010083678.6A Pending CN111340695A (en) | 2020-02-10 | 2020-02-10 | Super-resolution reconstruction method of dome screen video |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111340695A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111832463A (en) * | 2020-07-07 | 2020-10-27 | 哈尔滨理工大学 | Deep learning-based traffic sign detection method |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102663734A (en) * | 2012-03-15 | 2012-09-12 | 天津理工大学 | Fish eye lens calibration and fish eye image distortion correction method |
| CN103533235A (en) * | 2013-09-17 | 2014-01-22 | 北京航空航天大学 | Quick digital panoramic device based on linear array charge coupled device (CCD) for great case/event scene |
| CN104067372A (en) * | 2012-01-27 | 2014-09-24 | 塞莫费雪科学（不来梅）有限公司 | Multi-reflection mass spectrometer |
| CN106469448A (en) * | 2015-06-26 | 2017-03-01 | 康耐视公司 | Carry out automatic industrial inspection using 3D vision |
| CN106842520A (en) * | 2017-03-30 | 2017-06-13 | 中山联合光电科技股份有限公司 | A kind of high definition panorama looks around optical imaging system |
| CN107240066A (en) * | 2017-04-28 | 2017-10-10 | 天津大学 | Image super-resolution rebuilding algorithm based on shallow-layer and deep layer convolutional neural networks |
| CN108961151A (en) * | 2018-05-08 | 2018-12-07 | 中德（珠海）人工智能研究院有限公司 | A method of the three-dimensional large scene that ball curtain camera obtains is changed into sectional view |
| CN108983394A (en) * | 2018-08-23 | 2018-12-11 | 上海帛视光电科技有限公司 | A kind of fish eye lens system, image acquisition device |
-
2020
- 2020-02-10 CN CN202010083678.6A patent/CN111340695A/en active Pending
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104067372A (en) * | 2012-01-27 | 2014-09-24 | 塞莫费雪科学（不来梅）有限公司 | Multi-reflection mass spectrometer |
| CN102663734A (en) * | 2012-03-15 | 2012-09-12 | 天津理工大学 | Fish eye lens calibration and fish eye image distortion correction method |
| CN103533235A (en) * | 2013-09-17 | 2014-01-22 | 北京航空航天大学 | Quick digital panoramic device based on linear array charge coupled device (CCD) for great case/event scene |
| CN106469448A (en) * | 2015-06-26 | 2017-03-01 | 康耐视公司 | Carry out automatic industrial inspection using 3D vision |
| CN106842520A (en) * | 2017-03-30 | 2017-06-13 | 中山联合光电科技股份有限公司 | A kind of high definition panorama looks around optical imaging system |
| CN107240066A (en) * | 2017-04-28 | 2017-10-10 | 天津大学 | Image super-resolution rebuilding algorithm based on shallow-layer and deep layer convolutional neural networks |
| CN108961151A (en) * | 2018-05-08 | 2018-12-07 | 中德（珠海）人工智能研究院有限公司 | A method of the three-dimensional large scene that ball curtain camera obtains is changed into sectional view |
| CN108983394A (en) * | 2018-08-23 | 2018-12-11 | 上海帛视光电科技有限公司 | A kind of fish eye lens system, image acquisition device |
Non-Patent Citations (1)
| Title |
|---|
| JIAHUI YU.ETC: ""Wide Activation for Efficient and Accurate Image Super-Resolution"", <ARXIV>, 21 December 2018 (2018-12-21), pages 3 - 4 * |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111832463A (en) * | 2020-07-07 | 2020-10-27 | 哈尔滨理工大学 | Deep learning-based traffic sign detection method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US11037278B2 (en) | Systems and methods for transforming raw sensor data captured in low-light conditions to well-exposed images using neural network architectures | |
| US10257501B2 (en) | Efficient canvas view generation from intermediate views | |
| US10708525B2 (en) | Systems and methods for processing low light images | |
| US11880939B2 (en) | Embedding complex 3D objects into an augmented reality scene using image segmentation | |
| US20210124985A1 (en) | System and method for deep machine learning for computer vision applications | |
| Chen et al. | Deep exposure fusion with deghosting via homography estimation and attention learning | |
| US11871110B2 (en) | Single image ultra-wide fisheye camera calibration via deep learning | |
| CN111553841B (en) | Real-time video splicing method based on optimal suture line updating | |
| KR20230146649A (en) | Color and infrared 3D reconstruction using implicit radiance functions. | |
| EP3886044A1 (en) | Robust surface registration based on parameterized perspective of image templates | |
| CN115375536A (en) | Image processing method and apparatus | |
| Chang et al. | Deep learning based image Super-resolution for nonlinear lens distortions | |
| CN115115516A (en) | Real-world video super-resolution algorithm based on Raw domain | |
| CN111340695A (en) | Super-resolution reconstruction method of dome screen video | |
| JP2006221599A (en) | Method and apparatus for generating mapping function, and compound picture develop method, and its device | |
| Bergmann et al. | Gravity alignment for single panorama depth inference | |
| Hong et al. | PAR $^{2} $ Net: End-to-end Panoramic Image Reflection Removal | |
| Liang et al. | Multi-scale and multi-patch transformer for sandstorm image enhancement | |
| Yue et al. | Rvideformer: Efficient raw video denoising transformer with a larger benchmark dataset | |
| Okura et al. | Aerial full spherical HDR imaging and display | |
| CN114257733A (en) | Method and system for image processing of omni-directional image with viewpoint offset | |
| CN112991174A (en) | Method and system for improving resolution of single-frame infrared image | |
| CN111709880B (en) | Multi-path picture splicing method based on end-to-end neural network | |
| US11651475B2 (en) | Image restoration method and device | |
| Jung et al. | Deep low-contrast image enhancement using structure tensor representation |
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 |