CN113538247B - Super-resolution generation and conditional countermeasure network remote sensing image sample generation method - Google Patents
Super-resolution generation and conditional countermeasure network remote sensing image sample generation method Download PDFInfo
- Publication number
- CN113538247B CN113538247B CN202110923714.XA CN202110923714A CN113538247B CN 113538247 B CN113538247 B CN 113538247B CN 202110923714 A CN202110923714 A CN 202110923714A CN 113538247 B CN113538247 B CN 113538247B
- Authority
- CN
- China
- Prior art keywords
- resolution
- image
- remote sensing
- super
- sample
- 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
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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; 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
- G06T2207/10032—Satellite or aerial image; Remote sensing
-
- 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
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30168—Image quality inspection
Abstract
The invention discloses a super-resolution generation and conditional countermeasure network remote sensing image sample generation method which is characterized by comprising the following steps: step 1, preparing high and medium resolution image data, carrying out sample labeling on target area data, and training a super-resolution optimization model after the high and medium resolution image data are matched; and 2, performing reverse learning based on the conditional countermeasure network based on the high-resolution image data and the medium-resolution image data in the step 1 to generate a remote sensing image sample, performing quality detection on the generated remote sensing image sample, and screening out a final result. Compared with the problems of insufficient high quality, less labeling information, unstable quality of generated samples and the like of the traditional sample generation structure, the super-resolution generation structure effectively solves the problem of insufficient high-resolution data, intelligently generates and screens samples, and realizes the method for automatically generating high-resolution images by giving the samples by a computer.
Description
Technical Field
The invention relates to the field of remote sensing images, in particular to a super-resolution generation and conditional countermeasure network remote sensing image sample generation method.
Background
In recent years, deep learning is rapidly developed, deep learning methods are rapidly spread in various fields, various classification models are seen out at a glance, but considerable problems still exist, the generalization capability of the models is a relatively remarkable problem in deep learning application, and each change of a research area is a challenge to the sample set of the invention mainly because the remote sensing images and labels used for training have defects in quantity and geographic distribution. The difficulty of labeling the data set required by deep learning is almost a consensus, and in comparison, the labeling of the remote sensing image is more difficult, and the labeling is limited by a time phase problem due to the large application area of the remote sensing image. The remote sensing sample marking cost is very high, and the remote sensing sample marking cost is difficult to meet the requirements of practical application, so that a technology which does not rely on manual marking is very important.
Disclosure of Invention
The invention provides an intelligent sample generation scheme by combining a deep learning method, and a sample with higher quality is generated on the premise of using a small amount of samples as a guide.
The technical scheme of the invention is as follows: a method for generating a high-resolution image with a sample label by combining a super-resolution generation network and a conditional countermeasure network is used for expanding a remote sensing image sample label library. In order to overcome the bottleneck problems that the high-resolution remote sensing image is long in acquisition period and high in cost and restricts the generation of remote sensing image samples, a large number of high-resolution remote sensing images are obtained from low-resolution remote sensing images based on relatively few sample training of a super-resolution generation network.
The invention discloses a super-resolution generation and conditional countermeasure network remote sensing image sample generation method, which comprises the following steps of;
step 1, preparing high and medium resolution image data, carrying out sample labeling on target area data, and training a super-resolution optimization model after the high and medium resolution image data are matched;
and 2, performing reverse learning based on the conditional countermeasure network based on the high-resolution image data and the medium-resolution image data in the step 1 to generate a remote sensing image sample, performing quality detection on the generated remote sensing image sample, and screening out a final result.
Further, step 1, preparing high and medium resolution image data, performing sample labeling on the target area data, and after the high and medium resolution image data are matched, training a super-resolution optimization model, specifically including:
step 1.1, selecting corresponding medium-resolution and high-resolution image data according to a target, performing geographic matching on the two types of image data, and performing a small amount of sample marking on the image data of a target area;
step 1.2, after the medium-resolution and high-resolution image data are matched, segmenting the image data to prepare for constructing a super-resolution network;
and step 1.3, training to obtain a super-resolution optimization model by taking the medium-resolution image data as input and the high-resolution image data as output.
Further, the super-resolution optimization model in the step 1 specifically includes the following steps:
the input is a geographically registered low resolution image of the same spatial range;
the output is a high resolution image;
the loss function is the difference between the generated high-resolution image and the actual high-resolution image; the method comprises the steps of inputting a medium-resolution image, obtaining continuous low-level features through a low-resolution feature extraction network, respectively obtaining the features of each level from a multi-level structure in the extraction process, combining the features to form multi-scale network features, inputting the combined features into a high-resolution reconstruction network, and finally obtaining a required high-resolution output result through transposition convolution and upsampling in the high-resolution reconstruction network.
Further, the step 2 comprises:
inputting the label of the invention into a conditional antagonistic network, in the process, fitting a distribution function of a target by researching the size and the distribution of the target, taking the background and the foreground of label division as input, generating a high-resolution image through antagonistic training, comparing the error of the result with the high-resolution image generated by the super-resolution optimization model in the step 1, finally performing quality control by using a quality discrimination network, and outputting to obtain a final result.
Further, step 2, based on the data in step 1, performing reverse learning based on a conditional countermeasure network to generate a remote sensing image sample, and performing quality detection, wherein the sample generation mainly includes two parts, that is, the conditional countermeasure network generates the remote sensing image sample and the image quality detection, specifically, as follows:
step (2.1) the conditional countermeasure network generates a remote sensing image sample: obtaining mapping from a labeling vector to a remote sensing image based on a label obtained from a data preparation link and a high-resolution image generated by super-resolution optimization, applying the mapping, realizing reverse learning based on antagonistic learning neural network training, and generating a required remote sensing image sample;
step (2.2) using a conditional countermeasure network to detect the image quality, wherein the conditional countermeasure network specifically comprises the following steps:
the method comprises the following steps of inputting a position for providing a target, a position of a target contour, an angle of the target and a Gaussian random matrix, wherein the specific generation mode is as follows: setting a mean value and a variance, wherein the setting of the mean value and the variance requires iterative computation of optimal random function generation;
outputting a high-resolution remote sensing target map;
the loss function is: loss of generation of real images and generated images;
the final purpose of generating the labeling sample is achieved by compensating the difference between the generated image and the real image through the confrontation generation network.
Further, the step 2 performs quality detection on the generated sample, and includes:
the first round of automatic screening is completed in the stage of generating images, and results with generation scores larger than a threshold value are selected;
the second round of the method adopts PSNR indexes to compare the image with a real image, and the quality of the image is compared through a signal-to-noise ratio;
and the third round of manual screening is to remove obvious and unreasonable synthetic results through manual visual inspection on the automatic screening results of the first two rounds.
Furthermore, an image quality inspection picture classification network is set for the third manual screening inspection to perform quality screening on the generated image, and the structure of the network comprises: feature extraction is carried out by convolution-pooling structures which are continuously carried out three times, and then results are output through a 2-layer fully-connected classification structure.
Has the advantages that:
according to the invention, several deep learning methods are effectively combined, compared with the problems of insufficient high quality, less labeled information, unstable generated sample quality and the like of the traditional sample generation structure, the problem of insufficient high-resolution data is effectively solved through the super-resolution generation structure, and the process from artificial design labeling to actual images is realized through the high-resolution generation structure, so that the generated images of the invention implicitly contain the position and contour information of the target, and the quality of the generated samples can be more effectively controlled through the quality detection network.
Drawings
FIG. 1: a method flow diagram of the present invention;
FIG. 2: high and low resolution picture contrast;
FIG. 3: a super-resolution network structure;
FIG. 4: a countermeasure generation network;
FIG. 5: a conditional countermeasure network schematic;
FIG. 6: generating an initial condition and a result schematic diagram under the condition;
FIG. 7: a schematic diagram of a quality screening network.
Detailed Description
The technical solutions in the embodiments of the present invention will be described clearly and completely with reference to the accompanying drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, rather than all embodiments, and all other embodiments obtained by a person skilled in the art based on the embodiments of the present invention belong to the protection scope of the present invention without creative efforts.
As shown in fig. 1, a flow chart of a super-resolution generation and conditional countermeasure network remote sensing image sample generation method of the present invention includes the following steps;
step 1, preparing high and medium resolution data, carrying out sample labeling on target area data, and training a super-resolution optimization model after the high and medium resolution data are matched;
and 2, performing reverse learning based on the data in the step 1 and a conditional countermeasure network to generate a remote sensing image sample, and performing quality detection.
According to the embodiment of the present invention, step 1, preparing high and medium resolution data, performing sample labeling on target region data, and using the high and medium resolution data after matching for training a super-resolution optimization model specifically includes:
in the link, the invention firstly selects corresponding medium-resolution and high-resolution data according to a target, carries out geographic matching on two types of images, and carries out a small amount of sample labeling on target area data.
And then, inputting the label and the target image of the invention into a conditional countermeasure network, in the process, fitting a distribution function of the target by researching the size and the distribution of the target, taking the background and the foreground of the label division as input, generating a high-resolution image map of the invention through countermeasure training, comparing the result with the high-resolution image of the invention per se to obtain an error, finally, performing quality control by using a quality judgment network, and outputting to obtain a final result of the invention.
According to an embodiment of the present invention, the above-mentioned acquisition of data is not limited, and for the purpose of illustrating the applicability of the technology, two types of sourced data are taken as an example, and sentinel data and google data are mainly used. If higher quality data can be obtained in practical application, the method can also be taken as an implementation target of the technology.
The medium-resolution data is mainly a sentinel No. 2 satellite, and the sentinel-2 satellite is a high-resolution multispectral imaging satellite and is mainly used for global land observation including land vegetation, soil, water resources, inland waterway and coastal areas. The satellite has high resolution and high revisitation rate, the main payload of the sentinel-2 satellite is a multi-spectral imager (MSI), the working spectral bands are visible light, near infrared and short wave infrared, the ground resolution is 10m, 20m and 60m respectively, the breadth of a multi-spectral image is 290km, global land surface imaging data is updated once every 10 days, the average observation time of each orbit period is 16.3min, and the peak value is 31 min. High resolution google data is available from open sources, but the google earth collects data from various satellite and aerial photographic sources, which can take months to process, compare, and set up the data before it appears on the map. Pictures of Google Earth are from the "agile bird" satellite (QuickBird) and the "world window" satellite No. 1 (WorldView) launched in september of the last year (indeed, satellite shots) which provide Google Earth with clear images with image resolutions up to 0.61. However, the google data are different in position quality, and the values can only be used as reference.
The invention continuously generates a large amount of high-Resolution data through the Super-Resolution network, relatively speaking, the model SRCNN based on the CNN (convolutional neural network) introduces the CNN into SISR (Single-Image-Super-Resolution), and the model SRCNN only uses three layers of networks to obtain advanced results. In the SISR domain, the following two major directions are roughly classified. One is an algorithm based on evaluation criteria such as PSNR, SSIM, etc., in which the SRCNN model is used as a representative. Another algorithm is a series of algorithms represented by SRGAN, which aims to reduce the perceptual loss, pays attention to details, pays attention to the overall view, and is a method for reducing the perceptual loss. The two algorithms in different directions have different application fields. The invention focuses on different choices of targets, and in general, detail recovery is more important to the invention.
According to an embodiment of the present invention, the super-resolution optimization model, as shown in fig. 3, is as follows:
the input is as follows: a geographically registered low resolution image (RGB grid) of the same spatial extent;
the output is: high resolution imagery (RGB grid);
the loss function is: the difference between the generated high resolution video and the actual high resolution video.
The method comprises the steps of inputting a medium-resolution image, obtaining continuous low-level features through a low-resolution feature extraction network, respectively obtaining each level of features from a multi-level structure in the extraction process, combining the features to form multi-scale network features, inputting the combined features into a high-resolution reconstruction network, and finally obtaining a high-resolution output result required by the invention through sampling the result of the invention on a transposition convolution in the high-resolution reconstruction network.
The high-resolution image is low in time resolution, and the data volume in the acquired characteristic region may not be enough, so that the low-resolution image is introduced to be supplemented by using a super-resolution model.
According to the step 2 of the invention, based on the data in the step 1, reverse learning is carried out based on a conditional countermeasure network, a remote sensing image sample is generated, and quality detection is carried out, wherein the sample generation mainly comprises two parts, namely the generation of the remote sensing image sample and the generation of image quality detection by the conditional countermeasure network, and the method specifically comprises the following steps:
step (2.1) the conditional countermeasure network generates a remote sensing image sample: based on labels acquired from a data preparation link and a high-resolution image generated by super-resolution optimization, the method obtains a large amount of mappings from label vectors to remote sensing images, applies the mappings, realizes reverse learning based on antagonistic learning neural network training, generates remote sensing image samples required by the method, and aims to overcome the problem of insufficient antagonistic network robustness caused by randomness of input images.
Step (2.2) image quality detection: the method comprises the following steps that (1) primary remote sensing image samples are generated, but the quality is inevitably uneven in the generation process, and the quality of the samples is further detected.
In recent years, a counterproductive network has come to work in various fields, and as shown in fig. 4, GAN has two parts, a producer and a discriminator: the basic concept of the generator is simple, and a vector is Input, and a high-dimensional vector (which may be a picture or a word) is output through an NN. Meanwhile, the GAN has a part called "discriminator", whose Input is what you want to generate (which is actually the output generated by the generator), such as a picture, or a piece of speech.its output is a scalar which represents how good the Input is, and the larger the number is, the truer the Input is;
the main process can be described as follows: the generator generates a thing, inputs the thing into the arbiter, then judges whether the input is real data or machine-generated by the arbiter, if the arbiter has not been cheated, then the generator continues to evolve, outputs the second generation Output, inputs the arbiter again, the arbiter is also evolving at the same time, has stricter requirement to the generator Output. Thus, generators and discriminators have evolved with relationships that are somewhat like a competitive relationship, and therefore have a source of the name "generating an adversarial network (adversarial)".
The conditional countermeasure network is to add a certain condition to the original GAN so that the prediction result of GAN is narrowed to a smaller range, and focus more on the smaller range, according to an embodiment of the present invention, the conditional countermeasure network is specifically as shown in fig. 5 below:
the input is as follows: the method comprises the steps of providing targets, such as the positions (4) of buildings, the positions (a plurality of points) of the outlines of the targets, the angles (1) of the targets, and 1 Gaussian random matrix (used for adjusting the generation effect of generating the target image, wherein the generation method is to set the mean value and the variance, and the setting of the mean value and the variance requires iterative calculation of the optimal random function generation).
The output is: and (5) high-resolution remote sensing target map.
The loss function is: real images and loss of production of produced images.
The method and the system make up the difference between the generated image and the real image through the confrontation generation network so as to achieve the final purpose of generating the labeled sample. Fig. 6 is a schematic diagram of the initial conditions and results of condition generation.
According to the embodiment of the invention, in the step 2, the quality detection is carried out on the generated sample;
aiming at various problems of generating the synthetic picture, the invention ensures the quality of the final reserved picture by two-round automatic screening and one-by-one manual screening.
The first round of automatic screening is completed in the stage of generating images, and the invention selects the result with the generation score larger than 0.85.
The second round of the invention adopts PSNR index, compares the generated image with the original image, and compares the quality of the image by signal-to-noise ratio. In order to control the overall quality of the image, the quality of the generated image is evaluated through a similarity index PSNR. PSNR (peak Signal to Noise ratio), the peak Signal-to-Noise ratio, i.e. the ratio of the energy of the peak Signal to the average energy of the Noise, is usually expressed by taking log as decibel (dB), since MSE is the energy average of the difference between the real image and the image containing Noise, and the difference between the two is the Noise, PSNR is the ratio of the peak Signal energy to MSE. The definition formula is as follows:
wherein:
m and n represent row and column numbers, I and K represent an original image and a generated image respectively, I and j represent pixel coordinates, MaxValue represents the maximum value of an image, and bits represents the digit of the image;
and the third round of manual screening is to remove obvious and unreasonable synthetic results through manual visual inspection on the automatic screening results of the first two rounds.
An image quality inspection picture classification network is arranged for the third manual screening inspection to perform quality screening on the image generated by the invention, and the structure of the image is as shown in the following figure 7, feature extraction is performed by a convolution-pooling structure which is performed three times in succession, and then the result is output through a full-connection classification structure with 2 layers. The advantage of this design is that after the above screening, the result of the manual inspection will become the sample of the present invention to train the network structure, and as the training of the present invention proceeds, the model will become more and more intelligent, thereby reducing the work pressure for the manual screening.
Although illustrative embodiments of the present invention have been described above to facilitate the understanding of the present invention by those skilled in the art, it should be understood that the present invention is not limited to the scope of the embodiments, but various changes may be apparent to those skilled in the art, and it is intended that all inventive concepts utilizing the inventive concepts set forth herein be protected without departing from the spirit and scope of the present invention as defined and limited by the appended claims.
Claims (6)
1. A super-resolution generation and conditional countermeasure network remote sensing image sample generation method is characterized by comprising the following steps;
step 1, preparing high and medium resolution image data, carrying out sample labeling on target area data, and training a super-resolution optimization model after the high and medium resolution image data are matched;
step 2, reverse learning is carried out on the basis of the high-resolution image data and the medium-resolution image data in the step 1 and on the basis of a conditional countermeasure network, a remote sensing image sample is generated, quality detection is carried out on the generated remote sensing image sample, and a final result is screened out;
and 2, performing reverse learning based on the conditional countermeasure network based on the data in the step 1 to generate a remote sensing image sample, and performing quality detection, wherein the sample generation mainly comprises two parts, namely the conditional countermeasure network generation remote sensing image sample and the image quality detection, and specifically comprises the following steps:
step (2.1) the conditional countermeasure network generates a remote sensing image sample: obtaining mapping from a labeling vector to a remote sensing image based on a label obtained from a data preparation link and a high-resolution image generated by super-resolution optimization, applying the mapping, realizing reverse learning based on antagonistic learning neural network training, and generating a required remote sensing image sample;
step (2.2) using a conditional countermeasure network to detect the image quality, wherein the conditional countermeasure network specifically comprises the following steps:
the method comprises the following steps of inputting a position for providing a target, a position of a target contour, an angle of the target and a Gaussian random matrix, wherein the specific generation mode is as follows: setting a mean value and a variance, wherein the setting of the mean value and the variance requires iterative computation of optimal random function generation;
outputting a high-resolution remote sensing target map;
the loss function is: loss of generation of real images and generated images;
the final purpose of generating the labeling sample is achieved by compensating the difference between the generated image and the real image through the confrontation generation network.
2. The method for generating the super-resolution and conditional countermeasure network remote sensing image sample according to claim 1, wherein the step 1 of preparing high and medium resolution image data, labeling the target area data with the sample, and matching the high and medium resolution image data for training the super-resolution optimization model specifically comprises:
step 1.1, selecting corresponding medium-resolution and high-resolution image data according to a target, performing geographic matching on the two types of image data, and performing a small amount of sample marking on the image data of a target area;
step 1.2, after the medium-resolution and high-resolution image data are matched, segmenting the image data to prepare for constructing a super-resolution network;
and step 1.3, training to obtain a super-resolution optimization model by taking the medium-resolution image data as input and the high-resolution image data as output.
3. The method for generating the super-resolution generation and conditional countermeasure network remote sensing image sample according to claim 1, wherein the super-resolution optimization model in the step 1 is specifically as follows:
the input is a geographically registered low resolution image of the same spatial range;
the output is a high resolution image;
the loss function is the difference between the generated high-resolution image and the actual high-resolution image; the method comprises the steps of inputting a medium-resolution image, obtaining continuous low-level features through a low-resolution feature extraction network, respectively obtaining the features of each level from a multi-level structure in the extraction process, combining the features to form multi-scale network features, inputting the combined features into a high-resolution reconstruction network, and finally obtaining a required high-resolution output result through transposition convolution and upsampling in the high-resolution reconstruction network.
4. The super-resolution generation and conditional countermeasure network remote sensing image sample generation method according to claim 1, wherein the step 2 includes:
inputting the label into a conditional countermeasure network, in the process, fitting a distribution function of the target by researching the size and the distribution of the target, taking the background and the foreground divided by the label as input, generating a high-resolution image through countermeasure training, comparing the result with the high-resolution image generated by the super-resolution optimization model in the step 1 to obtain an error, finally performing quality control by using a quality judgment network, and outputting to obtain a final result.
5. The super-resolution generation and conditional countermeasure network remote sensing image sample generation method according to claim 1, wherein the step 2 of performing quality inspection on the generated sample comprises:
the first round of automatic screening is completed in the stage of generating images, and results with generation scores larger than a threshold value are selected;
comparing the image with a real image by adopting a PSNR index in the second round, and comparing the quality of the image by a signal-to-noise ratio;
and the third round of manual screening is to remove obvious and unreasonable synthetic results through manual visual inspection on the automatic screening results of the first two rounds.
6. The method for generating super-resolution and conditional countermeasure network remote sensing image samples as claimed in claim 5, wherein an image quality inspection picture classification network is provided for a third manual screening inspection to perform quality screening of the generated images, and the structure thereof comprises: feature extraction is carried out by convolution-pooling structures which are continuously carried out three times, and then results are output through a 2-layer fully-connected classification structure.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110923714.XA CN113538247B (en) | 2021-08-12 | 2021-08-12 | Super-resolution generation and conditional countermeasure network remote sensing image sample generation method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110923714.XA CN113538247B (en) | 2021-08-12 | 2021-08-12 | Super-resolution generation and conditional countermeasure network remote sensing image sample generation method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113538247A CN113538247A (en) | 2021-10-22 |
| CN113538247B true CN113538247B (en) | 2022-04-15 |
Family
ID=78090949
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110923714.XA Active CN113538247B (en) | 2021-08-12 | 2021-08-12 | Super-resolution generation and conditional countermeasure network remote sensing image sample generation method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113538247B (en) |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107729922A (en) * | 2017-09-20 | 2018-02-23 | 千寻位置网络有限公司 | Remote sensing images method for extracting roads based on deep learning super-resolution technique |
| CN110136063A (en) * | 2019-05-13 | 2019-08-16 | 南京信息工程大学 | A kind of single image super resolution ratio reconstruction method generating confrontation network based on condition |
| CN111899172A (en) * | 2020-07-16 | 2020-11-06 | 武汉大学 | Vehicle target detection method oriented to remote sensing application scene |
| CN112001847A (en) * | 2020-08-28 | 2020-11-27 | 徐州工程学院 | Method for generating high-quality image by relatively generating antagonistic super-resolution reconstruction model |
Family Cites Families (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2018053340A1 (en) * | 2016-09-15 | 2018-03-22 | Twitter, Inc. | Super resolution using a generative adversarial network |
| CN109978762B (en) * | 2019-02-27 | 2023-06-16 | 南京信息工程大学 | Super-resolution reconstruction method based on condition generation countermeasure network |
| CN110490802B (en) * | 2019-08-06 | 2021-01-19 | 北京观微科技有限公司 | Super-resolution-based satellite image airplane target model identification method |
| CN110599401A (en) * | 2019-08-19 | 2019-12-20 | 中国科学院电子学研究所 | Remote sensing image super-resolution reconstruction method, processing device and readable storage medium |
| CN110728658A (en) * | 2019-09-16 | 2020-01-24 | 武汉大学 | High-resolution remote sensing image weak target detection method based on deep learning |
| CN110992262B (en) * | 2019-11-26 | 2023-04-07 | 南阳理工学院 | Remote sensing image super-resolution reconstruction method based on generation countermeasure network |
| CN111242839B (en) * | 2020-01-13 | 2023-04-21 | 华南理工大学 | Image scaling and clipping method based on scale level |
| CN111428792B (en) * | 2020-03-26 | 2023-04-18 | 中国科学院空天信息创新研究院 | Remote sensing information image sample marking method and device |
-
2021
- 2021-08-12 CN CN202110923714.XA patent/CN113538247B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107729922A (en) * | 2017-09-20 | 2018-02-23 | 千寻位置网络有限公司 | Remote sensing images method for extracting roads based on deep learning super-resolution technique |
| CN110136063A (en) * | 2019-05-13 | 2019-08-16 | 南京信息工程大学 | A kind of single image super resolution ratio reconstruction method generating confrontation network based on condition |
| CN111899172A (en) * | 2020-07-16 | 2020-11-06 | 武汉大学 | Vehicle target detection method oriented to remote sensing application scene |
| CN112001847A (en) * | 2020-08-28 | 2020-11-27 | 徐州工程学院 | Method for generating high-quality image by relatively generating antagonistic super-resolution reconstruction model |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113538247A (en) | 2021-10-22 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Song et al. | Spatiotemporal satellite image fusion using deep convolutional neural networks | |
| Yang et al. | Building extraction at scale using convolutional neural network: Mapping of the united states | |
| Zhang et al. | Remote sensing image spatiotemporal fusion using a generative adversarial network | |
| Yang et al. | CDnet: CNN-based cloud detection for remote sensing imagery | |
| CN108573276B (en) | Change detection method based on high-resolution remote sensing image | |
| Pan | Cloud removal for remote sensing imagery via spatial attention generative adversarial network | |
| CN111652193A (en) | Wetland classification method based on multi-source images | |
| CN112949414B (en) | Intelligent surface water body drawing method for wide-vision-field high-resolution six-satellite image | |
| Liu et al. | Statistical convolutional neural network for land-cover classification from SAR images | |
| CN111640116B (en) | Aerial photography graph building segmentation method and device based on deep convolutional residual error network | |
| Meng et al. | Multilayer feature fusion network with spatial attention and gated mechanism for remote sensing scene classification | |
| CN116052016A (en) | Fine segmentation detection method for remote sensing image cloud and cloud shadow based on deep learning | |
| CN114494821A (en) | Remote sensing image cloud detection method based on feature multi-scale perception and self-adaptive aggregation | |
| CN115861756A (en) | Earth background small target identification method based on cascade combination network | |
| Mu et al. | A climate downscaling deep learning model considering the multiscale spatial correlations and chaos of meteorological events | |
| Li et al. | A pseudo-siamese deep convolutional neural network for spatiotemporal satellite image fusion | |
| Gu et al. | A classification method for polsar images using SLIC superpixel segmentation and deep convolution neural network | |
| Zhang et al. | Dense haze removal based on dynamic collaborative inference learning for remote sensing images | |
| Song et al. | Remote sensing image spatiotemporal fusion via a generative adversarial network with one prior image pair | |
| Wang | Remote sensing image semantic segmentation algorithm based on improved ENet network | |
| CN113538247B (en) | Super-resolution generation and conditional countermeasure network remote sensing image sample generation method | |
| Song et al. | HDTFF-Net: Hierarchical deep texture features fusion network for high-resolution remote sensing scene classification | |
| CN111191704A (en) | Foundation cloud classification method based on task graph convolutional network | |
| CN116563103A (en) | Remote sensing image space-time fusion method based on self-adaptive neural network | |
| CN115909077A (en) | Hyperspectral image change detection method based on unsupervised spectrum unmixing neural network |
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 |