CN114882252A - Semi-supervised remote sensing image change detection method and device and computer equipment - Google Patents
Semi-supervised remote sensing image change detection method and device and computer equipment Download PDFInfo
- Publication number
- CN114882252A CN114882252A CN202210550967.1A CN202210550967A CN114882252A CN 114882252 A CN114882252 A CN 114882252A CN 202210550967 A CN202210550967 A CN 202210550967A CN 114882252 A CN114882252 A CN 114882252A
- Authority
- CN
- China
- Prior art keywords
- remote sensing
- change detection
- sensing image
- detection result
- pair
- 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
- 230000008859 change Effects 0.000 title claims abstract description 292
- 238000001514 detection method Methods 0.000 title claims abstract description 261
- 238000000034 method Methods 0.000 claims abstract description 41
- 238000012549 training Methods 0.000 claims abstract description 18
- 238000012545 processing Methods 0.000 claims description 55
- 230000006870 function Effects 0.000 claims description 37
- 238000013519 translation Methods 0.000 claims description 24
- 238000004364 calculation method Methods 0.000 claims description 14
- 238000001914 filtration Methods 0.000 claims description 13
- 230000004927 fusion Effects 0.000 claims description 10
- 238000004590 computer program Methods 0.000 claims description 9
- 238000012544 monitoring process Methods 0.000 claims description 8
- 230000002708 enhancing effect Effects 0.000 claims description 3
- 239000000284 extract Substances 0.000 claims 1
- 238000010586 diagram Methods 0.000 description 6
- 230000008569 process Effects 0.000 description 6
- 238000010008 shearing Methods 0.000 description 6
- 238000004422 calculation algorithm Methods 0.000 description 5
- 238000002372 labelling Methods 0.000 description 4
- 238000013527 convolutional neural network Methods 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 238000004519 manufacturing process Methods 0.000 description 3
- 230000007246 mechanism Effects 0.000 description 3
- 238000005457 optimization Methods 0.000 description 3
- 238000000926 separation method Methods 0.000 description 3
- 238000003860 storage Methods 0.000 description 3
- 230000000007 visual effect Effects 0.000 description 3
- 238000013135 deep learning Methods 0.000 description 2
- 230000007306 turnover Effects 0.000 description 2
- 238000013528 artificial neural network Methods 0.000 description 1
- 238000005520 cutting process Methods 0.000 description 1
- 230000002950 deficient Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000003707 image sharpening Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 239000004973 liquid crystal related substance Substances 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000000717 retained effect Effects 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/74—Image or video pattern matching; Proximity measures in feature spaces
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/80—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level
- G06V10/806—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level of extracted features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Multimedia (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Medical Informatics (AREA)
- Software Systems (AREA)
- Databases & Information Systems (AREA)
- Computing Systems (AREA)
- Artificial Intelligence (AREA)
- Health & Medical Sciences (AREA)
- Image Analysis (AREA)
Abstract
The application relates to a semi-supervised remote sensing image change detection method and device and computer equipment. The method comprises the following steps: inputting a labeled remote sensing image pair, a real label and a non-labeled remote sensing image pair; calculating according to the remote sensing image pair with the label and the real label to obtain supervision loss; calculating according to the label-free remote sensing image pair to obtain unsupervised loss; and calculating according to the supervision loss and the unsupervised loss to obtain the total loss, taking the minimum total loss as a target function, and training a change detection model by using an optimizer to obtain a semi-supervised remote sensing image change detection model. By adopting the method, the change detection precision of the remote sensing image can be effectively improved by utilizing a small number of labeled remote sensing image pairs and a large number of non-labeled remote sensing image pairs, the problem of insufficient label data is solved, and the potential of remote sensing big data is released.
Description
Technical Field
The application relates to the technical field of remote sensing image processing, in particular to a semi-supervised remote sensing image change detection method and device and computer equipment.
Background
The task of remote sensing image Change Detection (CD) is to identify changes occurring in remote sensing images of the same area taken at different times. The variations herein generally refer to semantic variations. For many years, remote sensing image change detection is one of the research hotspots in the field of remote sensing. With the increasing number of remote sensing images, the improvement of shooting precision and the development of deep learning technology, the change information of the concerned area, including the change of natural ground objects and the change of artificial buildings, can be rapidly acquired by using the remote sensing image change detection technology, and powerful support is provided for the decision of governments, companies and organizations. So far, remote sensing image change detection technology has been widely applied in the fields of ecosystem monitoring, land resource and land utilization mapping, damage assessment, city expansion monitoring and the like. Currently, the mainstream CD algorithm is based on a fully supervised deep learning method, and is mostly based on a Convolutional Neural Network (CNN). Among them, UNet based on full convolution network is the most popular and becomes one of the standard CNN architectures for CD tasks, and the satisfactory effect is achieved.
However, the use of the remote sensing image change algorithm based on the fully supervised learning requires that the neural network training is performed on the double-temporal remote sensing image pair with a large number of labels, and the labeling of the labels needs to consume a large amount of labor and time cost. If the model is trained by using an unsupervised learning method, the change detection precision of the model is often low due to the lack of guidance of label data, and the change detection precision is not enough to support the practical change detection application.
Disclosure of Invention
Therefore, in order to solve the technical problems, a semi-supervised remote sensing image change detection method, a semi-supervised remote sensing image change detection device and computer equipment are provided, wherein the problems that a remote sensing image change detection label is deficient and a large amount of labor and time cost are consumed for labeling the label.
A semi-supervised remote sensing image change detection method comprises the following steps:
inputting a labeled remote sensing image pair, a real label and a non-labeled remote sensing image pair, wherein the real label is a size-matched change gray level image of the labeled remote sensing image pair;
carrying out weak enhancement processing on the labeled remote sensing image pair to obtain a weak enhanced labeled remote sensing image pair, inputting the weak enhanced labeled remote sensing image pair into a change detection model to obtain a first change detection result, and calculating according to the first change detection result and a real label to obtain supervision loss;
carrying out weak enhancement processing on the non-tag remote sensing image pair to obtain a weak enhancement non-tag remote sensing image pair, carrying out strong enhancement processing on the weak enhancement non-tag remote sensing image pair to obtain a distorted image, and inputting the distorted image into a change detection model to obtain a second change detection result;
inputting the weakly-enhanced label-free remote sensing image pair into a change detection model to obtain a third change detection result, carrying out strong enhancement processing on the third change detection result, filtering by using a confidence threshold filter to obtain a pseudo label, and calculating according to the second change detection result and the pseudo label to obtain unsupervised loss;
and calculating according to the supervision loss and the unsupervised loss to obtain the total loss, taking the minimum total loss as a target function, and training a change detection model by using an optimizer to obtain a semi-supervised remote sensing image change detection model.
In one embodiment, inputting the labeled remote sensing image pair, the real label and the unlabeled remote sensing image pair further comprises:
inputting a labeled remote sensing image pair, wherein the same pair of labeled remote sensing images consists of two labeled remote sensing images in the same area and at different time phases, and the different pair of labeled remote sensing images consists of two labeled remote sensing images in different areas and at different time phases;
inputting a real label, wherein the real label is a change gray image with a label remote sensing image pair matched with the size;
inputting a pair of non-tag remote sensing images, wherein the same pair of non-tag remote sensing images consists of two non-tag remote sensing images in the same area and at different time phases, and the different pair of non-tag remote sensing images consists of two non-tag remote sensing images in different areas and at different time phases.
In one embodiment, the weak enhancement processing of the labeled remote sensing image pair to obtain a weak enhanced labeled remote sensing image pair further includes:
carrying out translation and/or overturning with preset amplitude on the labeled remote sensing image pair to obtain a weakly enhanced labeled remote sensing image pair, wherein the same pair of labeled remote sensing images are subjected to translation and/or overturning with the same preset amplitude to obtain the same pair of weakly enhanced labeled remote sensing images, and different pairs of labeled remote sensing images are subjected to translation and/or overturning with random preset amplitude to obtain different pairs of weakly enhanced labeled remote sensing images;
carrying out weak enhancement processing on the unlabeled remote sensing image pair to obtain a weak enhancement unlabeled remote sensing image pair, which comprises the following steps:
and carrying out translation and/or overturning at the preset amplitude on the non-tag remote sensing image pair to obtain a weak enhancement non-tag remote sensing image pair, wherein the same pair of non-tag remote sensing images are carried out translation and/or overturning at the same preset amplitude to obtain the same pair of weak enhancement non-tag remote sensing images, and different pairs of non-tag remote sensing images are carried out translation and/or overturning at random preset amplitudes to obtain different pairs of weak enhancement non-tag remote sensing images.
In one embodiment, the change detection model comprises a siamese encoder and a decoder, wherein the encoder further comprises a graph attention module:
the Siamese encoder carries out downsampling processing on an image through shared weight and parameters, image features of the image are extracted, the image features are used as nodes, image fusion feature values of the nodes are obtained through image features corresponding to neighbor nodes of a graph attention module fusion node, the image fusion feature values are input into an attention unit, attention features fused into an attention mechanism are obtained, a decoder obtains a change detection result through extracting the attention features, the image comprises a weak enhanced labeled remote sensing image pair, a distorted image and a weak enhanced non-labeled remote sensing image pair, and the change detection result comprises a first change detection result, a second change detection result and a third change detection result.
In one embodiment, calculating the supervision loss according to the first change detection result and the real tag further includes: and calculating the first change detection result and the real label according to the cross entropy loss function and the dice loss function to obtain the supervision loss.
In one embodiment, the strong enhancement processing of the weakly enhanced unlabeled remote sensing image pair to obtain a distorted image further includes: carrying out color enhancement and/or shape enhancement on the weakly enhanced non-tag remote sensing images with preset intensity and preset combination to obtain distorted images, wherein the preset intensity and the preset combination for carrying out color enhancement and/or shape enhancement on the same pair of weakly enhanced non-tag remote sensing images are the same, and the preset intensity and the preset combination for carrying out color enhancement and/or shape enhancement on different pairs of weakly enhanced non-tag remote sensing images are random;
in one embodiment, the strongly enhancing the third change detection result and performing confidence threshold filter to obtain the pseudo tag further includes:
performing color enhancement and/or shape enhancement of preset intensity and preset combination on the third change detection result to obtain a strongly enhanced third change detection result;
and inputting the strongly enhanced third change detection result into a confidence coefficient threshold filter, setting a confidence coefficient threshold of the confidence coefficient threshold filter, eliminating the pixel prediction value of the strongly enhanced third change detection result smaller than the confidence coefficient threshold, and reserving the pixel prediction value of the strongly enhanced third change detection result larger than the confidence coefficient threshold to obtain the pseudo label.
In one embodiment, the calculating the unsupervised loss according to the second change detection result and the pseudo tag further includes: and calculating the second change detection result and the pseudo label according to the cross entropy loss function and the dice loss function to obtain the unsupervised loss.
A semi-supervised remote sensing image change detection apparatus, the apparatus comprising:
the image input module is used for inputting a labeled remote sensing image pair, a real label and a non-labeled remote sensing image pair, wherein the real label is a size-matched change gray level image of the labeled remote sensing image pair;
the monitoring loss calculation module is used for carrying out weak enhancement processing on the labeled remote sensing image pair to obtain a weak enhanced labeled remote sensing image pair, inputting the weak enhanced labeled remote sensing image pair into the change detection model to obtain a first change detection result, and calculating according to the first change detection result and the real label to obtain monitoring loss;
the unsupervised loss calculation module is used for carrying out weak enhancement processing on the non-tag remote sensing image pair to obtain a weak enhancement non-tag remote sensing image pair, carrying out strong enhancement processing on the weak enhancement non-tag remote sensing image pair to obtain a distorted image, and inputting the distorted image into the change detection model to obtain a second change detection result; inputting the weakly-enhanced label-free remote sensing image pair into a change detection model to obtain a third change detection result, carrying out strong enhancement processing on the third change detection result, filtering by using a confidence threshold filter to obtain a pseudo label, and calculating according to the second change detection result and the pseudo label to obtain unsupervised loss;
and the model training module is used for obtaining total loss through weighting calculation according to the supervised loss and the unsupervised loss, using the minimum total loss as a target function, and training the change detection model by using the optimizer to obtain a semi-supervised remote sensing image change detection model.
A computer device comprising a memory and a processor, the memory storing a computer program, the processor implementing the following steps when executing the computer program:
inputting a labeled remote sensing image pair, a real label and a non-labeled remote sensing image pair, wherein the real label is a size-matched change gray level image of the labeled remote sensing image pair;
carrying out weak enhancement processing on the labeled remote sensing image pair to obtain a weak enhanced labeled remote sensing image pair, inputting the weak enhanced labeled remote sensing image pair into a change detection model to obtain a first change detection result, and calculating according to the first change detection result and a real label to obtain supervision loss;
carrying out weak enhancement processing on the non-tag remote sensing image pair to obtain a weak enhancement non-tag remote sensing image pair, carrying out strong enhancement processing on the weak enhancement non-tag remote sensing image pair to obtain a distorted image, and inputting the distorted image into a change detection model to obtain a second change detection result;
inputting the weakly-enhanced label-free remote sensing image pair into a change detection model to obtain a third change detection result, carrying out strong enhancement processing on the third change detection result, filtering by using a confidence threshold filter to obtain a pseudo label, and calculating according to the second change detection result and the pseudo label to obtain unsupervised loss;
and calculating according to the supervision loss and the unsupervised loss to obtain the total loss, taking the minimum total loss as a target function, and training a change detection model by using an optimizer to obtain a semi-supervised remote sensing image change detection model.
According to the semi-supervised remote sensing image change detection method, the semi-supervised remote sensing image change detection device and the computer equipment, the remote sensing image change detection model is trained by using a small number of labeled remote sensing image pairs and a large number of unlabelled remote sensing image pairs, so that the problem of insufficient label data is solved, and the problem that a large amount of labor and time cost are consumed for labeling labels is solved; distortion change generated after image strong enhancement processing is ignored by training the remote sensing image change detection model, so that the change of an image object is concentrated, and the semantic understanding capability of the remote sensing image change detection model is improved; the remote sensing image change detection model is optimized by reducing the total loss, and the semi-supervised remote sensing image change detection model is obtained, so that the robustness and the generalization of the model are improved, and the change detection precision of the remote sensing image is also improved. Compared with the prior art, the semi-supervised remote sensing image change detection method, the semi-supervised remote sensing image change detection device and the computer equipment effectively utilize a small number of labeled remote sensing image pairs and a large number of unlabeled remote sensing image pairs to improve the change detection precision of the remote sensing images, overcome the problem of insufficient label data and release the potential of remote sensing big data.
Drawings
FIG. 1 is a schematic flow chart illustrating a method for detecting semi-supervised remote sensing image changes in one embodiment;
FIG. 2 is a schematic diagram illustrating an embodiment of a process for detecting changes in satellite images by using a semi-supervised remote sensing image change detection method;
FIG. 3 is a schematic diagram of a process for detecting image changes by the change detection model in one embodiment;
FIG. 4 is a visual display of a method of strong enhancement in one embodiment;
FIG. 5 is a visual display of randomly combined strong enhancement methods in one embodiment;
FIG. 6 is a visual presentation of confidence threshold filter filtering in one embodiment;
FIG. 7 is a diagram showing the test results of the change detection of the remote sensing image to be detected to the input semi-supervised remote sensing image change detection model in one embodiment;
FIG. 8 is a diagram illustrating an internal structure of a computer device according to an embodiment.
Detailed Description
In order to make the objects, technical solutions and advantages of the present application more apparent, the present application is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present application and are not intended to limit the present application.
In one embodiment, as shown in fig. 1, a semi-supervised remote sensing image change detection method is provided, which includes the following steps:
and step S1, inputting a labeled remote sensing image pair, a real label and a non-labeled remote sensing image pair, wherein the real label is a change gray level image of the labeled remote sensing image pair with matched size.
It can be understood that a small amount of (for example, dozens of) input labeled remote sensing image pairs and real labels can train a change detection model, the change detection model is guided to understand tasks, the manual work and time cost for manufacturing the labels can be reduced, and the input huge unlabeled remote sensing image pairs can effectively prevent the change detection model from being over-fitted on the small amount of labeled remote sensing image pairs, so that the robustness and the generalization of the change detection model are improved.
And step S2, carrying out weak enhancement processing on the labeled remote sensing image pair to obtain a weak enhanced labeled remote sensing image pair, inputting the weak enhanced labeled remote sensing image pair into a change detection model to obtain a first change detection result, and calculating according to the first change detection result and the real label to obtain the supervision loss.
It is understood that the weak enhancement processing refers to an image enhancement method with weak interference on the remote sensing image, and comprises translation and overturning.
It can be understood that the change detection model aims to identify semantic changes of remote sensing images acquired in the same area at different time phases to obtain a change detection result, wherein the change detection result is a change gray image of the remote sensing images matched with the size.
It can be understood that according to the characteristic that the number of unchanged pixels is far larger than the number of changed pixels in the remote sensing image change detection, the first change detection result and the real label can be calculated according to the cross entropy loss function and the dice loss function, and the supervision loss is obtained.
And step S3, carrying out weak enhancement processing on the unlabeled remote sensing image pair to obtain a weak enhancement unlabeled remote sensing image pair, carrying out strong enhancement processing on the weak enhancement unlabeled remote sensing image pair to obtain a distorted image, and inputting the distorted image into the change detection model to obtain a second change detection result.
It is understood that the strong enhancement process is a relatively weak enhancement process, which means an enhancement capable of generating a large distortion in the remote sensing image, and is composed of 3 to 4 color enhancement and/or shape enhancement methods.
And step S4, inputting the weakly-enhanced unlabeled remote sensing image pair into a change detection model to obtain a third change detection result, carrying out strong enhancement processing on the third change detection result, filtering by using a confidence threshold filter to obtain a pseudo label, and calculating according to the second change detection result and the pseudo label to obtain unsupervised loss.
It can be understood that, for the same pair of unlabeled remote sensing image pairs, the third change detection result is subjected to strong enhancement processing and confidence threshold filter filtering to obtain a pseudo label in step S4, and the weak enhancement unlabeled remote sensing image pair is subjected to strong enhancement processing in step S3 to obtain a distorted image, and the preset combination and the preset intensity of the two strong enhancement processing are consistent, so that the pseudo label is ensured to correspond to the pixel of the third change detection result one by one; otherwise, the preset combination and the preset intensity of the two times of strong enhancement processing are random, and the pseudo label does not correspond to the pixel of the third change detection result.
It can be understood that the third change detection result is subjected to strong enhancement processing to obtain a strong enhancement third change detection result, then the strong enhancement third change detection result is input into a confidence threshold filter for filtering, a pixel prediction value of the strong enhancement third change detection result which is larger than the confidence threshold is reserved, the pixel prediction value of the strong enhancement third change detection result which is larger than the confidence threshold is used as a pseudo label, and according to the characteristic that the number of unchanged pixels is far larger than the number of changed pixels in the remote sensing image change detection, the second change detection result and the pseudo label can be calculated according to a cross entropy loss function and a dice loss function to obtain unsupervised loss.
And step S5, calculating according to the supervised loss and the unsupervised loss to obtain the total loss, taking the minimum total loss as a target function, and training a change detection model by using an optimizer to obtain a semi-supervised remote sensing image change detection model.
It can be understood that the total loss is obtained through weighting calculation according to the supervised loss and the unsupervised loss, the minimum total loss is used as an objective function, parameters in the change detection model are continuously optimized through an AdamW optimization algorithm of an optimizer, and when the total loss does not decrease any more, the parameters in the change detection model at the moment are stored as final parameters, so that the semi-supervised remote sensing image change detection model is obtained.
According to the semi-supervised remote sensing image change detection method, the remote sensing image change detection model is trained by using a small number of labeled remote sensing image pairs and a large number of unlabelled remote sensing image pairs, so that the problem of insufficient label data is solved, and the problem that a large amount of labor and time cost are consumed for labeling labels is solved; distortion change generated after image strong enhancement processing is ignored by training the remote sensing image change detection model, so that the change of an image object is concentrated, and the semantic understanding capability of the remote sensing image change detection model is improved; the remote sensing image change detection model is optimized by reducing the total loss, and the semi-supervised remote sensing image change detection model is obtained, so that the robustness and the generalization of the model are improved, and the change detection precision of the remote sensing image is also improved. Compared with the prior art, the semi-supervised remote sensing image change detection method provided by the invention effectively utilizes a small number of labeled remote sensing image pairs and a large number of unlabeled remote sensing image pairs to improve the change detection precision of the remote sensing image, overcomes the problem of insufficient label data, and releases the potential of large remote sensing data.
In one embodiment, as shown in FIG. 2, a labeled remote sensing image pair X is input l The remote sensing images with the labels in the same pair are composed of two remote sensing images with the labels in the same region and different time phases, and the remote sensing images with the labels in different pairs are composed of two remote sensing images with the labels in different regions and different time phases;
inputting a real label Y l The real label is a labeled remote sensing image pair X l A size-matched change grayscale image;
input label-free remote sensing image pair X u The same pair of non-tag remote sensing images consists of two non-tag remote sensing images in the same area and at different time phases, and the different pair of non-tag remote sensing images consists of two non-tag remote sensing images in different areas and at different time phases.
It can be understood that the change detection model can be trained by inputting a small number (such as dozens of) labeled remote sensing image pairs and real labels, the change detection model is guided to understand tasks, the manual work and time cost for manufacturing the labels can be reduced, and the input huge unlabeled remote sensing image pairs can effectively prevent the change detection model from being over-fitted on the small number of labeled remote sensing image pairs, so that the robustness and the generalization of the change detection model are improved.
In one embodiment, the tagged remote sensing image pair is paired with X l Carrying out translation and/or turnover with preset amplitude to obtain weakly enhanced labeled remote sensing image pair X lw The same pair of labeled remote sensing images are translated and/or turned over at the same preset amplitude to obtain the same pair of weakly enhanced labeled remote sensing images, and different pairs of labeled remote sensing images are translated and/or turned over at random preset amplitudes to obtain different pairs of weakly enhanced labeled remote sensing images, so that the positions corresponding to the same pixel in the same pair of weakly enhanced labeled remote sensing images are ensured to be consistent;
taking unlabelled remote sensing image to X u Carrying out translation and/or turnover with preset amplitude to obtain weak enhancement label-free remote sensing image pair X w And carrying out translation and/or overturning at the same preset amplitude on different pairs of unlabeled remote sensing images to obtain different pairs of weakly-enhanced unlabeled remote sensing images, thereby ensuring that the corresponding positions of the same pixel in the same pair of weakly-enhanced unlabeled remote sensing images are consistent.
In one embodiment, as shown in fig. 3, the change detection model is composed of a siamese encoder and a decoder, wherein the encoder further comprises a drawing attention module composed of a multi-head drawing attention layer and a simple drawing attention layer;
the Siamese encoder carries out downsampling processing on the image through sharing weight and parameters, and the image characteristics of the image are extracted
N is the number of the image characteristics extracted by the Siam encoder, and the image characteristics are extracted
As nodes, image fusion characteristic values of the nodes are obtained by using image characteristics corresponding to neighbor nodes of the graph attention module fusion node
In particular, an image feature is identified
As a node i of the graph, fusing image characteristics corresponding to a neighbor node j of the node i by using a graph attention module
Obtaining the image fusion characteristic value of the node i
Is shown as
Wherein W represents a weight coefficient, N i Indicates the number of nodes, a ij Denotes the attention coefficient, expressed as
Wherein e is ij Representing the importance of a neighbor node j to node i, k another neighbor node, e ik Denotes the importance of the neighbor node k to the node i, a T Representing the inverse vector of the weight vector, | | represents the splicing operation;
the multi-head graph attention layer in the graph attention module fuses images of a node i output by each attention layer by using N groups of mutually independent attention layers sharing weight and parametersCharacteristic value
Splicing to obtain the final image fusion characteristic value of the node i
Is composed of
Wherein,
denotes the attention coefficient, W, corresponding to the nth group of attention layers n Representing the weight coefficient corresponding to the nth group attention layer;
inputting the final image fusion characteristic values of all nodes into an attention unit for convolution to obtain attention characteristics fused with an attention mechanism, and extracting the attention characteristics by a decoder to obtain a change detection result, wherein the image comprises a weak-enhancement labeled remote sensing image pair X lw Distorted image X s Weak enhancement of unlabeled remote sensing image pairs X w The change detection result includes a first change detection result
Second change detection result
And third change detection result
The method can be understood, the change detection model obtains image characteristics through a sample unloading of the Siamese encoder, the change detection result of the image is obtained by reconstructing the image characteristics through sampling on the decoder, the Siamese encoder is connected with the decoder through a jump connection mode, deep semantic information and shallow space information of the image can be better integrated, wherein a graph attention module in the Siamese encoder is fused with the image characteristics corresponding to neighbor nodes through an attention mechanism, weight self-adaption matching of different neighbors is realized, and therefore the accuracy of the change detection model is improved.
In one embodiment, the cross-entropy loss function L is used ce And die loss function L dice Detecting the first change
And a genuine label Y l Calculating to obtain supervision loss of
Wherein the cross entropy loss function L ce And die loss function L dice Are respectively represented as
In the formula, Y represents a label,
indicates the change detection result, where H is the height of the image, W is the width of the image,
a pixel value of an m-th pixel of the image, c is 0 or 1, and c indicates whether the m-th pixel of the image is changed or not.
It can be understood that the monitoring loss is calculated by adopting a mixed cross entropy loss function and a dice loss function, so that the influence of unbalanced change caused by the fact that the unchanged pixels are far larger than the changed pixels in the remote sensing image change detection is weakened.
In one embodimentIn, weak enhancement of unlabeled remote sensing image to X w Performing color enhancement and/or shape enhancement of preset intensity and preset combination to obtain distorted image X s Will distort the image X s Inputting the first change detection result into the change detection model to obtain a second change detection result
Specifically, as shown in fig. 4, the color enhancement and the shape enhancement mainly include several enhancement methods of brightness variation, color variation, contrast variation, equalization, hue separation, horizontal translation, rotation, sharpening, horizontal clipping, vertical clipping, exposure, vertical translation, and clipping, and specifically, as shown in fig. 5, the color enhancement and/or the shape enhancement of the preset intensity and the preset combination means that three enhancement methods other than clipping and a clipping method are selected and combined to form a strong enhancement method, wherein the intensity ranges of the three enhancement methods other than clipping and the clipping method are shown in table 1
| Enhancement method | Range of intensity | Description of the invention |
| Brightness of light | [0.05,0.95] | Changing the brightness of an image |
| Colour(s) | [0.05,0.95] | Changing color balance of an image |
| Contrast ratio | [0.05,0.95] | Changing contrast of an image |
| Equalization | / | Histogram equalization of images |
| Tone separation | [4,8] | Reducing the number of bits per color channel |
| Rotate | [-30,30] | Rotating an image |
| Sharpening | [0.05,0.95] | Adjusting a degree of sharpening of an image |
| Transverse shearing | [-0.3,0.3] | Shearing an image along a horizontal axis |
| Longitudinal shearing | [-0.3,0.3] | Shearing an image along a vertical axis |
| Exposure method | [0,256] | Inverting all pixel values that exceed a threshold |
| Transverse translation | [-0.3,0.3] | In waterIs translated in a horizontal direction |
| Longitudinal translation | [-0.3,0.3] | Translation in the vertical direction |
| Cutting out | [0.25,0.35] | Cropping an area from an image |
TABLE 1 enhancement method and Strength Range and description thereof
Specifically, the preset intensity and the preset combination for carrying out color enhancement and/or shape enhancement on the same pair of weakly enhanced unlabeled remote sensing images are the same, and the preset intensity and the preset combination for carrying out color enhancement and/or shape enhancement on different pairs of weakly enhanced unlabeled remote sensing images are random.
It can be understood that the strong enhancement is mainly used for two purposes, one is to increase or reduce the color difference between image features through color enhancement, and the other is to generate deformation through shape enhancement, and the change of the image object is absorbed through training the remote sensing image change detection model to ignore the distortion change of the image generated through color enhancement and/or shape enhancement, so as to improve the semantic understanding capability of the remote sensing image change detection model.
In one embodiment, weakly enhanced unlabeled remote sensing images are imaged to X w Inputting the third change detection result into the change detection model
Detecting the third change
Color enhancement and/or shape enhancement of a preset intensity and a preset combination are carried out to obtain a strong enhancement third change detection result;
specifically, as shown in fig. 4, the color enhancement and the shape enhancement mainly include several enhancement methods of brightness change, color change, contrast change, equalization, hue separation, horizontal translation, rotation, sharpening, horizontal shearing, vertical shearing, exposure, vertical translation, and cropping, specifically, as shown in fig. 5, the color enhancement and/or the shape enhancement of the preset intensity and the preset combination refers to selecting three enhancement methods other than cropping and combining the three enhancement methods into one strong enhancement method, wherein the intensity ranges of the three enhancement methods other than cropping and the cropping method are shown in table 1;
as shown in fig. 6, the strongly enhanced third change detection result is input into the confidence threshold filter, the confidence threshold τ of the confidence threshold filter is set, the pixel prediction value of the strongly enhanced third change detection result smaller than the confidence threshold τ is removed, the pixel prediction value of the strongly enhanced third change detection result larger than the confidence threshold τ is retained, and the pseudo tag is obtained
Wherein the size of the pixel prediction value is between 0 and 1.
It can be understood that by reserving the pixel prediction value of the strongly enhanced third change detection result which is greater than the confidence threshold τ, that is, reserving the white part in fig. 6, it is avoided that the pixel prediction value of the strongly enhanced third change detection result which is less than the confidence threshold τ is regarded as a pseudo label to mislead the precision of the change detection of the remote sensing image.
In one embodiment, the cross-entropy loss function L is used ce And die loss function L dice Detecting the second change
And a pseudo tag
Calculating to obtain unsupervised loss L u Is composed of
According to the loss of supervision L s And supervision loss L u Carrying out weighted calculation to obtain total loss L CD Is composed of
L CD ＝L s +λL u
In the formula, λ represents a weighting coefficient;
will lose L CD The minimum is used as an objective function, parameters in the change detection model are continuously optimized by using an AdamW optimization algorithm of an optimizer, and when the total loss L is CD And when the image does not descend any more, saving the parameters in the change detection model at the moment as final parameters to obtain a semi-supervised remote sensing image change detection model M.
It can be understood that the monitoring loss is calculated by adopting a mixed cross entropy loss function and a dice loss function, so that the influence of change imbalance caused by the fact that the unchanged pixels are far larger than the changed pixels in the remote sensing image change detection is weakened.
To further verify the semi-supervised remote sensing image change detection method provided by the invention, 50 pairs of labeled remote sensing images and 1950 pair of unlabelled remote sensing images are input to train to obtain a semi-supervised change detection model, the pairs of input to-be-detected remote sensing images are input to the semi-supervised change detection model, and the change detection results of the pairs of to-be-detected remote sensing images are output to obtain the change detection results of the pairs of to-be-detected remote sensing images, as shown in FIG. 7, in the figure, the first two columns are ImageA and ImageB of the input to-be-detected remote sensing images, the third column is a real label, the fourth column is the change detection results of the model using 50 pairs of labeled remote sensing images and 1950 pair of unlabelled remote sensing images, the fifth column is the change detection results of the model using 50 pairs of labeled remote sensing images and 1950 pair of unlabelled remote sensing images to perform semi-supervised training, and comparison can find that the change detection results of the remote sensing images obtained by the semi-supervised remote sensing image change detection method provided by the invention are closer to the real label, the accuracy of the change detection is higher.
It should be understood that although the various steps in the flow charts of fig. 1-3 are shown in order as indicated by the arrows, the steps are not necessarily performed in order as indicated by the arrows. The steps are not performed in the exact order shown and described, and may be performed in other orders, unless explicitly stated otherwise. Moreover, at least some of the steps in fig. 1-3 may include multiple sub-steps or multiple stages that are not necessarily performed at the same time, but may be performed at different times, and the order of performance of the sub-steps or stages is not necessarily sequential, but may be performed in turn or alternating with other steps or at least some of the sub-steps or stages of other steps.
In one embodiment, a semi-supervised remote sensing image change detection device is provided, comprising: the device comprises an image input module, a supervision loss calculation module, an unsupervised loss calculation module and a model training module, wherein:
and the image input module is used for inputting a labeled remote sensing image pair, a real label and a non-labeled remote sensing image pair, wherein the real label is a size-matched change gray level image of the labeled remote sensing image pair.
It can be understood that a small amount of (for example, dozens of) input labeled remote sensing image pairs and real labels can train a change detection model, the change detection model is guided to understand tasks, the manual work and time cost for manufacturing the labels can be reduced, and the input huge unlabeled remote sensing image pairs can effectively prevent the change detection model from being over-fitted on the small amount of labeled remote sensing image pairs, so that the robustness and the generalization of the change detection model are improved.
And the supervision loss calculation module is used for carrying out weak enhancement processing on the labeled remote sensing image pair to obtain a weak enhanced labeled remote sensing image pair, inputting the weak enhanced labeled remote sensing image pair into the change detection model to obtain a first change detection result, and calculating according to the first change detection result and the real label to obtain the supervision loss.
It is understood that the weak enhancement processing refers to an image enhancement method with weak interference on the remote sensing image, and comprises translation and overturning.
It can be understood that the change detection model aims to identify semantic changes of remote sensing images acquired in the same region at different time phases to obtain a change detection result, wherein the change detection result is a change gray level image of the remote sensing images matched with the size.
It can be understood that according to the characteristic that the number of unchanged pixels is far larger than the number of changed pixels in the remote sensing image change detection, the first change detection result and the real label can be calculated according to the cross entropy loss function and the dice loss function, and the supervision loss is obtained.
The unsupervised loss calculation module is used for carrying out weak enhancement processing on the non-tag remote sensing image pair to obtain a weak enhancement non-tag remote sensing image pair, carrying out strong enhancement processing on the weak enhancement non-tag remote sensing image pair to obtain a distorted image, and inputting the distorted image into the change detection model to obtain a second change detection result; and inputting the weakly-enhanced label-free remote sensing image pair into a change detection model to obtain a third change detection result, carrying out strong enhancement processing on the third change detection result, filtering by using a confidence threshold filter to obtain a pseudo label, and calculating according to the second change detection result and the pseudo label to obtain unsupervised loss.
It is understood that the strong enhancement process is a relatively weak enhancement process, which means an enhancement capable of generating a large distortion in the remote sensing image, and is composed of 3 to 4 color enhancement and/or shape enhancement methods.
It can be understood that for the same pair of unlabeled remote sensing image pairs, the third change detection result is subjected to strong enhancement processing and is filtered by a confidence threshold filter to obtain a pseudo label, the weak enhancement unlabeled remote sensing image pair is subjected to strong enhancement processing to obtain a distorted image, and the preset combination and the preset intensity of the two strong enhancement processing are consistent, so that the pseudo label is ensured to correspond to the pixels of the third change detection result one by one; otherwise, the preset combination and the preset intensity of the two times of strong enhancement processing are random, and the pseudo label does not correspond to the pixel of the third change detection result.
It can be understood that the third change detection result is subjected to strong enhancement processing to obtain a strong enhancement third change detection result, then the strong enhancement third change detection result is input into a confidence threshold filter for filtering, a pixel prediction value of the strong enhancement third change detection result which is larger than the confidence threshold is reserved, the pixel prediction value of the strong enhancement third change detection result which is larger than the confidence threshold is used as a pseudo label, and according to the characteristic that the number of unchanged pixels is far larger than the number of changed pixels in the remote sensing image change detection, the second change detection result and the pseudo label can be calculated according to a cross entropy loss function and a dice loss function to obtain unsupervised loss.
And the model training module is used for obtaining total loss through weighting calculation according to the supervised loss and the unsupervised loss, using the minimum total loss as a target function, and training the change detection model by using the optimizer to obtain a semi-supervised remote sensing image change detection model.
It can be understood that the total loss is obtained through weighting calculation according to the supervised loss and the unsupervised loss, the minimum total loss is used as an objective function, parameters in the change detection model are continuously optimized through an AdamW optimization algorithm of an optimizer, and when the total loss does not decrease any more, the parameters in the change detection model at the moment are stored as final parameters, so that the semi-supervised remote sensing image change detection model is obtained.
The specific limitations of the semi-supervised remote sensing image change detection device can be referred to the limitations of the semi-supervised remote sensing image change detection method in the above, and are not described herein again. All modules in the semi-supervised remote sensing image change detection device can be completely or partially realized by software, hardware and a combination thereof. The modules can be embedded in a hardware form or independent from a processor in the computer device, and can also be stored in a memory in the computer device in a software form, so that the processor can call and execute operations corresponding to the modules.
In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as shown in fig. 8. The computer device includes a processor, a memory, a network interface, a display screen, and an input device connected by a system bus. Wherein the processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device comprises a nonvolatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of an operating system and computer programs in the non-volatile storage medium. The network interface of the computer device is used for communicating with an external terminal through a network connection. The computer program is executed by a processor to realize a semi-supervised remote sensing image change detection method. The display screen of the computer equipment can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer equipment can be a touch layer covered on the display screen, a key, a track ball or a touch pad arranged on the shell of the computer equipment, an external keyboard, a touch pad or a mouse and the like.
Those skilled in the art will appreciate that the architecture shown in fig. 8 is merely a block diagram of some of the structures associated with the disclosed aspects and is not intended to limit the computing devices to which the disclosed aspects apply, as particular computing devices may include more or less components than those shown, or may combine certain components, or have a different arrangement of components.
In one embodiment, there is provided a computer device comprising a memory storing a computer program and a processor implementing the following steps when the processor executes the computer program:
inputting a labeled remote sensing image pair, a real label and a non-labeled remote sensing image pair, wherein the real label is a size-matched change gray level image of the labeled remote sensing image pair; carrying out weak enhancement processing on the labeled remote sensing image pair to obtain a weak enhanced labeled remote sensing image pair, inputting the weak enhanced labeled remote sensing image pair into a change detection model to obtain a first change detection result, and calculating according to the first change detection result and a real label to obtain supervision loss; carrying out weak enhancement processing on the non-tag remote sensing image pair to obtain a weak enhancement non-tag remote sensing image pair, carrying out strong enhancement processing on the weak enhancement non-tag remote sensing image pair to obtain a distorted image, and inputting the distorted image into a change detection model to obtain a second change detection result; inputting the weakly-enhanced label-free remote sensing image pair into a change detection model to obtain a third change detection result, carrying out strong enhancement processing on the third change detection result, filtering by using a confidence threshold filter to obtain a pseudo label, and calculating according to the second change detection result and the pseudo label to obtain unsupervised loss; and calculating according to the supervision loss and the unsupervised loss to obtain the total loss, taking the minimum total loss as a target function, and training a change detection model by using an optimizer to obtain a semi-supervised remote sensing image change detection model.
Claims (10)
1. A semi-supervised remote sensing image change detection method is characterized by comprising the following steps:
inputting a labeled remote sensing image pair, a real label and a non-labeled remote sensing image pair, wherein the real label is a size-matched change gray level image of the labeled remote sensing image pair;
carrying out weak enhancement processing on the labeled remote sensing image pair to obtain a weak enhanced labeled remote sensing image pair, inputting the weak enhanced labeled remote sensing image pair into a change detection model to obtain a first change detection result, and calculating according to the first change detection result and the real label to obtain supervision loss;
carrying out weak enhancement processing on the unlabeled remote sensing image pair to obtain a weak enhancement unlabeled remote sensing image pair, carrying out strong enhancement processing on the weak enhancement unlabeled remote sensing image pair to obtain a distorted image, and inputting the distorted image into the change detection model to obtain a second change detection result;
inputting the weakly-enhanced label-free remote sensing image pair into the change detection model to obtain a third change detection result, carrying out the strongly-enhanced processing on the third change detection result, filtering by using a confidence threshold filter to obtain a pseudo label, and calculating according to the second change detection result and the pseudo label to obtain unsupervised loss;
and calculating according to the supervision loss and the unsupervised loss to obtain total loss, taking the minimum total loss as a target function, and training the change detection model by using an optimizer to obtain a semi-supervised remote sensing image change detection model.
2. The method of claim 1, wherein inputting the pair of tagged remote sensing images, the pair of real tagged remote sensing images, and the pair of untagged remote sensing images comprises:
inputting the labeled remote sensing image pairs, wherein the same pair of labeled remote sensing images consists of two labeled remote sensing images in the same area and at different time phases, and the different pair of labeled remote sensing images consists of two labeled remote sensing images in different areas and at different time phases;
inputting the real label, wherein the real label is a change gray image with the label remote sensing image matched with the size;
and inputting the label-free remote sensing image pairs, wherein the same pair of label-free remote sensing images consists of two label-free remote sensing images in the same area and at different time phases, and the different pair of label-free remote sensing images consists of two label-free remote sensing images in different areas and at different time phases.
3. The method of claim 1, wherein weakly enhancing the labeled remote sensing image pair to obtain a weakly enhanced labeled remote sensing image pair comprises:
carrying out translation and/or overturning of the labeled remote sensing image pair with a preset amplitude to obtain a weakly enhanced labeled remote sensing image pair, wherein the same pair of labeled remote sensing images are carried out translation and/or overturning with the same preset amplitude to obtain the same pair of weakly enhanced labeled remote sensing images, and the different pairs of labeled remote sensing images are carried out translation and/or overturning with random preset amplitudes to obtain different pairs of weakly enhanced labeled remote sensing images;
and carrying out weak enhancement processing on the unlabeled remote sensing image pair to obtain a weak enhancement unlabeled remote sensing image pair, which comprises the following steps:
and carrying out translation and/or overturning with preset amplitude on the unlabeled remote sensing image pair to obtain a weak enhancement unlabeled remote sensing image pair, wherein carrying out translation and/or overturning with the same preset amplitude on the same pair of unlabeled remote sensing images to obtain the same pair of weak enhancement unlabeled remote sensing images, and carrying out translation and/or overturning with random preset amplitude on different pairs of unlabeled remote sensing images to obtain different pairs of weak enhancement unlabeled remote sensing images.
4. The method of claim 1, wherein said change detection model is comprised of a siamese encoder and a decoder, wherein said encoder further comprises a graph attention module:
the siamesed encoder carries out downsampling processing on an image through shared weight and parameters, extracts image characteristics of the image, takes the image characteristics as nodes, utilizes an image attention module to fuse image characteristics corresponding to neighbor nodes of the nodes to obtain image fusion characteristic values of the nodes, inputs the image fusion characteristic values into an attention unit to obtain attention characteristics fused with an attention system, and a decoder obtains a change detection result through extracting the attention characteristics, wherein the image comprises a weakly enhanced labeled remote sensing image pair, a distorted image and a weakly enhanced non-labeled remote sensing image pair, and the change detection result comprises a first change detection result, a second change detection result and a third change detection result.
5. The method of claim 1, wherein calculating a surveillance loss based on the first change detection and the authentic tag comprises:
and calculating the first change detection result and the real label by using a cross entropy loss function and a dice loss function to obtain the supervision loss.
6. The method of claim 1, wherein strongly enhancing the weakly enhanced unlabeled remote sensing image pair to obtain a distorted image comprises:
and carrying out color enhancement and/or shape enhancement on the weakly enhanced unlabeled remote sensing images with preset intensity and preset combination to obtain a distorted image, wherein the preset intensity and the preset combination for carrying out color enhancement and/or shape enhancement on the same pair of weakly enhanced unlabeled remote sensing images are the same, and the preset intensity and the preset combination for carrying out color enhancement and/or shape enhancement on the different pairs of weakly enhanced unlabeled remote sensing images are random.
7. The method of claim 1, wherein performing the strong enhancement processing and the confidence threshold filter filtering on the third change detection result to obtain a pseudo tag comprises:
carrying out color enhancement and/or shape enhancement of preset intensity and preset combination on the third change detection result to obtain a strong enhanced third change detection result;
inputting the strongly enhanced third change detection result into the confidence coefficient threshold filter, setting the confidence coefficient threshold of the confidence coefficient threshold filter, eliminating the pixel prediction value of the strongly enhanced third change detection result smaller than the confidence coefficient threshold, and reserving the pixel prediction value of the strongly enhanced third change detection result larger than the confidence coefficient threshold to obtain the pseudo label.
8. The method of claim 1, wherein calculating an unsupervised loss from the second change detection and the pseudo tag comprises:
and calculating the second change detection result and the pseudo label according to the cross entropy loss function and the dice loss function to obtain the unsupervised loss.
9. A semi-supervised remote sensing image change detection device, characterized in that the device comprises:
the image input module is used for inputting a labeled remote sensing image pair, a real label and a non-labeled remote sensing image pair, wherein the real label is a change gray level image with the matched size of the labeled remote sensing image pair;
the monitoring loss calculation module is used for carrying out weak enhancement processing on the labeled remote sensing image pair to obtain a weak enhanced labeled remote sensing image pair, inputting the weak enhanced labeled remote sensing image pair into a change detection model to obtain a first change detection result, and calculating according to the first change detection result and the real label to obtain monitoring loss, wherein the first change detection result is a change gray level image matched with the weak enhanced labeled remote sensing image pair in size;
the unsupervised loss calculation module is used for carrying out weak enhancement processing on the unlabeled remote sensing image pair to obtain a weak enhancement unlabeled remote sensing image pair, carrying out strong enhancement processing on the weak enhancement unlabeled remote sensing image pair to obtain a distorted image, and inputting the distorted image into the change detection model to obtain a second change detection result, wherein the second change detection result is a change gray level image matched with the distorted image in size; inputting the weakly-enhanced label-free remote sensing image pair into the change detection model to obtain a third change detection result, carrying out strong enhancement processing on the third change detection result, filtering by using a confidence threshold filter to obtain a pseudo label, and calculating according to the second change detection result and the pseudo label to obtain unsupervised loss, wherein the third change detection result is a change gray level image of the weakly-enhanced label-free remote sensing image pair with matched size;
and the model training module is used for obtaining total loss through weighting calculation according to the supervised loss and the unsupervised loss, using the minimum total loss as a target function, and training the change detection model by using an optimizer to obtain a semi-supervised remote sensing image change detection model.
10. A computer device comprising a memory and a processor, the memory storing a computer program, wherein the processor implements the steps of the method of any one of claims 1 to 8 when executing the computer program.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210550967.1A CN114882252B (en) | 2022-05-20 | 2022-05-20 | Semi-supervised remote sensing image change detection method, device and computer equipment |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210550967.1A CN114882252B (en) | 2022-05-20 | 2022-05-20 | Semi-supervised remote sensing image change detection method, device and computer equipment |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114882252A true CN114882252A (en) | 2022-08-09 |
| CN114882252B CN114882252B (en) | 2024-07-26 |
Family
ID=82677842
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210550967.1A Active CN114882252B (en) | 2022-05-20 | 2022-05-20 | Semi-supervised remote sensing image change detection method, device and computer equipment |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114882252B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115170575A (en) * | 2022-09-09 | 2022-10-11 | 阿里巴巴（中国）有限公司 | Method and equipment for remote sensing image change detection and model training |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110826458A (en) * | 2019-10-31 | 2020-02-21 | 河海大学 | Multispectral remote sensing image change detection method and system based on deep learning |
| WO2021022752A1 (en) * | 2019-08-07 | 2021-02-11 | 深圳先进技术研究院 | Multimodal three-dimensional medical image fusion method and system, and electronic device |
| CN112906638A (en) * | 2021-03-19 | 2021-06-04 | 中山大学 | Remote sensing change detection method based on multi-level supervision and depth measurement learning |
| CN113688665A (en) * | 2021-07-08 | 2021-11-23 | 华中科技大学 | Remote sensing image target detection method and system based on semi-supervised iterative learning |
-
2022
- 2022-05-20 CN CN202210550967.1A patent/CN114882252B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2021022752A1 (en) * | 2019-08-07 | 2021-02-11 | 深圳先进技术研究院 | Multimodal three-dimensional medical image fusion method and system, and electronic device |
| CN110826458A (en) * | 2019-10-31 | 2020-02-21 | 河海大学 | Multispectral remote sensing image change detection method and system based on deep learning |
| CN112906638A (en) * | 2021-03-19 | 2021-06-04 | 中山大学 | Remote sensing change detection method based on multi-level supervision and depth measurement learning |
| CN113688665A (en) * | 2021-07-08 | 2021-11-23 | 华中科技大学 | Remote sensing image target detection method and system based on semi-supervised iterative learning |
Non-Patent Citations (1)
| Title |
|---|
| 马博博: "基于深度学习的高分辨率遥感图像场景分类技术研究", 《CNKI电子期刊》, 31 May 2021 (2021-05-31), pages 1 - 78 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115170575A (en) * | 2022-09-09 | 2022-10-11 | 阿里巴巴（中国）有限公司 | Method and equipment for remote sensing image change detection and model training |
| CN115170575B (en) * | 2022-09-09 | 2022-12-23 | 阿里巴巴（中国）有限公司 | Method and equipment for remote sensing image change detection and model training |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114882252B (en) | 2024-07-26 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111553929B (en) | Mobile phone screen defect segmentation method, device and equipment based on converged network | |
| Hou et al. | High-resolution triplet network with dynamic multiscale feature for change detection on satellite images | |
| CN111080628B (en) | Image tampering detection method, apparatus, computer device and storage medium | |
| CN108647585B (en) | Traffic identifier detection method based on multi-scale circulation attention network | |
| Liu et al. | MSC-DNet: An efficient detector with multi-scale context for defect detection on strip steel surface | |
| CN111179217A (en) | Attention mechanism-based remote sensing image multi-scale target detection method | |
| CN112580439A (en) | Method and system for detecting large-format remote sensing image ship target under small sample condition | |
| CN114863236B (en) | Image target detection method based on dual-attention mechanism | |
| CN111257341A (en) | Underwater building crack detection method based on multi-scale features and stacked full convolution network | |
| CN109886159B (en) | Face detection method under non-limited condition | |
| Li et al. | A review of deep learning methods for pixel-level crack detection | |
| Yin et al. | Visual attention dehazing network with multi-level features refinement and fusion | |
| CN113487610B (en) | Herpes image recognition method and device, computer equipment and storage medium | |
| Li et al. | Small target deep convolution recognition algorithm based on improved YOLOv4 | |
| Guo et al. | A novel transformer-based network with attention mechanism for automatic pavement crack detection | |
| Chu et al. | Fine‐grained crack segmentation for high‐resolution images via a multiscale cascaded network | |
| Yu et al. | Progressive refined redistribution pyramid network for defect detection in complex scenarios | |
| CN114429577B (en) | Flag detection method, system and equipment based on high confidence labeling strategy | |
| Dong et al. | Field-matching attention network for object detection | |
| CN116665054A (en) | Remote sensing image small target detection method based on improved YOLOv3 | |
| Li et al. | A grid‐based classification and box‐based detection fusion model for asphalt pavement crack | |
| Zhang et al. | Residual attentive feature learning network for salient object detection | |
| CN115984226A (en) | Insulator defect detection method, device, medium, and program product | |
| Wang et al. | SLMS-SSD: Improving the balance of semantic and spatial information in object detection | |
| CN114882252A (en) | Semi-supervised remote sensing image change detection method and device and computer equipment |
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 |