CN115587964A - Entropy screening-based pseudo label cross consistency change detection method - Google Patents

Entropy screening-based pseudo label cross consistency change detection method Download PDF

Info

Publication number
CN115587964A
CN115587964A CN202211010881.6A CN202211010881A CN115587964A CN 115587964 A CN115587964 A CN 115587964A CN 202211010881 A CN202211010881 A CN 202211010881A CN 115587964 A CN115587964 A CN 115587964A
Authority
CN
China
Prior art keywords
entropy
pseudo
change
change detection
consistency
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202211010881.6A
Other languages
Chinese (zh)
Inventor
骆春波
徐加朗
罗杨
刘翔
孙文健
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Yangtze River Delta Research Institute of UESTC Huzhou
Original Assignee
Yangtze River Delta Research Institute of UESTC Huzhou
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Yangtze River Delta Research Institute of UESTC Huzhou filed Critical Yangtze River Delta Research Institute of UESTC Huzhou
Priority to CN202211010881.6A priority Critical patent/CN115587964A/en
Publication of CN115587964A publication Critical patent/CN115587964A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/0002—Inspection of images, e.g. flaw detection
    • 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
    • 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
    • G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
    • 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
    • G06V10/7753—Incorporation of unlabelled data, e.g. multiple instance learning [MIL]
    • 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
    • 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
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
    • 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
    • G06T2207/20—Special algorithmic details
    • G06T2207/20084—Artificial neural networks [ANN]

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Computing Systems (AREA)
  • Databases & Information Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • Medical Informatics (AREA)
  • Multimedia (AREA)
  • Quality & Reliability (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Molecular Biology (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Image Analysis (AREA)

Abstract

The invention belongs to the technical field of optical remote sensing image change detection, and discloses a pseudo label cross consistency change detection method based on entropy screening, which comprises the following steps: constructing a pseudo label cross consistency change detection model EPCCDM based on entropy screening; inputting the labeled double-time phase image pair and the unlabeled double-time phase image pair into the EPCCDM to obtain four change probability maps; executing an entropy-based dynamic threshold strategy on the change probability graph to obtain a pseudo label; training the EPCCDM by jointly using supervision loss and pseudo label cross consistency loss; after the training is finished, argmax processing is carried out on the change probability graph, the class with the highest probability is used as the class of the pixel, the class to which each pixel belongs is obtained, and the prediction change graph is generated. According to the invention, on the premise of only using a small amount of labeled data, effective change detection of the ultra-high definition remote sensing image is realized by using label-free data.

Description

Entropy screening-based pseudo label cross consistency change detection method
Technical Field
The invention belongs to the technical field of optical remote sensing image change detection, and particularly relates to a pseudo label cross consistency change detection method based on entropy screening.
Background
At present, the purpose of optical remote sensing image change detection is to identify significant changes between a pair of images taken at different times in the same region, and the method is mainly used for analyzing the change condition of the earth surface, such as the evolution of a water body, the development trend of buildings, the change of roads and the like. As an important and challenging task in earth observation, change detection has wide application in natural disaster assessment, city planning, resource management, forest felling monitoring, and the like.
For the task of detecting the change of the optical remote sensing image, the number of high-quality labeled data sets is small, and the practical application of the deep learning model to the task is limited. With the improvement of the earth observation level, it is not difficult to acquire images of different time phases and the same region required by a change detection task, but due to the remarkable improvement of the image resolution, the detailed labeling of the image pairs is increasingly difficult. The training data for change detection needs to compare images of different time phases at the same time and label the images with pixel-level precision, which is more expensive than other tasks (such as image classification and target detection). Currently, there are a large number of two-phase image pairs without true tags, and how to effectively utilize these unlabeled data is a major focus of the change detection task in recent years. Recently, with the excellent performance of semi-supervised learning in the deep learning field, it is also introduced into the change detection task, which makes it possible to train a change detection model with excellent performance with a small number of labeled data sets.
At present, semi-supervised learning work for change detection is mainly divided into two categories: self-training and consistency regularization. Self-training has three steps: (1) training a model on the labeled data; (2) Generating a pseudo label on the label-free data set by using a pre-training model; (3) The model is retrained again using the labeled dataset, unlabeled dataset, and their pseudo-labels. In the three steps, the pseudo label is a key factor influencing the final performance of the model, and the higher the quality of the generated pseudo label is, the more easily the performance of the model is improved during retraining. For example, prior art 1 first converts a bi-temporal image pair into a graph, which includes labeled nodes and unlabeled nodes, and then trains a countermeasure generation network to generate pseudo labels for the unlabeled nodes.
However, self-training is off-line, requiring training in steps. To enable online training, consistency regularization is proposed. The main idea of consistency regularization is: after the same sample is disturbed, the original sample before disturbance and the sample after disturbance are input into the network, and two outputs of the network should be similar. Consistency regularization is based on two assumptions: a smoothness assumption and a clustering assumption, wherein the smoothness assumption means that samples with similar distances tend to have the same class label in a feature space; clustering assumption means that the decision boundary for the model to predict should be in the region where the sample distribution density is low. Consistency regularization imposes consistency on each perturbed sample, encouraging the network to produce similar distributions for the samples before and after perturbation, thereby optimizing in the direction of reducing the distance between samples of the same category. Prior art 2 uses two discriminators to impose a feature consistency constraint between the entropy of tagged and untagged data sets. In prior art 3, a semi-supervised change detection method based on Mean Teacher is proposed, which inputs a sample after disturbance to a student network, simultaneously inputs a sample before disturbance to a Teacher network and applies the same disturbance to the output of the Teacher network, and finally realizes consistency regularization by reducing the Mean square error between the output of the student network and the output of the Teacher network.
The current semi-supervised change detection method has two key problems:
(1) The quality of the false label is low. However, because the available labeled data amount is small, the quality of the pseudo label generated by the model is often low, the existing semi-supervised change detection method does not screen the pseudo label, so that the low-quality pseudo label directly participates in the model training, extra noise is introduced, the training difficulty is improved, and the performance is reduced.
(2) The problem of consistency regularization in combination with false tags. Semi-supervised learning change detection methods based on consistency regularization, while enabling on-line training, do not make good use of pseudo-label information. As shown in fig. 1 (a), this kind of method often uses random gaussian noise, rotation, flip, clipping, scaling transformation, RGB offset, brightness and contrast offset, etc. to perturb the samples, the structure and initialization parameter values of the two nets are the same, and the consistency regularization is achieved by directly constraining the outputs of the two nets.
Through the above analysis, the problems and defects of the prior art are as follows: the existing semi-supervised learning change detection method has the disadvantages of low quality of pseudo labels, incapability of combining consistency regularization with the pseudo labels, incapability of realizing and effectively detecting the change of ultra-high-definition remote sensing images; the existing supervised learning change detection method can only use labeled data for training.
Disclosure of Invention
Aiming at the problems in the prior art, the invention provides a pseudo label cross consistency change detection method based on entropy screening.
The invention is realized in such a way that a pseudo label cross consistency change detection method based on entropy screening comprises the following steps:
constructing a pseudo label cross consistency change detection model EPCCDM based on entropy screening;
inputting a labeled double-time phase image pair and a non-labeled double-time phase image pair into two change detection branches of the EPCCDM, and performing SoftMax processing on the outputs of the two change detection branches to obtain four change probability graphs;
step three, executing an entropy-based dynamic threshold strategy on the four obtained change probability graphs to obtain corresponding pseudo labels; training the EPCCDM by jointly using supervision loss and pseudo label cross consistency loss based on the change probability map, the pseudo labels and the real labels;
and step four, after the training is finished, performing Argmax processing on the change probability graph, taking the category with the highest probability as the category of the pixels, obtaining the category to which each pixel belongs, and generating a prediction change graph.
Further, the entropy-screening-based pseudo tag cross consistency change detection model EPCCDM comprises: two variation detection branches N having the same structure but different initialization parameter values theta θ1 、N θ2 PLCC and EDT;
the two change detection branches N θ1 、N θ2 Based on the codec architecture, comprises two ResNet-18 encoders sharing weight and one decoder;
each encoder for outputting 5 levels of features
Figure BDA0003809130760000041
Where T = { T1, T2}, i = {1,2,3,4,5}, { C 1 ,C 2 ,C 3 ,C 4 ,C 5 }＝{64,64,128,256,512}；
The decoder is used for acquiring the characteristics output by the encoder and compressing the channel number of the characteristics into 256 by utilizing a 1 multiplied by 1 convolutional layer; simultaneously fusing different characteristics by directly adding for each fusion node, and passing through a group of 3 × 3, 1 × 1, 3 × 3 convolution layers after adding; the number of convolution kernels of the convolution layer is 256, 16 and 256 respectively;
the EPCCDM initializes the two change detection branches N with different parameter values θ1 And N θ2 The decoder of (1).
Further, the second step comprises:
first, a labeled two-time phase image pair is acquired (T1) l ,T2 l ) And unlabeled dual-temporal image pair (T1) ul ,T2 ul )；
Second, branch N is detected by two changes of the EPCCDM θ1 、N θ2 And SoftMax processing
Figure BDA0003809130760000042
Obtaining four variation probability maps
Figure BDA0003809130760000043
And
Figure BDA0003809130760000044
Figure BDA0003809130760000045
Figure BDA0003809130760000046
wherein the content of the first and second substances,
Figure BDA0003809130760000047
2 represents the number of categories; the number of categories is changed/unchanged; h O And W O Representing the original height and width of the input image, respectively.
Further, the third step includes:
(1) For four variation probability maps
Figure BDA0003809130760000048
Executing entropy-based dynamic threshold strategy screening to obtain reliable pixel points and obtain corresponding pseudo labels
Figure BDA0003809130760000049
(2) The PLCC of the EPCCDM carries out cross supervision on the change probability chart through a pseudo label, and two change detection branches N θ1 、N θ2 Carrying out consistency regularization; at the same time, the EPCCDM will supervise the loss L sup Cross-tag consistency loss L plcc Training is performed as a loss function.
Further, the step (1) includes:
1) Calculating the entropy of all pixel points in the change probability map by using the following formula to obtain an entropy map H iter ：
Figure BDA0003809130760000051
Wherein the content of the first and second substances,
Figure BDA0003809130760000052
two values at the j-th pixel point position in the change probability map are obtained;
2) Entropy is larger than beta using an entropy-based dynamic threshold strategy EDT iter All pixel point positions of the percentile are excluded as unreliable positions, and Argmax processing is carried out on the rest positions in the change probability diagram to obtain a pseudo label P;
each pixel point j in the pseudo label P is calculated by the following formula:
Figure BDA0003809130760000053
wherein, γ iter Representing the threshold of the iter iteration, taking an entropy diagram H iter Middle beta iter Entropy of percentile as gamma iter ；
3) Threshold value gamma using EDT iter And (3) carrying out dynamic adjustment:
Figure BDA0003809130760000054
where iter represents the number of current training iterations, and max _ iter represents the maximum number of iterations.
Further, the step (2) comprises:
(2.1) utilization ofThrough N θ1 And pseudo tag (P) generated after EDT 1 l ,P 1 ul ) Supervision N θ2 Output change probability map
Figure BDA0003809130760000055
And use of N θ2 Pseudo tag of
Figure BDA0003809130760000056
Supervision N θ1 Change probability map of (2)
Figure BDA0003809130760000057
(2.2) determining the EPCCDM Total loss function L PCCDM ＝L sup +L plcc ；
Wherein L is plcc Indicating a false tag cross-consistency loss; l is sup Representing two variation detection branches N θ1 、N θ2 Two cross entropy losses of, said two change detection branches N θ1 、N θ2 Two cross entropy losses of
Figure BDA0003809130760000058
And its corresponding real label (Y) l ,Y l ) Calculated supervision loss:
Figure BDA0003809130760000061
wherein the content of the first and second substances,
Figure BDA0003809130760000062
representing a labeled training set;
Figure BDA0003809130760000063
representing a label-free training set; l is CE Represents the cross entropy loss; { (T1, T2), Y } represents a training sample; j represents the total number of pixels; (T1, T2) represents a two-phase image pair; pr (Y) j = c | (T1, T2); θ) represents the probability that the jth pixel belongs to the c-th class; the c-th class is changed/unchangedMelting; y denotes a real tag.
Further, in the third step, training the EPCCDM by jointly using a supervision loss and a pseudo tag cross-consistency loss includes:
firstly, initializing encoders of two change detection branches of the EPCCDM by using ImageNet pre-training weights, and using a Kaiming random initialization method for decoders of the two branches to enable initialization parameters between the decoders to be different;
secondly, two change detection branches both adopt Adam optimizer and apply initial learning rate lr o Is set to be 1 x 10 -4 (ii) a Using a multivariate learning rate attenuation strategy, setting the batch size to be 4 and the maximum iteration number to be 20000, training:
Figure BDA0003809130760000064
wherein, lr n Indicates a new learning rate, lr o Denotes the initial learning rate, iter denotes the current number of iterations, and max _ iter denotes the maximum number of iterations.
Another object of the present invention is to provide a computer device, which includes a memory and a processor, the memory stores a computer program, and the computer program, when executed by the processor, causes the processor to execute the steps of the entropy-screening-based pseudo tag cross consistency change detection method.
Another object of the present invention is to provide a computer readable storage medium, which stores a computer program, and when the computer program is executed by a processor, the processor executes the steps of the pseudo tag cross consistency change detection method based on entropy screening.
Another object of the present invention is to provide an information data processing terminal for executing the entropy-screening-based pseudo tag cross consistency change detection method.
In combination with the technical solutions and the technical problems to be solved, please analyze the advantages and positive effects of the technical solutions to be protected in the present invention from the following aspects:
first, aiming at the technical problems existing in the prior art and the difficulty in solving the problems, the technical problems to be solved by the technical scheme of the present invention are closely combined with results, data and the like in the research and development process, and some creative technical effects are brought after the problems are solved. The specific description is as follows:
the invention provides a pseudo label cross consistency change detection method based on entropy screening, which solves the problems of low pseudo label quality and consistency regularization and pseudo label combination in a semi-supervised learning change detection method, and realizes effective ultra-high definition remote sensing image change detection by using label-free data on the premise of only using a small amount of labeled data.
The invention trains a deep learning model by using a small amount of labeled data and a large amount of unlabeled data, and realizes the change detection of the double-temporal ultrahigh remote sensing image pair. Aiming at the problem of low quality of pseudo labels in the semi-supervised learning change detection method, the EPCCDM comprises an entropy-based dynamic threshold strategy, namely EDT, EDT filters unreliable pixel points with high entropy, the pixel points with the entropy lower than the threshold are selected as relatively reliable pseudo labels (because the entropy describes the uncertainty of a random variable value, the lower the entropy is, the smaller the uncertainty is represented), and the threshold can be dynamically adjusted along with the rise of training iteration times, so that the loss of pseudo label information can be reduced.
The EPCCDM comprises Pseudo-label Cross Consistency (PLCC), and the PLCC restrains two change detection branches with the same structure and different initialization parameter values, so that two inputs subjected to different disturbances through the two branches generate a change graph which tends to be consistent.
The EPCCDM can be trained by using the labeled data and the unlabeled data at the same time, so that better change detection performance is obtained, and experimental results show that compared with the three prior art, the EPCCDM obtains the highest F1 and Kappa indexes in four proportions of a small sample reference data set, and the change graph obtained by the method has better details, clearer boundaries and higher internal integrity. The entropy-based dynamic threshold strategy included in the EPCCDM can ignore the false detection area of the pseudo label to a certain extent, and screens out the pseudo label with relatively high quality, so that the training difficulty is reduced, the model is optimized towards the correct direction, and the smaller the proportion of the labeled data amount in all the training data is, the more remarkable the improvement effect of the entropy-based dynamic threshold strategy on the change detection performance is. The cross consistency of the pseudo labels contained in the EPCCDM can be better combined with the pseudo labels and consistency regularization, and the high-quality pseudo labels are screened out through the EDT, so that the PLCC can more accurately execute the consistency regularization, and the performance of semi-supervised learning change detection is improved.
Secondly, considering the technical scheme as a whole or from the perspective of products, the technical effect and advantages of the technical scheme to be protected by the invention are specifically described as follows:
the difference between the present invention and the prior art is as follows:
the invention is similar to Cross Pseudo Supervision provided by the prior art, and the difference is as follows: cross Pseudo Supervision of the prior art applies to semantic segmentation tasks, whereas the Pseudo-tag Cross-consistency of the present invention applies to change detection tasks. The Cross Pseudo super vision in the prior art directly executes Argmax operation on the output of the network to obtain a Pseudo label, but the Pseudo label Cross consistency of the invention needs to execute entropy-based dynamic threshold strategy on the output of the network to obtain the Pseudo label. The encoder in the change detection branch of the invention directly uses the ResNet-18 network without the last AvgPool, softMax and full connection layer, and the decoder is built by itself. In general, the method for detecting the cross consistency change of the pseudo label based on entropy screening is original in a change detection task as a whole.
The invention designs pseudo label cross consistency, as shown in (b) of fig. 2, the pseudo label cross consistency applies different disturbances to the same input sample through two network branches with the same structure and different initialization parameters, screens out high-quality pseudo labels through an entropy-based dynamic threshold strategy, and then utilizes the pseudo labels to constrain the output of the two branches in a cross supervision manner.
Third, as an inventive supplementary proof of the claims of the present invention, there are also presented several important aspects:
the technical scheme of the invention fills the technical blank in the industry at home and abroad:
the invention provides a pseudo label cross consistency change detection method based on entropy for the first time, and the method uses pseudo label cross consistency to carry out consistency regularization on the output of a network and the obtained higher-quality pseudo label after screening the pseudo label by using a dynamic threshold strategy based on entropy, thereby realizing high-performance end-to-end semi-supervised change detection, improving the performance of the semi-supervised change detection, filling the blank of reasonably screening the pseudo label in the field of change detection, and providing a guidance method for effectively utilizing non-label change detection data to improve the change detection performance.
Drawings
FIG. 1 is a basic framework diagram of a consistency regularization change detection method provided by an embodiment of the present invention; in the figure, → represents the data at the tail part and loss supervision is carried out on the data at the arrow, and → is provided with// does not carry out back propagation;
FIG. 1 (a) is a basic framework diagram of a consistency regularization change detection method provided by an embodiment of the present invention;
fig. 1 (b) is a basic framework diagram of a method for detecting cross-consistency change of a pseudo tag according to an embodiment of the present invention;
FIG. 2 is a flowchart of a pseudo tag cross consistency change detection method based on entropy screening according to an embodiment of the present invention;
FIG. 3 is an overall framework diagram of an EPCCDM provided by the embodiment of the invention; in the figure: different colors of the fusion nodes represent different initialization parameter values; the presence// at → denotes no back propagation; - → shows the data of the tail part and carries out loss supervision on the data of the arrow; (T1) l ,T2 l ) And Y l Is a double epoch in a tagged data setAgainst an image pair and its true label; (T1) ul ,T2 ul ) Is a dual-temporal image in the unlabeled dataset;
Figure BDA0003809130760000091
and
Figure BDA0003809130760000092
is a variation probability map; (P) 1 l ,P 1 ul ) And
Figure BDA0003809130760000093
is the generated pseudo tag;
FIG. 4 shows an entropy-based dynamic threshold strategy according to an embodiment of the present invention iter A change curve graph along with the training iteration number iter;
FIG. 5 is an overall framework of the comparison method Mean Teacher provided by the embodiments of the present invention;
FIG. 6 is a schematic diagram of pseudo-label visualization results provided by an embodiment of the invention when training on a Google dataset using 1/32 labeled data; in the figure: black indicates an unchanged area, white indicates a changed area, and gray indicates an ignored area;
FIG. 6 (a) is a schematic diagram of a real tag provided by an embodiment of the present invention;
FIG. 6 (b) is a diagram of the pre-training period (. Beta.) provided by an embodiment of the present invention iter 70) no EDT schematic;
FIG. 6 (c) is a graph of the pre-training period (. Beta.) provided by an embodiment of the present invention iter About 70) with EDT scheme;
FIG. 6 (d) shows the middle training period (β) provided by the embodiment of the present invention iter 80) EDT free schematic;
FIG. 6 (e) is a diagram of the mid-training period (β) provided by an embodiment of the present invention iter 80) with EDT scheme;
FIG. 6 (f) is a diagram of the late training period (β) provided by an embodiment of the present invention iter 97) no EDT schematic;
FIG. 6 (g) is a diagram of the late training period (β) provided by an embodiment of the present invention iter H 97) EDT scheme;
FIG. 7 is a graphical representation of the predicted results of a Google data set trained using 1/8 of the labeled data according to an embodiment of the present invention;
FIG. 7 (a) is a schematic diagram of the T1 image result provided by the embodiment of the present invention;
FIG. 7 (b) is a diagram illustrating the result of T2 image provided by the embodiment of the present invention;
FIG. 7 (c) is a diagram illustrating the result of the true tag provided by the embodiment of the present invention;
FIG. 7 (d) is a diagram showing EPCCDM results provided by the embodiment of the present invention;
FIG. 7 (e) is a diagram illustrating the EPCCDM _ MET result of the comparison method provided by the embodiment of the present invention;
FIG. 7 (f) is a SNUNet-CD/48 result diagram of the comparison method provided by the embodiment of the invention;
FIG. 7 (g) is a diagram illustrating the results of the FC-Sim-conc comparison method provided by the embodiment of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail with reference to the following embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
1. The embodiments are explained. This section is an explanatory embodiment expanding on the claims so as to fully understand how the present invention is embodied by those skilled in the art.
As shown in fig. 2, the method for detecting cross consistency change of pseudo tag based on entropy screening according to the embodiment of the present invention includes:
s101, constructing a pseudo label cross consistency change detection model EPCCDM based on entropy screening;
s102, inputting a labeled double-time phase image pair and a non-labeled double-time phase image pair into two change detection branches of the EPCCDM, and performing SoftMax processing on the outputs of the two change detection branches to obtain four change probability graphs;
s103, executing an entropy-based dynamic threshold strategy on the four obtained change probability graphs to obtain corresponding pseudo labels; training the EPCCDM by jointly using supervision loss and pseudo label cross consistency loss based on the change probability map, the pseudo labels and the real labels;
and S104, after the training is finished, executing Argmax processing on the change probability graph, taking the class with the highest probability as the class of the pixel, obtaining the class to which each pixel belongs, and generating a prediction change graph.
The method for detecting the cross consistency change of the pseudo label based on entropy screening specifically comprises the following steps:
1. integral frame
The overall framework of the EPCCDM is shown in FIG. 3, and includes two change detection branches N having the same structure but different initialization parameter values θ θ1 、N θ2 PLCC and EDT.
Given a labeled two-phase image pair (T1) l ,T2 l ) And unlabeled dual-phase image pair (T1) ul ,T2 ul ) Through N θ1 、N θ2 And SoftMax operation
Figure BDA0003809130760000111
Obtaining four variation probability maps
Figure BDA0003809130760000112
And
Figure BDA0003809130760000113
Figure BDA0003809130760000114
wherein 2 represents the number of classes (changed/unchanged), H O And W O Representing the original height and width of the input image, respectively. The overall process is shown as follows:
Figure BDA0003809130760000115
Figure BDA0003809130760000121
then the four change probability maps
Figure BDA0003809130760000122
Executing EDT, screening out reliable pixel points, and obtaining corresponding pseudo labels
Figure BDA0003809130760000123
The PLCC carries out cross supervision on the change probability chart through a pseudo label, thereby carrying out cross supervision on two branches N θ1 、N θ2 A consistency regularization is performed. EPCCDM will supervise loss L sup Loss of cross-consistency with a pseudo-label L plcc Training is performed as a loss function. Finally, to
Figure BDA0003809130760000124
And applying Argmax operation to pick the class with the highest probability as the class of the pixel, thereby generating the prediction change map. In the test phase, branch N is detected using only one change in EPCCDM θ1 And (5) predicting the result.
2. Change detection Branch N θ1 、N θ2
A simple change detection branch is designed for EPCCDM, as shown in fig. 3. The change detection branch of the EPCCDM is based on a codec architecture, comprising two ResNet-18 encoders sharing weights and one decoder. Each encoder outputs 5 levels of features
Figure BDA0003809130760000125
Where T = { T1, T2}, i = {1,2,3,4,5}, { C 1 ,C 2 ,C 3 ,C 4 ,C 5 } = {64,64,128,256,512}. The decoder receives the features from the encoder and first compresses the number of channels of the features to 256 using 1 x 1 convolutional layers. Then, each fusion node fuses different features through a direct addition mode, and passes through a group of convolution layers of 3 × 3, 1 × 1 and 3 × 3 after addition, and the number of convolution kernels of the group of convolution layers is 256, 16 and 256 respectively. Notably, EPCCDM initializes two change checks with different parameter valuesSide branch N θ1 And N θ2 Thereby adding different perturbations to the input samples.
3. Entropy-based dynamic threshold policy
The quality of the pseudo label can influence the effect of semi-supervised learning, additional noise can be introduced into an error pseudo label, the training difficulty is improved, and even the performance is reduced. In order to obtain a high-quality pseudo label, an Entropy-based Dynamic Threshold (EDT) strategy is designed to screen out reliable pixel points. Specifically, the EDT selects a pixel point with an entropy lower than a threshold value as a pseudo label, and the threshold value is dynamically adjusted along with the rise of the training iteration number, so as to reduce the loss of pseudo label information. The entropy describes the uncertainty of a random variable value, and the lower the entropy, the smaller the uncertainty. Since the change probability map CD is an output obtained by subjecting the input image pair (T1, T2) to the change detection branch and SoftMax operation, where each pixel point has two values representing the probabilities of the pixel point belonging to the changed and unchanged categories, respectively, the uncertainty of the pixel point can be measured by calculating the entropy at each pixel point in the change probability map, specifically, defining
Figure BDA0003809130760000131
For two values at the j-th pixel position in the change probability map, the calculation process of the entropy of the pixel is as follows:
Figure BDA0003809130760000132
in each training iteration, the entropy of all pixel positions in the change probability graph is calculated to obtain an entropy graph H iter EDT will have entropy greater than beta iter All pixel point positions of the percentile are excluded as unreliable positions, argmax operation is carried out on the rest positions in the change probability graph, and a pseudo label P is obtained, namely the calculation process of each pixel point j in the pseudo label P is as follows:
Figure BDA0003809130760000133
wherein gamma is iter Representing the threshold of the iter iteration, taking entropy diagram H iter Middle beta iter Entropy of percentile as gamma iter 。
Under the supervision of the loss function, the prediction result gradually tends to be reliable along with the increase of the training iteration number. Because the neglected pixel points in the pseudo label do not participate in the calculation of the loss function, in order to utilize more pseudo label information in the later training period, the EDT is used for the threshold value gamma iter Making dynamic adjustments, i.e. beta iter The curve rises with the increase of the number of training iterations, and the rising curve is shown in fig. 4, so that more pixel points are taken into consideration in the later period of training, and the calculation process is as follows:
Figure BDA0003809130760000134
where iter is the current number of training iterations and max _ iter is the maximum number of iterations.
In summary, the overall flow of EDT is shown in algorithm 1.
Figure BDA0003809130760000135
Figure BDA0003809130760000141
4. Pseudo tag cross-consistency
As shown in fig. 3, the PLCC detects a branch N for two changes of the same structure and different initialization parameter values θ1 、 N θ2 And (4) mutually constraining, so that the two inputs subjected to different disturbances through the two branches generate a variation graph which tends to be consistent. Specifically, the use of a passthrough N θ1 And pseudo tag (P) generated after EDT 1 l ,P 1 ul ) To supervise N θ2 Variation of outputProbability map
Figure BDA0003809130760000142
For the same reason, use N θ2 Pseudo tag of
Figure BDA0003809130760000143
To supervise N θ1 Change probability map of (2)
Figure BDA0003809130760000144
Consider a labeled training set
Figure BDA0003809130760000145
And a label-free training set
Figure BDA0003809130760000146
False tag cross-consistency loss L plcc Is defined as:
Figure BDA0003809130760000151
in the formula, L CE For cross entropy loss, consider a training sample { (T1, T2), Y }, L { (T1, T2), L }, L { (L2) } for cross entropy loss CE Is defined as:
Figure BDA0003809130760000152
where J is the total number of pixels, (T1, T2) is a bi-temporal image pair, pr (Y) j = c | (T1, T2); θ) represents the probability that the jth pixel belongs to the c-th class (i.e., changed/unchanged), and Y is the true label.
5. Loss function
For a given labeled training set
Figure BDA0003809130760000153
And a label-free training set
Figure BDA0003809130760000154
The proposed semi-supervised training of EPCCDM involves two types of losses: firstly, the
Figure BDA0003809130760000155
And its corresponding genuine label (Y) l ,Y l ) Calculated supervision loss L sup I.e. two change detection branches N θ1 、N θ2 Two cross entropy losses of (a); secondly, by
Figure BDA0003809130760000156
And corresponding pseudo label
Figure BDA0003809130760000157
Calculated pseudo label cross consistency loss L plcc ，L plcc The calculation process of (2) is as follows:
Figure BDA0003809130760000158
to sum up, EPCCDM Total loss function L PCCDM ＝L sup +L plcc 。
6. Training process
In the training stage of the EPCCDM, the encoders of the two branches are initialized by ImageNet pre-training weights, and the decoders of the two branches are initialized by a Kaiming random initialization method, so that the initialization parameters of the decoders are different. Both branches adopt Adam optimizer and will initial learning rate lr o Is set to be 1 x 10 -4 . Using a multivariate learning rate decay strategy, that is, for each iteration, the decay formula of the learning rate is as follows:
Figure BDA0003809130760000159
in the formula lr n Representing a new learning rate, lr o Represents the initial learning rate, iter represents the current number of iterations, and max _ iter represents the maximum number of iterations.
The batch size is set to 4, i.e. 4 pairs of tagged dual phase images and 4 pairs of untagged dual phase images are read per iteration. The maximum number of iterations of training is set to 20000.
2. Application examples. In order to prove the creativity and the technical value of the technical scheme of the invention, the part is an application example of the technical scheme of the claims to a specific product or related technology.
The pseudo label cross consistency change detection method based on entropy screening provided by the embodiment of the invention is applied to computer equipment, the computer equipment comprises a memory and a processor, the memory stores a computer program, and the computer program is executed by the processor, so that the processor executes the steps of the pseudo label cross consistency change detection method based on entropy screening.
The method for detecting the pseudo label cross consistency change based on the entropy screening is applied to a computer-readable storage medium, and a computer program is stored, and when the computer program is executed by a processor, the processor executes the steps of the method for detecting the pseudo label cross consistency change based on the entropy screening.
The pseudo label cross consistency change detection method based on entropy screening provided by the embodiment of the invention is applied to an information data processing terminal, and the information data processing terminal is used for executing the pseudo label cross consistency change detection method based on entropy screening.
The pseudo label cross consistency change detection method based on entropy screening provided by the embodiment of the invention is applied to a computer cloud, and the change condition (namely, a change map) between two images can be obtained only by uploading a double-time phase image pair locally.
The pseudo label cross consistency change detection method based on entropy screening provided by the embodiment of the invention is applied to an automatic driving automobile, street information is shot in real time through a camera, and the spatial change condition (namely a change map) between two moments is obtained by using the method.
The pseudo label cross consistency change detection method based on entropy screening provided by the embodiment of the invention is applied to a remote sensing satellite, images at different moments are shot through the satellite, and the change conditions (namely change maps) of the same region at different moments are obtained by using the method.
3. Evidence of the relevant effects of the examples. The embodiment of the invention achieves some positive effects in the process of research and development or use, and has great advantages compared with the prior art, and the following contents are described by combining data, diagrams and the like in the test process.
1. Experimental setup
1.1.1 data set and preprocessing method thereof
The experiment used a Google dataset consisting of 20 pairs of seasonal VHR remote sensing images with a spatial resolution of 55 cm/pixel and a size distribution between 1006 x 1168 to 5184 x 6736 pixels. The data set covers suburbs in Guangzhou city of China, and focuses on the change situation of buildings. Each image is divided into 256 x 256 non-overlapping sub-images, each pair of sub-images containing at least one pixel variation, resulting in 1156 pairs of sub-images. These sub-image pairs are in accordance with 7:2: a scale of 1 is randomly divided to generate training, validation and test sets, i.e., training, validation and test sets containing 808, 232, 116 pairs of 256 x 256 images, respectively. 1/4, 1/8, 1/16 and 1/32 of the Google training set are used as labeled data, and the rest are used as unlabeled data, namely the numbers of the labeled image pairs and the unlabeled image pairs in the four training sets are respectively {202,606}, {101,707}, {50,758}, and {24,784}.
1.1.2 comparative methods
In order to compare the performance of the EPCCDM proposed from multiple angles, two leading edge existing supervised learning change detection technologies, FC-Sim-conc, SNUNet-CD/48 and a semi-supervised learning change detection technology Mean Teacher, are selected as comparison methods. To ensure fairness, mean Teacher uses the change detection branch of EPCCDM for semi-supervised learning, as shown in fig. 5. Mean Teacher is a popular semi-supervised learning method, comprising a Teacher network N θt And student network N θs It enforces a coherence constraint on both networks aiming to align the output of the student network with the output of the teacher network. In the training stage, each iteration is performed by a teacher networkParameter of the complex theta t By calculating student network parameters theta only s Is updated by Exponential Moving Average (EMA), the update procedure is as follows:
Figure BDA0003809130760000171
where α is a smoothing hyperparameter and is set to 0.99.
Parameter theta of student network s By calculating a loss function
Figure BDA0003809130760000172
The updating is carried out, and the updating is carried out,
Figure BDA0003809130760000173
the loss involves two parts: one is tagged data (T1) l ,T2 l ) Obtaining a prediction result through a student network
Figure BDA0003809130760000174
Rear and real label Y l Calculated supervision loss part L sup (ii) a Second, label-free data (T1) before/after disturbance ul ,T2 ul ) And tagged data (T1) l ,T2 l ) Inputting the result into the student/teacher network to obtain the prediction result of the student/teacher network
Figure BDA0003809130760000181
For is to
Figure BDA0003809130760000182
And
Figure BDA0003809130760000183
applying a consistency constraint as an unsupervised loss part L unsup Loss function
Figure BDA0003809130760000184
The calculation process is as follows:
Figure BDA0003809130760000185
in the formula L CE For cross entropy loss, L MSE Is the mean square error loss.
1.1.3 evaluation index
By changing the common evaluation indexes in the detection task: f1 fraction and Kappa coefficient to measure the performance of the change detection method.
(1) Fraction of F1
F1 is a weighted harmonic mean of precision and recall, which takes into account both precision and recall to balance the contradiction between the two, so as to better reflect the change detection capability of the method. The calculation of F1 is as follows:
Figure BDA0003809130760000186
in the formula, TP, FP, TN and FN are the numbers of true cases, false positives, true negatives and false negatives respectively, and Precision and Recall are Precision and Recall respectively.
(2) Kappa coefficient
Kappa reflects the consistency of the predicted value and the true value, the larger the value is, the closer the predicted value and the true value are represented, and the calculation process is as follows:
Figure BDA0003809130760000187
Figure BDA0003809130760000188
Figure BDA0003809130760000189
in the formula, OA and PRE represent the overall accuracy and the desired accuracy, respectively.
1.2 analysis of results
1.2.1 ablation test results
To perform ablation experiments, the EDT and PLCC were removed from the EPCCDM and the resulting reference model was only supervised learning. The results of the ablation experiments are shown in Table 1, and it can be seen that PLCC resulted in a significant performance improvement in all ratios, specifically, in 1/4, 1/8, 1/16, 1/32 ratios, the addition of PLCC resulted in a performance improvement of 2.74%, 3.12%, 1.39%, 2.45% F1 (3.18%, 5.13%, 4.03%, 3.78% kappa), respectively. The results show that semi-supervised learning has great potential in the task of change detection, and the designed PLCC is effective.
TABLE 1
Figure BDA0003809130760000191
The addition of EDT further improved performance, bringing about performance gains of 0.53%, 0.77%, 2.47%, 3.04% of F1 (0.85%, 0.65%, 3.1%, 3.52%) in the ratios 1/4, 1/8, 1/16, 1/32, respectively. It is worth noting that the performance gain of EDT when the proportion of the labeled training set is low (i.e. the proportion of 1/16 and 1/32) is much larger than that when the proportion is high, because the quality of the pseudo label predicted by the model is often relatively low when there is too little labeled data (e.g. 50 pairs at the proportion of 1/16 and 24 pairs at the proportion of 1/32), which causes more pseudo label noise to be introduced into the training process, at this time, it is more critical to filter noise and screen out more reliable pseudo labels by means of EDT, thereby reducing the training difficulty and optimizing the model to the correct direction. Fig. 6 visualizes the false label situation before and after EDT is used in a ratio of 1/32, and it can be seen that the false detection area of the false label can be ignored to some extent by using EDT in the early stage, the middle stage and the later stage of training, and the higher quality false label is obtained. The above experimental results show the importance of the quality of the pseudo labels and the effectiveness of the EDT, and the screening of the high-quality pseudo labels through the EDT can enable the PLCC to more accurately execute consistency regularization, thereby being beneficial to improving the performance of the semi-supervised learning change detection method.
In conclusion, the designed PLCC and EDT significantly improved the performance of the change detection model, and EPCCDM using PLCC and EDT improved the performance of 3.27%, 3.89%, 3.86%, 5.49%, F1%, and 4.03%, 5.78%, 7.13%, 7.3% kappa, in the ratios 1/4, 1/8, 1/16, 1/32, respectively, compared to the reference model.
1.2.2 comparative Experimental results
To further evaluate the effectiveness of the proposed EPCCDM, three leading edge methods were compared on the Google dataset as described in 1.1.3 and the quantitative results are shown in table 2. It can be seen that EPCCDM achieves the best performance at all scales. Compared with the existing semi-supervised learning technology Mean Teacher, the PLCC and EDT semi-supervised learning method used by the EPCCDM exceeds Mean Teacher, and F1 indexes of 0.93%, 1.07%, 2.8% and 3.52% and Kappa indexes of 0.94%, 1.93%, 4.55% and 4.33% are respectively improved in four proportions of 1/4, 1/8, 1/16 and 1/32, because PLCC carries out additional cross consistency constraint on a model through high-quality pseudo labels screened by EDT. Compared with two existing change detection technologies FC-Sim-conc and SNunet-CD/48 using supervised learning, EPCCDM is improved by at least 2.09% (3.28%), 1.44% (2.05%), 3.21% (4.7%), and 4.03% (4.4%) F1 (Kappa) in four proportions, which shows that the method can effectively use unlabeled data for training and improve the change detection result.
TABLE 2
Figure BDA0003809130760000201
For visual comparison, FIG. 7 shows a typical prediction after training at a 1/8 scale. It can be observed that there are many missed detections and false positives in the comparison method, and the proposed EPCCDM achieves the best visual performance, with a predicted variation graph more conforming to the real label. Specifically, EPCCDM greatly reduces missing detection and holes in a changed area of a building compared with the semi-supervised learning change detection technology Mean Teacher. Compared with two supervised learning change detection technologies, namely FC-Sim-conc and SNunet-CD/48, the EPCCDM has the advantages that the detection of the change area is more accurate, false alarms and uncertain areas are reduced, and a change graph with more accurate boundaries is generated.
It should be noted that embodiments of the present invention can be realized in hardware, software, or a combination of software and hardware. The hardware portion may be implemented using dedicated logic; the software portions may be stored in a memory and executed by a suitable instruction execution system, such as a microprocessor or specially designed hardware. It will be appreciated by those skilled in the art that the apparatus and methods described above may be implemented using computer executable instructions and/or embodied in processor control code, for example such code provided on a carrier medium such as a diskette, CD-or DVD-ROM, a programmable memory such as read-only memory (firmware) or a data carrier such as an optical or electronic signal carrier. The apparatus of the present invention and its modules may be implemented by hardware circuits such as very large scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field programmable gate arrays, programmable logic devices, or software executed by various types of processors, or a combination of hardware circuits and software, e.g., firmware.
The above description is only for the purpose of illustrating the present invention and the appended claims are not to be construed as limiting the scope of the invention, which is intended to cover all modifications, equivalents and improvements that are within the spirit and scope of the invention as defined by the appended claims.

Claims (10)

1. A pseudo label cross consistency change detection method based on entropy screening is characterized by comprising the following steps:
constructing a pseudo label cross consistency change detection model EPCCDM based on entropy screening;
inputting the labeled double-time phase image pair and the unlabeled double-time phase image pair into two change detection branches of the EPCCDM, and performing SoftMax processing on the outputs of the two change detection branches to obtain four change probability graphs;
step three, executing an entropy-based dynamic threshold strategy on the four obtained change probability graphs to obtain corresponding pseudo labels; training the EPCCDM by jointly using supervision loss and pseudo label cross consistency loss based on the change probability map, the pseudo labels and the real labels;
and step four, after the training is finished, executing Argmax processing on the change probability graph, taking the class with the highest probability as the class of the pixel, obtaining the class to which each pixel belongs, and generating a prediction change graph.
2. The entropy-screening-based pseudo tag cross consistency change detection method as claimed in claim 1, wherein the entropy-screening-based pseudo tag cross consistency change detection model EPCCDM includes: two variation detection branches N having the same structure but different initialization parameter values theta θ1 、N θ2 PLCC and EDT;
the two change detection branches N θ1 、N θ2 Based on the codec architecture, the ResNet-18 encoder and a decoder which comprise two shared weights;
each encoder for outputting 5 levels of features
Figure FDA0003809130750000011
Where T = { T1, T2}, i = {1,2,3,4,5}, { C 1 ,C 2 ,C 3 ,C 4 ,C 5 }＝{64,64,128,256,512}；
The decoder is used for acquiring the characteristics output by the encoder and compressing the channel number of the characteristics into 256 by utilizing a 1 multiplied by 1 convolutional layer; simultaneously fusing different characteristics by directly adding for each fusion node, and passing through a group of 3 × 3, 1 × 1, 3 × 3 convolution layers after adding; the number of convolution kernels of the convolution layer is 256, 16 and 256 respectively;
the EPCCDM initializes the two change detection scores with different parameter valuesBranch N θ1 And N θ2 The decoder of (1).
3. The method for detecting cross-consistency change of pseudo labels based on entropy screening as claimed in claim 1, wherein the second step comprises:
first, a labeled two-time phase image pair (T1) is acquired l ,T2 l ) And unlabeled dual-phase image pair (T1) ul ,T2 ul )；
Second, branch N is detected by two changes of the EPCCDM θ1 、N θ2 And SoftMax processing S (-) to obtain four variation probability graphs
Figure FDA0003809130750000021
And
Figure FDA0003809130750000022
Figure FDA0003809130750000023
Figure FDA0003809130750000024
wherein the content of the first and second substances,
Figure FDA0003809130750000025
2 represents the number of categories; the number of categories is changed/unchanged; h O And W O Representing the original height and width of the input image, respectively.
4. The method for detecting cross-consistency change of pseudo labels based on entropy screening as claimed in claim 1, wherein the third step comprises:
(1) For four variation probability maps
Figure FDA0003809130750000026
Executing entropy-based dynamic threshold strategy screening to obtain reliable pixel points and obtain corresponding pseudo labels
Figure FDA0003809130750000027
(2) The PLCC of the EPCCDM carries out cross supervision on the change probability chart through a pseudo label, and two change detection branches N θ1 、N θ2 Carrying out consistency regularization; while the EPCCDM will supervise the loss L sup Cross-tag consistency loss L plcc Training is performed as a loss function.
5. The method for detecting cross-consistency variation of pseudo labels based on entropy screening as claimed in claim 4, wherein the step (1) comprises:
1) Calculating the entropy of all pixel points in the change probability diagram by using the following formula to obtain an entropy diagram H iter ：
Figure RE-FDA0003972779250000028
Wherein, the first and the second end of the pipe are connected with each other,
Figure RE-FDA0003972779250000029
two values at the j-th pixel point position in the change probability map are obtained;
2) Entropy is larger than beta using an entropy-based dynamic threshold strategy EDT iter All pixel point positions of the percentile are excluded as unreliable positions, and Argmax processing is carried out on the rest positions in the change probability diagram to obtain a pseudo label P;
each pixel point j in the pseudo label P is calculated by the following formula:
Figure RE-FDA0003972779250000031
wherein, γ iter Representing the threshold of the iter iteration, taking entropy diagramH iter Middle beta iter Entropy of percentile as gamma iter ；
3) Threshold value gamma using EDT iter And (3) carrying out dynamic adjustment:
Figure RE-FDA0003972779250000032
where iter represents the number of current training iterations, and max _ iter represents the maximum number of iterations.
6. The method for detecting cross-consistency variation of pseudo labels based on entropy screening as claimed in claim 4, wherein the step (2) comprises:
(2.1) Using the passage of N θ1 And pseudo tag (P) generated after EDT 1 l ,P 1 ul ) Supervision N θ2 Output change probability map
Figure FDA0003809130750000033
And use of N θ2 Pseudo tag of
Figure FDA0003809130750000034
Supervision N θ1 Change probability map of
Figure FDA0003809130750000035
(2.2) determining the EPCCDM Total loss function L PCCDM ＝L sup +L plcc ；
Wherein L is plcc Indicating a false tag cross-consistency loss; l is a radical of an alcohol sup Representing two variation detection branches N θ1 、N θ2 Two cross entropy losses of, said two change detection branches N θ1 、N θ2 Two cross entropy losses of
Figure FDA00038091307500000310
And its corresponding real label (Y) l ,Y l ) Computed supervisionLoss:
Figure FDA0003809130750000036
wherein the content of the first and second substances,
Figure FDA0003809130750000037
representing a labeled training set;
Figure FDA0003809130750000038
representing unlabeled training sets
Figure FDA0003809130750000039
L CE Represents cross entropy loss; { (T1, T2), Y } represents a training sample; j represents the total number of pixels; (T1, T2) represents a two-phase image pair; pr (Y) j = c | (T1, T2); θ) represents the probability that the jth pixel belongs to the c-th class; the c-th class is changed/unchanged; y denotes a real tag.
7. The method for detecting change in pseudo-tag cross-consistency based on entropy screening of claim 1, wherein in the third step, training the EPCCDM by using a combination of a supervision loss and a pseudo-tag cross-consistency loss comprises:
firstly, initializing encoders of two change detection branches of the EPCCDM by using ImageNet pre-training weight, and using a Kaiming random initialization method for decoders of the two branches to enable initialization parameters between the decoders to be different;
secondly, two change detection branches both adopt Adam optimizer and apply initial learning rate lr o Is set to be 1 x 10 -4 (ii) a Using a multivariate learning rate attenuation strategy, setting the batch size to be 4 and the maximum iteration number to be 20000, training:
Figure FDA0003809130750000041
wherein, lr n Indicates a new learning rate, lr o Denotes the initial learning rate, iter denotes the current number of iterations, and max _ iter denotes the maximum number of iterations.
8. A computer arrangement, characterized in that the computer arrangement comprises a memory and a processor, the memory storing a computer program which, when executed by the processor, causes the processor to carry out the steps of the entropy screening-based pseudo tag cross consistency change detection method according to any one of claims 1 to 7.
9. A computer-readable storage medium storing a computer program which, when executed by a processor, causes the processor to perform the steps of the entropy screening-based pseudo tag cross consistency change detection method according to any one of claims 1 to 7.
10. An information data processing terminal, characterized in that the information data processing terminal is used for executing the entropy screening-based pseudo label cross consistency change detection method according to any one of claims 1 to 7.
CN202211010881.6A 2022-08-22 2022-08-22 Entropy screening-based pseudo label cross consistency change detection method Pending CN115587964A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202211010881.6A CN115587964A (en) 2022-08-22 2022-08-22 Entropy screening-based pseudo label cross consistency change detection method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202211010881.6A CN115587964A (en) 2022-08-22 2022-08-22 Entropy screening-based pseudo label cross consistency change detection method

Publications (1)

Publication Number Publication Date
CN115587964A true CN115587964A (en) 2023-01-10

Family

ID=84771270

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202211010881.6A Pending CN115587964A (en) 2022-08-22 2022-08-22 Entropy screening-based pseudo label cross consistency change detection method

Country Status (1)

Country Link
CN (1) CN115587964A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116310581A (en) * 2023-03-29 2023-06-23 南京信息工程大学 Semi-supervised change detection flood identification method
CN116665064A (en) * 2023-07-27 2023-08-29 城云科技（中国）有限公司 Urban change map generation method based on distillation generation and characteristic disturbance and application thereof

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116310581A (en) * 2023-03-29 2023-06-23 南京信息工程大学 Semi-supervised change detection flood identification method
CN116665064A (en) * 2023-07-27 2023-08-29 城云科技（中国）有限公司 Urban change map generation method based on distillation generation and characteristic disturbance and application thereof
CN116665064B (en) * 2023-07-27 2023-10-13 城云科技（中国）有限公司 Urban change map generation method based on distillation generation and characteristic disturbance and application thereof

Similar Documents

Publication Publication Date Title
CN110287927B (en) Remote sensing image target detection method based on depth multi-scale and context learning
CN111553929B (en) Mobile phone screen defect segmentation method, device and equipment based on converged network
CN112966684B (en) Cooperative learning character recognition method under attention mechanism
CN109671071B (en) Underground pipeline defect positioning and grade judging method based on deep learning
CN115587964A (en) Entropy screening-based pseudo label cross consistency change detection method
CN110956126B (en) Small target detection method combined with super-resolution reconstruction
CN112541508A (en) Fruit segmentation and recognition method and system and fruit picking robot
CN115393687A (en) RGB image semi-supervised target detection method based on double pseudo-label optimization learning
CN111915650B (en) Target tracking method and system based on improved twin network
CN113298058A (en) Water quality prediction inversion method and system based on hyperspectral image of unmanned aerial vehicle
CN109919302B (en) Training method and device for neural network of image
CN115565043A (en) Method for detecting target by combining multiple characteristic features and target prediction method
CN116168240A (en) Arbitrary-direction dense ship target detection method based on attention enhancement
CN117217368A (en) Training method, device, equipment, medium and program product of prediction model
CN116110022A (en) Lightweight traffic sign detection method and system based on response knowledge distillation
CN114565824A (en) Single-stage rotating ship detection method based on full convolution network
CN113591608A (en) High-resolution remote sensing image impervious surface extraction method based on deep learning
CN117437423A (en) Weak supervision medical image segmentation method and device based on SAM collaborative learning and cross-layer feature aggregation enhancement
CN116977710A (en) Remote sensing image long tail distribution target semi-supervised detection method
CN113344005B (en) Image edge detection method based on optimized small-scale features
CN116385466A (en) Method and system for dividing targets in image based on boundary box weak annotation
CN116596851A (en) Industrial flaw detection method based on knowledge distillation and anomaly simulation
CN111079807A (en) Ground object classification method and device
CN116363610A (en) Improved YOLOv 5-based aerial vehicle rotating target detection method
CN112396620A (en) Image semantic segmentation method and system based on multiple thresholds

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