CN114399687A - Semi-supervised self-training hyperspectral remote sensing image classification method based on spatial correction - Google Patents

Semi-supervised self-training hyperspectral remote sensing image classification method based on spatial correction Download PDF

Info

Publication number
CN114399687A
CN114399687A CN202111455830.XA CN202111455830A CN114399687A CN 114399687 A CN114399687 A CN 114399687A CN 202111455830 A CN202111455830 A CN 202111455830A CN 114399687 A CN114399687 A CN 114399687A
Authority
CN
China
Prior art keywords
local image
image blocks
semantic
spectral
cluster
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
CN202111455830.XA
Other languages
Chinese (zh)
Inventor
武越
白壮飞
肖郑磊
王善峰
公茂果
慕贵峰
苗启广
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.)
Xidian University
Original Assignee
Xidian University
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 Xidian University filed Critical Xidian University
Priority to CN202111455830.XA priority Critical patent/CN114399687A/en
Publication of CN114399687A publication Critical patent/CN114399687A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G06F18/2155—Generating training patterns; Bootstrap methods, e.g. bagging or boosting characterised by the incorporation of unlabelled data, e.g. multiple instance learning [MIL], semi-supervised techniques using expectation-maximisation [EM] or naïve labelling
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G06F18/232—Non-hierarchical techniques
    • G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
    • G06F18/23211—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with adaptive number of clusters
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G06F18/232—Non-hierarchical techniques
    • G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
    • G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Evolutionary Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Biophysics (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Probability & Statistics with Applications (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a semi-supervised self-training hyperspectral remote sensing image classification method based on spatial correction, which comprises the following steps of: segmenting the hyperspectral remote sensing images to be classified according to the number of spectral channels to obtain a spectral slice data set; extracting spectral space characteristics based on local image blocks in each spectral slice data to obtain a spectral space characteristic set corresponding to the hyperspectral remote sensing image to be classified; clustering all local image blocks based on a preset algorithm to obtain a plurality of initial clusters; merging the local image block clusters according to the semantic information to obtain a target cluster; calculating the confidence of each target cluster; each unmarked local image block in the target cluster is endowed with a pseudo mark set; and performing iterative correction on the pseudo mark set based on a spatial local constraint strategy to finish the classification of the hyperspectral remote sensing images. The invention can efficiently realize the image classification task under the condition that a small number of marked local image blocks exist or a large number of marked local image blocks cannot be obtained.

Description

Semi-supervised self-training hyperspectral remote sensing image classification method based on spatial correction
Technical Field
The invention belongs to the field of artificial intelligence, and particularly relates to a semi-supervised self-training hyperspectral remote sensing image classification method based on spatial correction.
Background
The hyperspectral remote sensing technology is a breakthrough of the optical remote sensing technology, and improves the whole image information chain of the optical remote sensing image from a black-white panchromatic image to a hyperspectral image through multispectral. Different from common remote sensing images, high-dimensional data of the hyperspectral remote sensing images have abundant spectrum and space information, so that the hyperspectral remote sensing images have high classification processing difficulty along with the problems of large data volume, much information redundancy and the like.
In order to overcome the problems of large data volume, more information redundancy and the like, the traditional method carries out dimension reduction operation on high-dimensional data and maps the high-dimensional data to a low-dimensional space so as to carry out image classification. Although the method solves the problems of high dimension and information redundancy, the original image space information is inevitably damaged by dimension reduction, and the accuracy of the classification result is reduced.
In order to fully utilize spectrum and spatial information without destroying original image spatial information, the prior art adopts a deep learning method to classify hyperspectral images. However, the existing hyperspectral image classification method based on deep learning depends on having enough labeled training sets, each pixel of the hyperspectral image needs one label, and the spatial resolution is low, so that the classification of each pixel is difficult to judge, the classification cannot be directly given to the pixel, the classification can be determined only by combining with on-site comparison, and the difficulty in labeling the image block is high. Therefore, the hyperspectral image classification method in the prior art needs high technical difficulty, time and labor consumption and large cost.
Disclosure of Invention
In order to solve the problems in the prior art, the invention provides a semi-supervised self-training hyperspectral remote sensing image classification method based on spatial correction. The technical problem to be solved by the invention is realized by the following technical scheme:
a semi-supervised self-training hyperspectral remote sensing image classification method based on spatial correction comprises the following steps: step 1: segmenting the hyperspectral remote sensing images to be classified according to the number of spectral channels to obtain a spectral slice data set, wherein the number of spectral slice data is the same as the number of spectral channels; step 2: extracting spectral space characteristics based on local image blocks in each spectral slice data to obtain a spectral space characteristic set corresponding to the hyperspectral remote sensing image to be classified; and step 3: clustering all local image blocks based on a preset algorithm to obtain a plurality of initial clusters; and 4, step 4: merging the local image block clusters according to the semantic information to obtain a target cluster; and 5: calculating the confidence of each target cluster; step 6: each unmarked local image block in the target cluster is endowed with a pseudo mark set; and 7: and (3) repeating the steps 2 to 7 based on a space local constraint strategy, and performing iterative correction on the pseudo mark set to finish the classification of the hyperspectral remote sensing images.
In one embodiment of the present invention, the step 2 comprises: step 2-1: taking each pixel in each spectral slice data as a center, and intercepting according to a preset size value to obtain a local image block set; the number of the local image blocks is the same as that of the pixels; step 2-2: and performing convolutional network coding on all local image blocks corresponding to each spectral slice data to obtain a spectral space feature set.
In an embodiment of the invention, the marked local image blocks and/or the unmarked local image blocks are comprised in the initial cluster.
The invention has the beneficial effects that:
the method can reduce the image blocks required by training, improve the classification precision and efficiency, and improve the classification efficiency.
The present invention will be described in further detail with reference to the accompanying drawings and examples.
Drawings
Fig. 1 is a schematic diagram of a semi-supervised self-training hyperspectral remote sensing image classification method based on spatial correction according to an embodiment of the invention.
Detailed Description
The present invention will be described in further detail with reference to specific examples, but the embodiments of the present invention are not limited thereto.
Examples
Referring to fig. 1, fig. 1 is a schematic diagram of a semi-supervised self-training hyperspectral remote sensing image classification method based on spatial correction according to an embodiment of the present invention, where the method includes:
step 1: and segmenting the hyperspectral remote sensing images to be classified according to the number of the spectral channels to obtain a spectral slice data set, wherein the number of the spectral slice data set is the same as the number of the spectral channels.
The method comprises the following steps of preprocessing a hyperspectral image, segmenting the hyperspectral remote sensing image according to the number of spectral channels in the preprocessing step, and inputting a new data set into a subsequent processing flow. Each segmented spectral slice data only contains information of one channel, and the method is more suitable for subsequent model training.
Step 2: and extracting spectral space characteristics based on the local image blocks in each spectral slice data to obtain a spectral space characteristic set corresponding to the hyperspectral remote sensing image to be classified.
Step 2 is a convolutional neural network characterization learning process. The method adopts the convolutional neural network model to extract the space-spectrum characteristics, can fully utilize the rich spectrum information and space information of the hyperspectral image, has strong learning capacity, and can efficiently mine the data distribution characteristics of the image in a high-dimensional space.
The local image blocks are three-dimensional image blocks in the invention.
Optionally, step 2 includes:
step 2-1: taking each pixel in each spectral slice data as a center, and intercepting according to a preset size value to obtain a local image block set; the number of the local image blocks is the same as that of the pixels.
Step 2-2: and performing convolutional network coding on all local image blocks corresponding to each spectral slice data to obtain a spectral space feature set.
Suppose that given a hyperspectral image dataset to be classified, P ═ P (P)1,P2,…,PN) N represents the total number of hyperspectral images, PnRepresenting the nth hyperspectral image. Firstly, each hyperspectral image is evenly divided into s spectral slices according to the number of spectral channels, wherein the number of spectral channels of each spectral slice is 1/s. Then, the same operation is done for each spectral slice: local image blocks with a certain size are selected by taking each pixel as a center, and if m pixels are provided, each spectral slice has m image blocks, and the whole image has m multiplied by s image blocks. Here, the size of each image block is e × e × t, e denotes the size of the length and width of the local image block, and t denotes the number of spectral channels of the spectral slice. The image blocks of all spectral slices of the nth hyperspectral image may be represented as:
Zn＝{zn,i∈RU|i＝1,2,…,m×s；U＝e×e×t}；
wherein Z isnE.z, which is defined as the set of image blocks of the entire dataset. XnThe spectral-spatial feature set is defined as the spectral-spatial feature set of the nth hyperspectral image after convolutional network coding. XnCan be expressed as:
Xn＝f(Zn；θ)；
where θ is defined as a parameter of the trained convolutional neural network, and f represents the forward process function. The final model outputs the generated spectral-spatial features in the second last layer, and outputs the class probabilities of the image blocks in the last layer, namely the softmax layer.
And step 3: and clustering all the local image blocks based on a preset algorithm to obtain a plurality of initial clusters.
And 3, performing semantic constraint clustering, wherein the local image block set corresponding to each spectral slice data can be subjected to independent clustering operation, and the whole image block set is clustered into a plurality of clusters with high internal confidence coefficients. Furthermore, the invention can utilize semantic information of a small number of marked image blocks to identify the semantic category of each cluster, and then combines the clusters with the same semantic category; similar to the calculation of the intersection between the sets, the clustering results of each spectral slice are compared, and the image blocks with the same semantic category are used as the result of the whole clustering.
The preset algorithm is selected by those skilled in the art according to the service requirement, and the present invention is not limited thereto. For example, the preset algorithm is a k-means algorithm.
For example, the nth spectral slice has m local image blocks Zn＝(zn1；zn2；…；znm),Xn＝(xn1；xn2；…；xnm) The spectral space feature set is obtained by extracting features of the convolution neural network from m local image blocks, wherein X isnCan be represented as Xn＝f(Zn(ii) a θ). We use the k-means algorithm to classify a set of spectral-spatial features XnClustering is performed to obtain k clusters with high internal confidence coefficient C ═ C1；C2；…；Ck}。
Optionally, the initial cluster includes marked local image blocks and/or unmarked local image blocks.
And 4, step 4: and merging the local image block clusters according to the semantic information to obtain a target cluster.
Optionally, the step 4 includes:
step 4-1: the marked local image blocks in each initial cluster are identified.
Step 4-2: when the initial cluster comprises the marked local image blocks, determining semantic information corresponding to the marked local image blocks as semantic tags of the initial cluster; or, when only unmarked local image blocks are included in the initial cluster, the semantic label of the initial cluster is determined to be unknown.
Step 4-3: and combining the initial clusters with the same semantic label into a target cluster.
And 4, introducing a semantic constraint strategy, calculating the confidence coefficient of the local image block with the semantic category by using the semantic label of the marked local image block as semantic information, and identifying the semantic category of each cluster. The number of clusters generated in the clustering operation process is dozens or even hundreds of times of the real categories, some clusters do not have corresponding semantic labels and semantic information to divide the clusters into unknown categories, the clusters with the semantic categories are merged together, and the number of the clusters is reduced. In the iterative optimization process, a termination condition is required for the clustering process, namely, the clustering process is terminated until the number of clusters is less than or equal to K, wherein the value of K is generally set to be two times to three times of the number of real semantic categories. Meanwhile, through semantic constraint, the labels in each cluster are based on semantic distribution and are used as next neural network training image blocks, better feature representation can be learned, and the next semantic constraint clustering is more efficient. And finally, comparing the clustering results of all the spectral slices, and taking the image blocks of which each spectral slice has the same semantic category as the final clustering result.
And 5: the confidence level of each target cluster is calculated.
Secondly, the confidence coefficient of the image block is calculated, the local image block is clustered into a plurality of clusters with high internal confidence coefficient by the clustering operation, and the information of the marked image block is required to be known when the confidence coefficient of the local image block is calculated. The labeled tile information is derived from the initial set of labeled tiles at the time of initial over-clustering, followed by a set of pseudo-labels generated at each iteration.
Optionally, step 5 includes:
step 5-1: calculating the number of marked local image blocks in each target cluster, which is expressed as:
Figure BDA0003386613310000061
step 5-2: calculating the average number of marked local image blocks in all target clusters, which is expressed as:
Figure BDA0003386613310000062
where K is the number of true semantic classes, NcIs the number of target clusters, Np,iIs the number of local image blocks contained in the ith cluster, and pt (i, j, q) indicates that the semantic label of the jth local image block in the ith cluster is q; when q is 0, it indicates that the semantic tag class is background, and pt (i, j, q) is equal to 0, or, equal to 1.
Step 5-3: calculating the purity of the marked local image block in the target cluster, and calculating the purity of the local image block with the semantic label f in the ith target cluster as follows:
Figure BDA0003386613310000071
wherein PUREi,maxIs PUREi,fMaximum value of (2).
When PUREi,fGreater than threshold values TH and Ni>NaveIt is shown that the local image blocks containing f types are the most in the cluster, and the cluster has higher confidence belonging to f types.
Step 6: each unmarked local image block in the target cluster is assigned a pseudo-mark set.
Step 6 is a space constraint process, step 6 can introduce a space local constraint strategy, the clustering result is calculated to generate a pseudo label set of the iteration, and the pseudo label set and the image block can be subsequently input into a CNN (Convolutional Neural network) network for next training to correct the pseudo label set.
Optionally, the step 6 includes:
step 6-1: determining a square field by taking each unmarked local image block in the target cluster as a center; the square neighborhood corresponds to a plurality of semantic tags, and each semantic tag corresponds to a weight value.
Step 6-2: and classifying all semantic labels.
Step 6-3: and scoring each type of semantic tag according to a preset weight and a weight value.
Step 6-4: and sequencing each type of semantic label according to the scoring result.
The highest scoring class indicates that the target cluster has a greater likelihood of belonging to the class, with weights associated with the two-dimensional euclidean distance, since labels closer to the target cluster should have greater decision-making power on the labels of the target cluster.
Step 6-5: and according to the sorting result, pseudo-marking sets are assigned to each unmarked local image block in the target cluster.
Examples of local decisions are: it is assumed that three marked local image blocks L1, L2, and L3 are present in the neighborhood of a square centered on the target cluster P, and a blank space indicates a local image block without a mark. And respectively calculating the scores of the three marked local image blocks according to the weight values, wherein the score of the ith marked local image block is calculated as:
Figure BDA0003386613310000081
wherein, the weight w1 is the weight of 8 local image blocks in the inner circle of the square neighborhood, and w2 is the weight of 16 local image blocks in the outer circle.
And 7: and (3) repeating the steps 2 to 7 based on a space local constraint strategy, and performing iterative correction on the pseudo mark set to finish the classification of the hyperspectral remote sensing images.
Step 7 is an iterative process based on self-training, and the step 7 can introduce a model training strategy based on self-training and continuously optimize a pseudo label set in an iterative mode to complete the classification of the hyperspectral remote sensing images.
The current unmarked local image block is endowed with a pseudo mark, which is used as a marked local image block in the next iteration process, and the semantic label is the endowed pseudo mark.
Optionally, step 7 includes:
and training to obtain a preset classifier through the marked local image blocks.
The invention firstly trains a classifier through the existing labeled image blocks, then utilizes the classifier to identify the unlabeled image blocks, each unlabeled image block has a certain confidence coefficient, selects the unlabeled image block with high confidence coefficient to give a distribution label, and then adds the label into the labeled image block set. The above process is continuously cycled until all unmarked tiles are added to the marked tile set.
Optionally, the iteratively correcting the pseudo mark set includes:
when the number of the target clusters is smaller than a preset threshold value, determining that the iterative correction of the pseudo mark set is completed; or all unmarked local image blocks are endowed with a pseudo mark set, and the iterative correction of the pseudo mark set is determined to be completed.
The invention can unify the whole supervision training process and the unsupervised clustering, the new mark local image block generated by each iteration can strengthen the training process of feature extraction, and the generated new feature further promotes the internal structure of the mining data of the clustering process to achieve a benign cycle, so that the iteration is carried out until the condition is met.
For example, each hyperspectral remote sensing image is firstly cut into 4 spectral slices, and then an original spectral space characteristic set is extracted from four spectral slice data
Figure BDA0003386613310000091
Figure BDA0003386613310000092
Wherein the set of spectral-spatial features represents the ith spectral slice data for the jth iteration processThe original spectral-spatial feature set is derived from a flattening of a local image block of size 5 x 5 centered on each sample.
Then, introducing an initial semantic mark set L, and respectively carrying out semantic constraint clustering on the original spectral space feature set of the four spectral slice data to obtain clustering results
Figure BDA0003386613310000093
Comparing the four clustering results, and taking the samples with the same semantic category as the final clustering result C0. Finally, for the clustering result C0Carrying out local decision to obtain a pseudo mark set Y0。
In each subsequent classification process, training the feature extraction network is performed first by using the pseudo label set obtained in the last iteration. Next, the features of each local image block are obtained by encoding the trained convolutional neural network, and the spectral space feature set of the four spectral slice data
Figure BDA0003386613310000094
Figure BDA0003386613310000095
Wherein
Figure BDA0003386613310000096
Represented is the spectral-spatial feature set of the 4 th iteration of the spectral slice data. The semantic constraint clustering and local decision operation are the same as the initial classification process, and the pseudo mark set Y of the time is obtainedtThe difference is that the labeled samples used by the semantic constraints and the samples of the training network are both pseudo-label sets generated by the last classification process. And iterating in the above way until the end condition is met, and obtaining a final classification result.
The invention mainly comprises two steps: feature representation and self-training. And performing local image block representation by using a convolution network in the feature representation process. In the self-training process, semantic information constraint of a small number of marked samples is utilized for clustering, pseudo labels with high confidence degree are gradually distributed to unmarked samples, a space constraint strategy is introduced, and the pseudo labels with errors are corrected and distributed by utilizing space consistency in the image. In the self-training process, local image blocks with high confidence level are gradually increased and added to corresponding semantic classes, so that semantic constraints are gradually enhanced. Meanwhile, the increase of the high-confidence-degree mark is also beneficial to the region consistency in the hyperspectral remote sensing image, so that the space constraint effect is highlighted, and the efficiency of hyperspectral image classification is improved.
In conclusion, the remote sensing image classification method can realize the effective remote sensing image classification task under the condition that only a small number of marked local image blocks exist or a large number of marked local image blocks cannot be obtained, can reduce the cost of image marking and has strong robustness.
The foregoing is a more detailed description of the invention in connection with specific preferred embodiments and it is not intended that the invention be limited to these specific details. For those skilled in the art to which the invention pertains, several simple deductions or substitutions can be made without departing from the spirit of the invention, and all shall be considered as belonging to the protection scope of the invention.

Claims (8)

1. A semi-supervised self-training hyperspectral remote sensing image classification method based on spatial correction is characterized by comprising the following steps:
step 1: segmenting the hyperspectral remote sensing images to be classified according to the number of spectral channels to obtain a spectral slice data set, wherein the number of spectral slice data is the same as the number of spectral channels;
step 2: extracting spectral space characteristics based on local image blocks in each spectral slice data to obtain a spectral space characteristic set corresponding to the hyperspectral remote sensing image to be classified;
and step 3: clustering all local image blocks based on a preset algorithm to obtain a plurality of initial clusters;
and 4, step 4: merging the local image block clusters according to the semantic information to obtain a target cluster;
and 5: calculating the confidence of each target cluster;
step 6: each unmarked local image block in the target cluster is endowed with a pseudo mark set;
and 7: and (3) repeating the steps 2 to 7 based on a space local constraint strategy, and performing iterative correction on the pseudo mark set to finish the classification of the hyperspectral remote sensing images.
2. The method of claim 1, wherein the step 2 comprises:
step 2-1: taking each pixel in each spectral slice data as a center, and intercepting according to a preset size value to obtain a local image block set; the number of the local image blocks is the same as that of the pixels;
step 2-2: and performing convolutional network coding on all local image blocks corresponding to each spectral slice data to obtain a spectral space feature set.
3. The method according to claim 1, wherein marked local image blocks and/or unmarked local image blocks are comprised in the initial cluster.
4. The method of claim 3, wherein the step 4 comprises:
step 4-1: identifying the marked local image blocks in each initial cluster;
step 4-2: when the initial cluster comprises the marked local image blocks, determining semantic information corresponding to the marked local image blocks as semantic tags of the initial cluster; or when the initial cluster only comprises unmarked local image blocks, determining the semantic labels of the initial cluster as unknown;
step 4-3: and combining the initial clusters with the same semantic label into a target cluster.
5. The method of claim 1, wherein the step 5 comprises:
step 5-1: calculating the number of marked local image blocks in each target cluster, which is expressed as:
Figure FDA0003386613300000021
step 5-2: calculating the average number of marked local image blocks in all target clusters, which is expressed as:
Figure FDA0003386613300000022
where K is the number of true semantic classes, NcIs the number of target clusters, Np,iIs the number of local image blocks contained in the ith cluster, and pt (i, j, q) indicates that the semantic label of the jth local image block in the ith cluster is q; when q is 0, it indicates that the semantic tag class is background, pt (i, j, q) is equal to 0, or, equal to 1;
step 5-3: calculating the purity of the marked local image block in the target cluster, and calculating the purity of the local image block with the semantic label f in the ith target cluster as follows:
Figure FDA0003386613300000023
wherein PUREi,maxIs PUREi,fMaximum value of (2).
6. The method of claim 1, wherein the step 6 comprises:
step 6-1: determining a square field by taking each unmarked local image block in the target cluster as a center; the square neighborhood corresponds to a plurality of semantic tags, and each semantic tag corresponds to a weight value;
step 6-2: classifying all semantic tags;
step 6-3: scoring each type of semantic tag according to a preset weight and a weight value;
step 6-4: sequencing each type of semantic tags according to the scoring result;
step 6-5: and according to the sorting result, pseudo-marking sets are assigned to each unmarked local image block in the target cluster.
7. The method of claim 1, wherein step 7 is preceded by:
and training to obtain a preset classifier through the marked local image blocks.
8. The method of claim 1, wherein iteratively correcting the set of pseudo labels comprises:
when the number of the target clusters is smaller than a preset threshold value, determining that the iterative correction of the pseudo mark set is completed; or all unmarked local image blocks are endowed with a pseudo mark set, and the iterative correction of the pseudo mark set is determined to be completed.
CN202111455830.XA 2021-12-01 2021-12-01 Semi-supervised self-training hyperspectral remote sensing image classification method based on spatial correction Pending CN114399687A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111455830.XA CN114399687A (en) 2021-12-01 2021-12-01 Semi-supervised self-training hyperspectral remote sensing image classification method based on spatial correction

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111455830.XA CN114399687A (en) 2021-12-01 2021-12-01 Semi-supervised self-training hyperspectral remote sensing image classification method based on spatial correction

Publications (1)

Publication Number Publication Date
CN114399687A true CN114399687A (en) 2022-04-26

Family

ID=81225638

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111455830.XA Pending CN114399687A (en) 2021-12-01 2021-12-01 Semi-supervised self-training hyperspectral remote sensing image classification method based on spatial correction

Country Status (1)

Country Link
CN (1) CN114399687A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117523409A (en) * 2023-11-10 2024-02-06 中国科学院空天信息创新研究院 Distributed collaborative incremental updating method and device based on model structure decoupling

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117523409A (en) * 2023-11-10 2024-02-06 中国科学院空天信息创新研究院 Distributed collaborative incremental updating method and device based on model structure decoupling
CN117523409B (en) * 2023-11-10 2024-06-07 中国科学院空天信息创新研究院 Distributed collaborative incremental updating method and device based on model structure decoupling

Similar Documents

Publication Publication Date Title
CN111695467B (en) Spatial spectrum full convolution hyperspectral image classification method based on super-pixel sample expansion
CN109165306B (en) Image retrieval method based on multitask Hash learning
CN113378632B (en) Pseudo-label optimization-based unsupervised domain adaptive pedestrian re-identification method
CN110334705B (en) Language identification method of scene text image combining global and local information
CN108830296B (en) Improved high-resolution remote sensing image classification method based on deep learning
CN110321967B (en) Image classification improvement method based on convolutional neural network
CN111797779A (en) Remote sensing image semantic segmentation method based on regional attention multi-scale feature fusion
CN111259828B (en) High-resolution remote sensing image multi-feature-based identification method
CN110647907B (en) Multi-label image classification algorithm using multi-layer classification and dictionary learning
CN111598004B (en) Progressive reinforcement self-learning unsupervised cross-domain pedestrian re-identification method
CN110414616B (en) Remote sensing image dictionary learning and classifying method utilizing spatial relationship
CN106874862B (en) Crowd counting method based on sub-model technology and semi-supervised learning
CN110443257B (en) Significance detection method based on active learning
CN105279519A (en) Remote sensing image water body extraction method and system based on cooperative training semi-supervised learning
CN110807485B (en) Method for fusing two-classification semantic segmentation maps into multi-classification semantic map based on high-resolution remote sensing image
CN108596240B (en) Image semantic segmentation method based on discriminant feature network
CN111967325A (en) Unsupervised cross-domain pedestrian re-identification method based on incremental optimization
CN111401380A (en) RGB-D image semantic segmentation method based on depth feature enhancement and edge optimization
CN115410059B (en) Remote sensing image part supervision change detection method and device based on contrast loss
CN116310350B (en) Urban scene semantic segmentation method based on graph convolution and semi-supervised learning network
CN115240024A (en) Method and system for segmenting extraterrestrial pictures by combining self-supervised learning and semi-supervised learning
CN113222072A (en) Lung X-ray image classification method based on K-means clustering and GAN
CN114677515A (en) Weak supervision semantic segmentation method based on inter-class similarity
CN114596463A (en) Image-based land parcel type classification method
CN114399687A (en) Semi-supervised self-training hyperspectral remote sensing image classification method based on spatial correction

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