High-resolution remote sensing sample labeling method based on topic model
By employing a semi-supervised learning method based on topic models, a weak classifier is trained using a small number of labeled samples. Combining traditional and deep learning features, this approach addresses the problem of remote sensing image annotation models' dependence on labeled data, achieving efficient and accurate remote sensing image annotation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- BEIJING DATA INTELLIGENCE INFORMATION TECH CO LTD
- Filing Date
- 2023-04-06
- Publication Date
- 2026-07-14
AI Technical Summary
Existing automatic annotation models for remote sensing images rely on a large amount of labeled data, resulting in high manual costs and poor generalization performance when data is insufficient, making it difficult to effectively utilize unlabeled data and limited prior knowledge.
A semi-supervised learning method based on topic models is adopted. A weak classifier is trained using a small number of labeled samples. By combining traditional features and deep learning features, the weak classifier model is constructed through annotation using a visual topic model and a weak classifier model to improve generalization ability.
It achieves efficient annotation of remote sensing images with limited labeled data, improves the accuracy and generalization ability of weak classifiers, and can classify remote sensing samples more accurately.
Smart Images
Figure CN116363460B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to remote sensing image annotation methods, and more particularly to a high-resolution remote sensing sample annotation method based on a topic model, belonging to the field of remote sensing image classification. Background Technology
[0002] Remote sensing technology, a new and comprehensive technology developed in the 1960s, is closely related to space science, electro-optics, computer science, geography, and other scientific technologies. It is one of the most powerful technical means for studying Earth's resources and environment. Automatic semantic annotation of remote sensing images is fundamental to the organization and indexing of these images. Especially with the explosive growth of remote sensing data in recent years, the sheer volume of high-resolution remote sensing images has made direct image retrieval and acquisition even more difficult, highlighting the significant importance of semantic annotation for high-resolution remote sensing images. However, existing manual annotation methods are time-consuming, labor-intensive, and not highly feasible, necessitating an efficient annotation method at the semantic level.
[0003] Current automatic annotation models can be mainly categorized into discriminative and generative annotation models. Discriminative annotation models are a type of supervised classification model that trains a classifier for each category of keywords. Based on the similarity between the visual content of images, keywords from labeled images are propagated to unlabeled images. Generative annotation models attempt to build a probabilistic association model between images and labeled words. This method assumes that the more similar the visual features of images, the greater the probability of labeling them with semantically similar keyword sequences.
[0004] However, many existing algorithms still rely on the quantity of labeled images, meaning that labeled models typically require a large amount of labeled data for training, which is too costly in terms of human intervention. When the amount of labeled data is insufficient, many models exhibit poor generalization performance. Furthermore, in specific tasks, obtaining labeled data is difficult, while unlabeled data is readily available. Therefore, fully utilizing a large amount of unlabeled data and limited prior knowledge, combined with the inherent characteristics of remote sensing images, is the main approach to improving the semantic annotation of remote sensing images. Summary of the Invention
[0005] To address the aforementioned technical problems, this invention provides a high-resolution remote sensing sample annotation method based on a topic model. Employing semi-supervised learning, it requires only a small subset of labeled samples for training and learning, thus assisting in training a weak classifier model with strong generalization ability. This method enables more accurate annotation of the samples to be labeled.
[0006] This invention provides a high-resolution remote sensing sample annotation method based on a topic model, the method comprising:
[0007] S1 acquires a high-resolution remote sensing sample set, including a training sample set and a sample set to be labeled, wherein the training sample set contains labeling information;
[0008] S2 uses traditional and deep learning methods to extract features from the training sample set, obtaining traditional and deep features of the training sample set;
[0009] S3 performs feature quantization on the traditional and deep features of the training sample set to obtain the bag-of-words representation of the training sample set;
[0010] S4 Construct a visual topic model by inputting the training sample set represented by bag of words into the visual topic model to obtain the topic distribution of the training sample set;
[0011] S5 utilizes the topic distribution and annotation information of the training sample set to construct and train a weak classifier model, wherein the weak classifier model includes at least two weak classifiers;
[0012] S6 inputs the sample set to be labeled into the weak classifier model to obtain the labeling information of the sample set to be labeled.
[0013] Further, step S2 includes:
[0014] Construct a deep convolutional neural network and use it to extract features from the training sample set to obtain deep features of the training sample set. The deep convolutional neural network can be VGG, ResNet, or AlexNet.
[0015] Traditional algorithms are used to extract features from the training sample set to obtain traditional features of the training sample set. These traditional algorithms include SIFT, Gray-Level Co-occurrence Matrix, or HOG feature extraction algorithm.
[0016] Further, step S3 includes:
[0017] Clustering methods are used to cluster the traditional features and deep features of the training sample set. Each cluster center is a visual word, and all visual words constitute a visual dictionary.
[0018] The traditional and deep features of the training sample set are mapped onto the visual dictionary, and histogram statistics are performed on the visual words of the training sample set to obtain the bag-of-words representation of the training sample set.
[0019] Further, step S4 includes:
[0020] A visual topic model is constructed by inputting the training sample set into the visual topic model and learning from the training sample set represented by bag of words to obtain the visual word topic distribution and potential topic distribution of the training sample set.
[0021] Furthermore, the weak classifier model includes T weak classifiers, and step S5 includes:
[0022] Train multiple weak classifiers based on the latent topic distribution and annotation information of each training sample, calculate the classification error rate of each weak classifier on the training sample set, and obtain the optimal weak classifier for this round of training.
[0023] Update the sample weights, and after T iterations, obtain T weak classifiers.
[0024] Further, step S6 includes:
[0025] Perform steps S2-S4 on the sample set to be labeled to obtain the topic distribution of the sample set to be labeled;
[0026] Learn the latent topic distribution of the unlabeled sample set based on the visual word topic distribution of the training sample set;
[0027] Input the latent topic distribution of the sample set to be labeled into the trained weak classifier model to obtain the word sequence to be labeled in the sample set;
[0028] Select the word with the highest probability to construct the set of labeled words for the sample set to be labeled, and obtain the labeling information of the sample set to be labeled.
[0029] The beneficial effects of this invention are as follows: This invention provides a high-resolution remote sensing sample annotation method based on a topic model. First, features are extracted from a labeled training sample set to obtain traditional features and deep features. These traditional and deep features, with their good properties, jointly represent the image, avoiding the problem of missing image information caused by a single feature. Then, the traditional and deep features are quantized to obtain a bag-of-words representation of the training sample set. A visual topic model is used to learn the topic distribution of the training sample set based on the bag-of-words representation. Multiple weak classifiers are trained based on the topic distribution of the training sample set, and these multiple weak classifiers form a weak classifier model. After the training process is completed, the trained weak classifier model is used to annotate the sample set to be labeled. This method utilizes a semi-supervised learning mode, using a small number of labeled training samples to assist in training multiple weak classifiers. During iterative training, the accuracy of the weak classifiers is improved, and the final weak classifier model has strong generalization ability, enabling more accurate classification of remote sensing samples and images. Attached Figure Description
[0030] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0031] Figure 1 This is a flowchart of a method according to an embodiment of the present invention;
[0032] Figure 2 This is a diagram representing the LDA graph model of an embodiment of the present invention. Detailed Implementation
[0033] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention are within the scope of protection of the present invention.
[0034] Please see Figure 1 This invention provides a high-resolution remote sensing sample annotation method based on a topic model, the method comprising:
[0035] S1 acquires a high-resolution remote sensing sample set, including a training sample set and a sample set to be labeled, wherein the training sample set contains labeling information;
[0036] S2 uses traditional and deep learning methods to extract features from the training sample set, obtaining traditional and deep features of the training sample set;
[0037] S3 performs feature quantization on the traditional and deep features of the training sample set to obtain the bag-of-words representation of the training sample set;
[0038] S4 Construct a visual topic model by inputting the training sample set represented by bag of words into the visual topic model to obtain the topic distribution of the training sample set;
[0039] S5 utilizes the topic distribution and annotation information of the training sample set to construct and train a weak classifier model, wherein the weak classifier model includes at least two weak classifiers;
[0040] S6 inputs the sample set to be labeled into the weak classifier model to obtain the labeling information of the sample set to be labeled.
[0041] High-resolution remote sensing sample sets are obtained, which are divided into a training sample set containing a small number of samples and a sample set to be labeled containing a large number of samples. The training sample set contains labeling information, while the sample set to be labeled does not contain labeling information.
[0042] Feature extraction is performed on the training sample set using both traditional algorithms and deep learning methods:
[0043] (1) The traditional algorithm is used to extract feature points from the training sample set. The resulting feature is the traditional feature. The traditional algorithm can be the SIFT algorithm, the gray-level co-occurrence matrix algorithm or the HOG feature extraction algorithm. The resulting features are SIFT features, texture features or HOG features, respectively. The traditional feature of this invention can be one or more of the above features.
[0044] (2) Construct a deep convolutional neural network and use the deep convolutional neural network to extract features from the training sample set to obtain the deep features of the training sample set. The deep convolutional neural network is VGG, ResNet, AlexNet or other deep learning networks.
[0045] Clustering methods are used to cluster the traditional features and deep features of the training sample set. The clustering method can be K-means clustering. Multiple cluster centers are obtained, and each cluster center represents a visual word. All visual words constitute a visual dictionary.
[0046] The traditional and deep features of the training sample set are quantized, and histogram statistics are performed on the visual words of each training sample set. The specific process of feature quantization is as follows: calculate the distance between two feature vectors, then find the cluster center in the visual dictionary that is closest to that feature vector, and map that feature vector to that visual word. By counting the occurrences of visual words in each training sample, a visual word histogram corresponding to that feature for a single training sample can be generated. Connecting the histograms of traditional and deep features yields the visual histogram for each training sample, i.e., the bag-of-words representation of the training sample set.
[0047] To construct a visual topic model, input the training sample set represented by a bag-of-words representation into the visual topic model to obtain the topic distribution of the training sample set. The visual topic model can be an LDA model, assuming that a training sample consists of N visual words. ,in It is the nth visual word in the training samples. The training sample set contains M training samples, denoted as... Each training sample is generated by a mixture of K potential topics z, and each topic is defined by parameters. A probability distribution generated on a visual dictionary. Parameters and parameters They are respectively subject to the parameter. The Dirichlet distribution, This indicates the mixing ratio of the distribution of image subjects. This represents the distribution of visual words across various themes, where w represents the visual words in the image. An LDA graphical model is as follows: Figure 2 As shown.
[0048] Input the training sample set into the LDA model to obtain the topic distribution of the training sample set. Given parameters... and Mixed-topic distribution The joint probability distribution of topic Z and training sample W is shown below:
[0049]
[0050] right By integrating and summing over z, we can obtain the marginal distribution of the training samples, as shown below:
[0051]
[0052] Finally, the marginal probabilities of each training sample are multiplied together to obtain the probability distribution of the training sample set:
[0053]
[0054] The parameters in the LDA model can be estimated using an empirical Bayesian method, which approximates the incalculable likelihood using a variational distribution. and maximize.
[0055] The LDA model learns topic distributions from a training sample set, including the latent topic distribution and the visual word topic distribution of the training sample set. The LDA model includes the independence assumption, meaning that the visual word topic distribution of the training sample set is independent of specific training samples; therefore, the visual word topic distribution also applies to the sample set to be labeled.
[0056] A weak classifier model is constructed based on the latent topic distribution and annotation information of the training sample set. The weak classifier model includes at least two weak classifiers, which can be random forest classifiers, decision tree classifiers, or other weak classifiers. The weak classifier model can be composed of the same type of weak classifier or different weak classifiers. In this embodiment, T random forest classifiers are used. The specific process is as follows: (1) Perform the first iteration, train multiple random forest classifiers based on the latent topic distribution and annotation information of each training sample, and then calculate the classification error rate of each random forest classifier on the training sample set to obtain the optimal random forest classifier for this round of training. (2) Update the sample weights. After T iterations, T random forest classifiers are obtained. The T random forest classifiers form the weak classifier model. Among them, the sample weights are the probability distribution of the training samples.
[0057] Feature extraction is performed on the unlabeled sample set to obtain traditional and deep features. These features are then quantized to obtain a bag-of-words representation. The latent topic distribution of the unlabeled sample set is learned based on the visual word topic distribution of the training sample set. This latent topic distribution is then input into a trained weak classifier model to obtain the unlabeled word sequence. Finally, the unlabeled word with the highest probability is selected to construct the labeled word set, thus obtaining the labeling information for the unlabeled sample set.
[0058] The beneficial effects of this invention are as follows: This invention provides a high-resolution remote sensing sample annotation method based on a topic model. First, features are extracted from a labeled training sample set to obtain traditional features and deep features. These traditional and deep features, with their good properties, jointly represent the image, avoiding the problem of missing image information caused by a single feature. Then, the traditional and deep features are quantized to obtain a bag-of-words representation of the training sample set. A visual topic model is used to learn the topic distribution of the training sample set based on the bag-of-words representation. Multiple weak classifiers are trained based on the topic distribution of the training sample set, and these multiple weak classifiers form a weak classifier model. After the training process is completed, the trained weak classifier model is used to annotate the sample set to be labeled. This method utilizes a semi-supervised learning mode, using a small number of labeled training samples to assist in training multiple weak classifiers. During iterative training, the accuracy of the weak classifiers is improved, and the final weak classifier model has strong generalization ability, enabling more accurate classification of remote sensing samples and images.
Claims
1. A high-resolution remote sensing sample annotation method based on a topic model, characterized in that, The method includes: S1 acquires a high-resolution remote sensing sample set, including a training sample set and a sample set to be labeled, wherein the training sample set contains labeling information; S2 uses traditional and deep learning methods to extract features from the training sample set, obtaining traditional and deep features of the training sample set; S3 performs feature quantization on the traditional and deep features of the training sample set to obtain the bag-of-words representation of the training sample set; S4 Construct a visual topic model by inputting the training sample set represented by bag of words into the visual topic model to obtain the topic distribution of the training sample set; S5 utilizes the topic distribution and annotation information of the training sample set to construct and train a weak classifier model, wherein the weak classifier model includes at least two weak classifiers; S6 Input the sample set to be labeled into the weak classifier model to obtain the labeling information of the sample set to be labeled; S2 includes: Construct a deep convolutional neural network and use it to extract features from the training sample set to obtain deep features of the training sample set. The deep convolutional neural network can be VGG, ResNet, or AlexNet. The traditional features of the training sample set are obtained by extracting features from the training sample set according to the traditional algorithm. The traditional algorithm is the SIFT algorithm, the gray-level co-occurrence matrix algorithm, or the HOG feature extraction algorithm. S3 includes: Clustering methods are used to cluster the traditional features and deep features of the training sample set. Each cluster center is a visual word, and all visual words constitute a visual dictionary. The traditional and deep features of the training sample set are mapped onto the visual dictionary, and histogram statistics are performed on the visual words of the training sample set to obtain the bag-of-words representation of the training sample set.
2. The high-resolution remote sensing sample annotation method based on topic model according to claim 1, characterized in that, Step S4 includes: A visual topic model is constructed by inputting the training sample set into the visual topic model and learning from the training sample set represented by bag of words to obtain the visual word topic distribution and potential topic distribution of the training sample set.
3. The high-resolution remote sensing sample annotation method based on topic model according to claim 2, characterized in that, The weak classifier model includes T weak classifiers. Step S5 includes: Train multiple weak classifiers based on the latent topic distribution and annotation information of each training sample, calculate the classification error rate of each weak classifier on the training sample set, and obtain the optimal weak classifier for this round of training. Update the sample weights, and after T iterations, obtain T weak classifiers.
4. The high-resolution remote sensing sample annotation method based on topic model according to claim 3, characterized in that, Step S6 includes: Perform steps S2-S4 on the sample set to be labeled to obtain the topic distribution of the sample set to be labeled; Learn the latent topic distribution of the unlabeled sample set based on the visual word topic distribution of the training sample set; Input the latent topic distribution of the sample set to be labeled into the trained weak classifier model to obtain the word sequence to be labeled in the sample set; Select the word with the highest probability to construct the set of labeled words for the sample set to be labeled, and obtain the labeling information of the sample set to be labeled.