A distribution alignment unmanned aerial vehicle remote sensing small sample generalization enhancement method and system

Abstract

The application discloses a distribution alignment unmanned aerial vehicle remote sensing small sample generalization enhancement method and system, belongs to the technical field of remote sensing detection, and is applied to unmanned aerial vehicle remote sensing image identification; the application realizes the method as follows: 1, optical detection is carried out on unmanned aerial vehicle remote sensing images, and remote sensing image data sets are acquired; further, the remote sensing image data sets are composed of remote sensing image labeled data sets and remote sensing image unlabeled data sets; 2, remote sensing image feature extraction is carried out on the remote sensing image data sets by using a deep neural network; 3, feature distribution statistics of unlabeled samples are used to calibrate feature distribution of labeled samples; 4, labeled samples are expanded by sampling from the calibrated distribution, and a robust classifier is trained. Compared with the prior art, the application improves the generalization ability of small sample labeled data in unmanned aerial vehicle remote sensing image identification.

Description

Technical Field

[0001] This invention relates to a distribution-aligned UAV remote sensing small-sample generalization enhancement method and system, belonging to the field of remote sensing technology, and applied to UAV remote sensing image recognition. Background Technology

[0002] Unmanned aerial vehicle (UAV) remote sensing image recognition technology is crucial in various fields such as agricultural monitoring, environmental protection, and disaster management. By analyzing remote sensing images captured by UAVs, it is possible to assess crop growth in real time, monitor environmental changes, and detect the impact of natural disasters, thereby providing data support for decision-making. In recent years, despite significant technological advancements, complex and variable geographical environments (such as terrain undulations, weather conditions, and changes in lighting) still pose challenges to the stability of recognition results.

[0003] Current UAV remote sensing image recognition primarily relies on deep learning methods. These methods extract image features through multi-layered network structures, progressively capturing information from low-level details to high-level semantics. Through downsampling and feature fusion, these models demonstrate excellent computational efficiency and recognition accuracy. However, these methods often require large amounts of labeled data for training to fully realize their performance.

[0004] To address the problem of insufficient data, few-shot learning techniques are gaining increasing attention in UAV remote sensing image recognition. Few-shot learning trains models with a very small number of labeled samples, enabling them to possess strong generalization capabilities and thus reducing reliance on large-scale datasets. Although deep learning techniques have made progress in this area, complex backgrounds, such as vegetation, buildings, or other environmental interference, still affect the model's effective extraction of target features. Especially under few-shot conditions, background interference has a more significant impact on the model's learning process, leading to reduced robustness and effectiveness of feature extraction. This challenge of background interference limits further improvements in UAV remote sensing image recognition performance, necessitating optimization through few-shot learning to enhance model stability and accuracy.

[0005] Therefore, improving the generalization ability of small sample labeled data in UAV remote sensing image recognition has become an urgent problem to be solved. Summary of the Invention

[0006] The purpose of this invention is to address the technical problem of improving the generalization ability of small sample labeled data in UAV remote sensing image recognition, and to propose a distribution-aligned UAV remote sensing small sample generalization enhancement method and system.

[0007] The objective of this invention is achieved through the following technical solution:

[0008] The present invention achieves the above objectives using the following technical solution.

[0009] A distribution-aligned UAV remote sensing small-sample generalization enhancement method includes the following steps:

[0010] Step 1: Perform optical detection on the UAV remote sensing images to obtain a remote sensing image dataset; the further remote sensing image dataset consists of a labeled remote sensing image dataset as shown in Equation (1) and an unlabeled remote sensing image dataset as shown in Equation (2);

[0011]

[0012] in, This represents the acquired labeled remote sensing image, y i Category labels representing remotely sensed images; The set of real numbers represents the remote sensing image data; w, h, and c represent the length, width, and number of channels of the input remote sensing image, respectively; and N represents the number of labeled remote sensing images.

[0013]

[0014] in, This represents the acquired unlabeled remote sensing images, where M represents the number of unlabeled remote sensing images;

[0015] Step 2: Use a deep neural network to extract remote sensing image features from the remote sensing image dataset in the manner shown in Equation (3);

[0016]

[0017] Where f represents the feature extraction performed on the acquired remote sensing image, Let n represent the set of real numbers. c The output channel dimension is represented by w, h, and c, which represent the length, width, and number of channels of the input remote sensing image, respectively.

[0018] Step 3: Calibrate the feature distribution extracted from the labeled samples using the feature distribution statistics extracted from the unlabeled samples;

[0019] Step 3.1: Perform Gaussian transformation on the remote sensing image features as shown in Equation (4);

[0020]

[0021] Where x' represents the transformed image feature, f(x i ) represents the input image features. γ represents the transformation magnitude; i represents the sample index;

[0022] Step 3.2: Perform K-means clustering on the features extracted from the unlabeled samples to form classification result clusters;

[0023] Step 3.3: Perform Gaussian distribution statistics on the classification result clusters;

[0024] Step 3.3.1: Perform mean statistics on the classification result clusters using the method shown in equation (5);

[0025]

[0026] Where, q′ j It is the j-th sample with Gaussian distribution characteristics in the N×M query set. The class distinction is represented by the method shown in equation (6) for the function of reference; c n This is represented as a cluster belonging to the nth category;

[0027]

[0028] Step 3.3.2: Perform covariance statistics on the nth cluster of the classification results using the method shown in Equation (7);

[0029]

[0030] Where T represents the matrix transpose operation;

[0031] Step 3.4: Match the feature distributions of labeled samples with those of unlabeled samples based on the similarity of their distribution means;

[0032] Calibration 3.5: Perform distribution calibration on the labeled feature distribution to obtain the mean and covariance of the calibrated distribution as shown in Equation (8);

[0033] μ′=βN(μ n )+(1-β)s′ i ,

[0034]

[0035] Where μ′ is the calibrated mean, ∑′ is the calibrated covariance, N is defined as the nearest neighbor function, β determines the weight of the Gaussian statistics of the nearest cluster in the Gaussian statistics of the calibrated distribution, and s′ i To support the Gaussian-like distribution of the set, α and β are hyperparameters, where α determines the degree of dispersion of the features sampled from the calibrated new class distribution;

[0036] Step 4: Expand labeled samples by sampling from the calibrated distribution to train a robust classifier.

[0037] Step 4.1: Expand the labeled samples to obtain the sampling dataset;

[0038] Step 4.1.1: Construct a mapping relationship between samples and labels, forming sample-label pairs (si ,y i );

[0039] Step 4.1.2: For label y i The labeled sample features are expanded to obtain the sampling dataset as shown in Equation (9);

[0040]

[0041] Step 4.2: Train the classifier using a fully connected network as shown in Equation (10);

[0042]

[0043] Where x′ is the sampled dataset x generated And samples randomly sampled from the original dataset x, This represents an indicator function, when y i When the label and class k are the same The result is 1 if there is inconsistency, and 0 if there is no consistency. N′ is the sampled dataset x. generated And the total number of samples in the original dataset x, where K is the total number of categories of the samples Pr(y) i |x;θ) is a classifier θ that classifies x as y. i The probability of;

[0044] A distribution-aligned UAV remote sensing small-sample generalization enhancement system is disclosed to implement the above-mentioned method. The system includes a remote sensing image acquisition device and a data processor.

[0045] Among them, the remote sensing image acquisition equipment is used to detect the target area and obtain a remote sensing image dataset; and uses it as input to the data processor.

[0046] The data processor includes an image feature extraction module, a distribution calibration module, a feature sampling module, and a category decision module;

[0047] The image feature extraction module is used to extract image features from the remote sensing image dataset and use them as input to the distribution calibration module;

[0048] The distribution calibration module is used to calibrate the feature distribution of the extracted small samples using the feature calibration method of Gaussian distribution statistics, thereby mitigating the distribution bias caused by sample noise in the small sample distribution and obtaining a robust feature distribution; it is then used as the input of the feature sampling module.

[0049] The feature sampling module is used to sample the calibrated distribution, and the sampled features and the features extracted from the real data are used together to train the classifier, which is then used as the input to the category decision module.

[0050] The category decision module is used to identify and classify the semantic features of the target and output the classification result of the target category.

[0051] Beneficial effects:

[0052] Compared with existing technologies, it has the following beneficial effects:

[0053] By employing a feature calibration method based on the Gaussian distribution statistics of unlabeled samples, distorted sample distributions can be effectively corrected. Compared to traditional remote sensing target recognition methods, this invention calibrates the sample distribution and expands the sample base by sampling from the calibrated distribution. This mitigates the negative impact of insufficient labeled samples on feature extraction in deep learning models, allowing the model to focus more on samples that conform to the true distribution. This improves the discriminative ability of extracted features and enhances the accuracy and robustness of target recognition. Furthermore, the data processor can utilize an embedded motherboard, enabling integration into various intelligent monitoring devices. It offers advantages such as simple implementation, low complexity, strong adaptability, and plug-and-play functionality. Attached Figure Description

[0054] Figure 1 This is a flowchart illustrating the present invention; Detailed Implementation

[0055] To better illustrate the purpose and advantages of this invention, the invention will be further described below with reference to the accompanying drawings and examples. It should be noted that the implementation of this invention is not limited to the following embodiments, and any modifications or alterations made to this invention will fall within the scope of protection of this invention.

[0056] Example

[0057] like Figure 1 As shown in this embodiment, a distribution-aligned UAV remote sensing small-sample generalization enhancement method is implemented using the following steps:

[0058] Step 1: Perform optical detection on the UAV remote sensing images to obtain a remote sensing image dataset; the further remote sensing image dataset consists of a labeled remote sensing image dataset as shown in Equation (1) and an unlabeled remote sensing image dataset as shown in Equation (2);

[0059]

[0060] in, This represents the acquired labeled remote sensing image, y i Category labels representing remotely sensed images; The set of real numbers represents the remote sensing image data; w, h, and c represent the length, width, and number of channels of the input remote sensing image, respectively; and N represents the number of labeled remote sensing images.

[0061]

[0062] in, This represents the acquired unlabeled remote sensing images, where M represents the number of unlabeled remote sensing images;

[0063] In this embodiment, the remote sensing image acquisition device performs optical detection on the target area, acquires remote sensing optical images, and constructs a remote sensing image dataset. The remote sensing image acquisition device is an airborne camera of an unmanned aerial vehicle.

[0064] Step 2: Use a deep neural network to extract remote sensing image features from the remote sensing image dataset in the manner shown in Equation (3);

[0065]

[0066] Where f represents the feature extraction performed on the acquired remote sensing image, Let n represent the set of real numbers. c The output channel dimension is represented by w, h, and c, which represent the length, width, and number of channels of the input remote sensing image, respectively.

[0067] In this embodiment, the image feature extraction module is a WideResNet28 pre-trained model, based on deep learning methods.

[0068] Step 3: Calibrate the feature distribution extracted from the labeled samples using the feature distribution statistics extracted from the unlabeled samples;

[0069] Step 3.1: Perform Gaussian transformation on the remote sensing image features as shown in Equation (4);

[0070]

[0071] Where x' represents the transformed image feature, f(x i ) represents the input image features. γ represents the transformation magnitude; i represents the sample index;

[0072] Step 3.2: Perform K-means clustering on the features extracted from the unlabeled samples to form classification result clusters;

[0073] Step 3.3: Perform Gaussian distribution statistics on the classification result clusters;

[0074] Step 3.3.1: Perform mean statistics on the classification result clusters using the method shown in equation (5);

[0075]

[0076] Where, q′ j It is the j-th sample with Gaussian distribution characteristics in the N×M query set. The class distinction is represented by the method shown in equation (6) for the function of reference; c n This is represented as a cluster belonging to the nth category;

[0077]

[0078] Step 3.3.2: Perform covariance statistics on the classification result clusters as shown in Equation (7);

[0079]

[0080] Where T represents the matrix transpose operation;

[0081] Step 3.4: Match the feature distributions of labeled samples with those of unlabeled samples based on the similarity of their distribution means;

[0082] Calibration 3.5: Perform distribution calibration on the labeled feature distribution to obtain the mean and covariance of the calibrated distribution as shown in Equation (8);

[0083] μ′=βN(μ n )+(1-β)s′ i ,

[0084]

[0085] Where N′ is defined as the nearest neighbor function, β determines the weight of the Gaussian statistics of the nearest cluster in the Gaussian statistics of the calibrated distribution, and s′ i To support the Gaussian-like distribution of the set, α and β are hyperparameters, where α determines the degree of dispersion of the features sampled from the calibrated new class distribution;

[0086] In this embodiment, the similarity of the distribution means is measured using Euclidean distance.

[0087] Step 4: Expand labeled samples by sampling from the calibrated distribution to train a robust classifier.

[0088] Step 4.1: Expand the labeled samples to obtain the sampling dataset;

[0089] Step 4.1.1: Construct a mapping relationship between samples and labels, forming sample-label pairs (s i ,y i );

[0090] Step 4.1.2: For label y i The labeled sample features are expanded to obtain the sampling dataset shown in Equation (9). Where x generated These are samples collected from the calibrated distribution;

[0091]

[0092] Step 4.2: Train the classifier using a fully connected classifier with the loss function shown in Equation (10);

[0093]

[0094] Where x′ is the sampled dataset x generated And samples randomly sampled from the original dataset x, This represents an indicator function, when y i When the label and class k are the same The result is 1 if there is inconsistency, and 0 if there is no consistency. N′ is the sampled dataset x. generated And the total number of samples in the original dataset x, where K is the total number of categories of the samples Pr(y) i |x;θ) is a classifier θ that classifies x as y. i The probability of;

[0095] In this embodiment, the sampling method directly adopts the Gaussian distribution sampling method with calibrated mean and variance. The classifier can also be replaced with other classifiers. The loss function adopts the commonly used classification loss function, cross-entropy loss.

[0096] A distribution-aligned UAV remote sensing small-sample generalization enhancement system is disclosed to implement the above-mentioned method. The system includes a remote sensing image acquisition device and a data processor.

[0097] Among them, the remote sensing image acquisition equipment is used to detect the target area and obtain a remote sensing image dataset; which is then used as input to the data processor.

[0098] The data processor includes an image feature extraction module, a distribution calibration module, a feature sampling module, and a category decision module;

[0099] The image feature extraction module is used to extract image features from the remote sensing image dataset and use them as input to the distribution calibration module;

[0100] The distribution calibration module is used to calibrate the feature distribution of the extracted small samples using the feature calibration method of Gaussian distribution statistics, thereby mitigating the distribution bias caused by sample noise in the small sample distribution and obtaining a robust feature distribution; it is then used as the input of the feature sampling module.

[0101] The feature sampling module is used to sample the calibrated distribution, and the sampled features and the features extracted from the real data are used together to train the classifier, which is then used as the input to the category decision module.

[0102] The category decision module is used to identify and classify the semantic features of the target and output the classification result of the target category.

[0103] In this embodiment, the remote sensing image acquisition device is an airborne camera of a drone. The data processor is a computer or an embedded motherboard. The image feature extraction module is a WideResNet28 pre-trained model based on deep learning methods.

[0104] The above detailed description further illustrates the purpose, technical solution, and beneficial effects of the invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A distribution-aligned UAV remote sensing small-sample generalization enhancement method, characterized in that: Includes the following steps, Step 1: Perform optical detection on the UAV remote sensing images to obtain a remote sensing image dataset; the further remote sensing image dataset consists of a labeled remote sensing image dataset as shown in Equation (1) and an unlabeled remote sensing image dataset as shown in Equation (2); （1） in, This refers to the acquired labeled remote sensing images. Category labels representing remotely sensed images; The set of real numbers representing remote sensing image data; These represent the length, width, and number of channels of the input remote sensing image, respectively. This indicates the number of labeled remote sensing images; （2） in, This represents the acquired unlabeled remote sensing image. Indicates the number of unlabeled remote sensing images; Step 2: Use a deep neural network to extract remote sensing image features from the remote sensing image dataset in the manner shown in Equation (3); （3） in, This indicates that feature extraction is performed on the acquired remote sensing images. Represents the set of real numbers. Indicates the output channel dimension. These represent the length, width, and number of channels of the input remote sensing image, respectively. Step 3: Calibrate the feature distribution extracted from the labeled samples using the feature distribution statistics extracted from the unlabeled samples; Step 3.1: Perform Gaussian transformation on the remote sensing image features as shown in Equation (4); （4） in, This represents the image features after transformation. Represents the features of the input image; This is expressed as the conversion range; Represented as a sample index; Step 3.2: Perform K-means clustering on the features extracted from the unlabeled samples to form classification result clusters; Step 3.3: Perform Gaussian distribution statistics on the classification result clusters; Step 3.4: Match the feature distributions of labeled samples with those of unlabeled samples based on the similarity of their distribution means; Calibration 3.5: Perform distribution calibration on the labeled feature distribution to obtain the mean and covariance of the calibrated distribution as shown in Equation (8); , （8） in, This is the calibrated mean. For the calibrated covariance, Defined as the nearest neighbor function, This determines the weight of the Gaussian statistics of the nearest cluster in the Gaussian statistics of the calibration distribution. To support the Gaussian-like distribution characteristics of the set, and It's a hyperparameter. This determines the degree of dispersion of features sampled from the calibrated new class distribution; Step 4: Expand labeled samples by sampling from the calibrated distribution to train a robust classifier.

2. The method for distribution-aligned UAV remote sensing small-sample generalization enhancement as described in claim 1, characterized in that: Step 3.3 is implemented as follows: Step 3.3.1: Perform mean statistics on the classification result clusters using the method shown in equation (5); , （5） in, yes The query for the characteristics of a Gaussian distribution in a set. One sample, The class distinction is represented by the method shown in equation (6) for the function of reference; Represented as the first Clusters of various categories; （6） Step 3.3.2: Perform covariance statistics on the nth cluster of the classification results using the method shown in equation (7); （7） in, This is the transpose operation for a matrix.

3. The method for distribution-aligned UAV remote sensing small-sample generalization enhancement as described in claim 1, characterized in that: Step 4 is implemented as follows: Step 4.1: Expand the labeled samples to obtain the sampling dataset; Step 4.2: Train the classifier using a fully connected network as shown in Equation (10); （10） in, For the sampled dataset And samples randomly sampled from the original dataset x, This represents an indicator function, when When the tag and class k are the same The result is 1, and 0 if they are inconsistent. For the sampled dataset The total number of samples in the original dataset x, where K is the total number of classes in the samples. For classifier Classify x as The probability of.

4. The method for distribution-aligned UAV remote sensing small-sample generalization enhancement as described in claim 3, characterized in that: Step 4.1 is implemented as follows: Step 4.1.1: Construct a mapping relationship between samples and labels to form sample-label pairs. ; Step 4.1.2: For the label... The labeled sample features are expanded to obtain the sampling dataset shown in equation (9): (9).

5. A distribution-aligned UAV remote sensing small-sample generalization enhancement system that implements the method described in claim 1, characterized in that: Includes remote sensing image acquisition equipment and data processors; Remote sensing image acquisition equipment is used to detect target areas and acquire remote sensing image datasets, which are then used as input to data processors.

6. The distributed-aligned UAV remote sensing small-sample generalization enhancement system as described in claim 5, characterized in that: The data processor includes an image feature extraction module, a distribution calibration module, a feature sampling module, and a category decision module; The image feature extraction module is used to extract image features from the remote sensing image dataset and use them as input to the distribution calibration module; The distribution calibration module is used to calibrate the feature distribution of extracted small samples using a feature calibration method based on Gaussian distribution statistics, mitigating distribution bias caused by sample noise in the small sample distribution, thereby obtaining a robust feature distribution; It serves as the input to the feature sampling module; The feature sampling module is used to sample the calibrated distribution, and the sampled features and the features extracted from the real data are used together to train the classifier, which is then used as the input to the category decision module. The category decision module is used to identify and classify the semantic features of the target and output the classification result of the target category.

7. The distributed-aligned UAV remote sensing small-sample generalization enhancement system as described in claim 5, characterized in that: The remote sensing image acquisition equipment is an airborne camera on a drone.

8. The distributed-aligned UAV remote sensing small-sample generalization enhancement system as described in claim 5, characterized in that: The data processor is a computer or an embedded motherboard.

9. The distributed-aligned UAV remote sensing small-sample generalization enhancement system as described in claim 6, characterized in that: The image feature extraction module is a WideResNet28 pre-trained model.

Images