A deep learning-based unmanned aerial vehicle polarization image reconstruction method and system

Abstract

This invention discloses a deep learning-based method and system for UAV polarization image reconstruction. By constructing a polarization image real degradation model that integrates Gaussian blur, downsampling, and Gaussian noise, the method simulates the resolution attenuation, blurring, and noise interference processes under non-ideal imaging systems, thereby improving the accuracy of the results. The method processes a predetermined number of UAV polarization images acquired using the polarization image real degradation model to generate a low-resolution polarization image dataset. A super-resolution reconstruction deep network integrating residual networks, dense networks, and sub-pixel convolutional layers is constructed. The super-resolution reconstruction deep network is trained and optimized using the low-resolution polarization image dataset through a loss function, resulting in an optimized super-resolution reconstruction deep network that can adapt to different target scenes. This improves the imaging pixels of small targets, ensures the integrity of polarization features, and thus enhances image reconstruction accuracy.

Description

Technical Field

[0002] This invention belongs to the field of UAV detection technology, specifically relating to a method and system for UAV polarization image reconstruction based on deep learning. Background Technology

[0004] In existing technologies, drone aerial images, due to their long shooting distance and wide field of view, generally suffer from problems such as low target pixel count, lack of detail information, and weak feature representation. Small targets are easily lost during downsampling. Traditional optical super-resolution methods only process intensity images and do not fully utilize multi-dimensional polarization information such as Stokes vectors, polarization degree, and polarization angle. After reconstruction, target edges are blurred and polarization features are distorted, failing to meet the requirements for high-precision detection and recognition of small targets. Existing super-resolution networks do not incorporate the real degradation process of drone imaging, resulting in poor model generalization and difficulty in adapting to complex environments such as backlighting and strong scattering.

[0005] A new method for reconstructing UAV polarization images is needed to solve the above-mentioned technical problems. Summary of the Invention

[0007] The purpose of this invention is to provide a deep learning-based method for reconstructing polarization images from unmanned aerial vehicles (UAVs), which addresses the technical problems of insufficient imaging pixels for small targets and incomplete polarization features in existing UAV image acquisition methods, resulting in low image reconstruction accuracy.

[0008] The present invention also aims to provide a deep learning-based UAV polarization image reconstruction system.

[0009] The technical solution of this invention to solve its technical problem is as follows:

[0010] A deep learning-based method for reconstructing polarization images from unmanned aerial vehicles (UAVs), characterized by the following steps:

[0011] S1: Construct a polarization image real degradation model that integrates Gaussian blur, downsampling, and Gaussian noise; process a set number of UAV polarization images collected using the polarization image real degradation model and generate a low-resolution polarization image dataset containing Stokes vectors, linear polarization degree, and polarization angle.

[0012] S2: Construct a super-resolution reconstruction deep network that integrates residual networks, dense networks, and sub-pixel convolutional layers;

[0013] S3: Using a low-resolution polarization image dataset, the super-resolution reconstruction deep network is trained and optimized through a loss function to obtain the optimized super-resolution reconstruction deep network.

[0014] S4: Input the UAV polarization image into the optimized super-resolution reconstruction depth network to obtain a high-resolution polarization image.

[0015] Preferably, the UAV polarization image is acquired by a focal plane polarization camera.

[0016] Preferably, the formula corresponding to the true degradation model of the polarization image is:

[0017] ;

[0018] in, This represents a true degradation model of polarization images. Represents a high-resolution polarization image. Represents a fuzzy matrix. This indicates a downsampling operation. This indicates the number of downsampling attempts, and N represents the noise level. Indicates the acquisition angle of the polarization image. .

[0019] The formula corresponding to the loss function is:

[0020] ;

[0021] ;

[0022] ;

[0023] in, Represents the loss function. Represents the content loss function. Represents the polarization loss function. Represents the polarization degree function of a polarization image. Represents the polarization angle function of a polarized image. The balance parameter represents the weighting of each loss term. Indicates trainable parameters, This represents the convolution operation in a feature extraction network.

[0024] Preferably, step S4 specifically involves: inputting a low-resolution polarization image into the network, extracting features in the low-resolution space, mapping the low-resolution features of the low-resolution image to the output of each layer through residual learning, having each layer interact with data through dense connections, and then performing an upsampling operation on the low-resolution features after feature extraction and convolving them through a convolutional layer to obtain a high-resolution polarization image.

[0025] A reconstruction system employing a deep learning-based UAV polarization image reconstruction method, characterized in that it comprises:

[0026] The polarization image real degradation module is used to construct a polarization image real degradation model that integrates Gaussian blur, downsampling and Gaussian noise. The polarization image real degradation model is used to process a set number of UAV polarization images and generate a low-resolution polarization image dataset containing Stokes vectors, linear polarization degree and polarization angle.

[0027] The super-resolution reconstruction deep network module is used to construct a super-resolution reconstruction deep network that integrates residual networks, dense networks, and sub-pixel convolutional layers.

[0028] The network training optimization and image processing module is used to train and optimize the super-resolution reconstruction depth network using a loss function on a low-resolution polarization image dataset to obtain an optimized super-resolution reconstruction depth network; the UAV polarization image is input into the optimized super-resolution reconstruction depth network to obtain a high-resolution polarization image.

[0029] The beneficial effects of this invention are as follows: By constructing a polarization image real degradation model that integrates Gaussian blur, downsampling, and Gaussian noise, the resolution attenuation, blurring, and noise interference processes under non-ideal imaging systems are simulated, thereby improving the accuracy of the results. The polarization image real degradation model processes a set number of UAV polarization images and generates a low-resolution polarization image dataset containing Stokes vectors, linear polarization degree, and polarization angle. A super-resolution reconstruction depth network integrating residual networks, dense networks, and sub-pixel convolutional layers is constructed. The super-resolution reconstruction depth network is trained and optimized using the low-resolution polarization image dataset through a loss function, resulting in an optimized super-resolution reconstruction depth network. The parameter configuration and local structure of the super-resolution reconstruction network at different scales are optimized using target objects of different scales, thereby enhancing the generalization ability of the super-resolution reconstruction depth network and enabling it to adapt to different target scenarios. This increases the number of pixels in small target imaging, ensures the integrity of polarization features, and thus improves image reconstruction accuracy. Attached Figure Description

[0031] Figure 1 This is a flowchart illustrating the method of the present invention;

[0032] Figure 2 This is a schematic diagram of the structure of the super-resolution reconstruction depth network for data processing in the method of this invention. Detailed Implementation

[0034] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0035] like Figure 1 As shown, this invention discloses a deep learning-based method for reconstructing polarization images of unmanned aerial vehicles (UAVs), characterized by the following steps:

[0036] S1: Construct a polarization image real degradation model that integrates Gaussian blur, downsampling, and Gaussian noise; process a set number of UAV polarization images acquired using the polarization image real degradation model to generate a low-resolution polarization image dataset containing Stokes vectors, linear polarization degree, and polarization angle; wherein, the UAV polarization images are acquired using a focal plane polarization camera. The formula corresponding to the polarization image real degradation model is:

[0037] ;

[0038] in, This represents a true degradation model of polarization images. Represents a high-resolution polarization image. Represents a fuzzy matrix. This indicates a downsampling operation. This indicates the number of downsampling attempts, and N represents the noise level. Indicates the acquisition angle of the polarization image. .

[0039] S2: Construct a super-resolution reconstruction deep network that integrates residual networks, dense networks, and sub-pixel convolutional layers;

[0040] S3: Using a low-resolution polarization image dataset, the super-resolution reconstruction deep network is trained and optimized using a loss function to obtain the optimized super-resolution reconstruction deep network; where the formula for the loss function is:

[0041] ;

[0042] ;

[0043] ;

[0044] in, Represents the loss function. Represents the content loss function. Represents the polarization loss function. Represents the polarization degree function of a polarization image. Represents the polarization angle function of a polarized image. The balance parameter represents the weighting of each loss term. Indicates trainable parameters, This represents the convolution operation in a feature extraction network.

[0045] S4: Input the UAV polarization image into the optimized super-resolution reconstruction depth network to obtain a high-resolution polarization image. For example... Figure 2As shown, specifically: a low-resolution polarization image is input into the network, features are extracted in the low-resolution space, and the low-resolution features of the low-resolution image are mapped to the output of each layer through residual learning. Each layer interacts with data through dense connections. The low-resolution features after feature extraction are upsampled and then convolved through convolutional layers to obtain a high-resolution polarization image.

[0046] A reconstruction system employing a deep learning-based UAV polarization image reconstruction method, characterized in that it comprises:

[0047] The polarization image real degradation module is used to construct a polarization image real degradation model that integrates Gaussian blur, downsampling and Gaussian noise. The polarization image real degradation model is used to process a set number of UAV polarization images and generate a low-resolution polarization image dataset containing Stokes vectors, linear polarization degree and polarization angle.

[0048] The super-resolution reconstruction deep network module is used to construct a super-resolution reconstruction deep network that integrates residual networks, dense networks, and sub-pixel convolutional layers.

[0049] The network training optimization and image processing module is used to train and optimize the super-resolution reconstruction depth network using a loss function on a low-resolution polarization image dataset to obtain an optimized super-resolution reconstruction depth network; the UAV polarization image is input into the optimized super-resolution reconstruction depth network to obtain a high-resolution polarization image.

[0050] Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

Claims

1. A deep learning-based method for reconstructing polarization images from unmanned aerial vehicles (UAVs), characterized in that, Includes the following steps: S1: Construct a polarization image real degradation model that integrates Gaussian blur, downsampling, and Gaussian noise; process a set number of UAV polarization images collected using the polarization image real degradation model and generate a low-resolution polarization image dataset containing Stokes vectors, linear polarization degree, and polarization angle. S2: Construct a super-resolution reconstruction deep network that integrates residual networks, dense networks, and sub-pixel convolutional layers; S3: Using a low-resolution polarization image dataset, the super-resolution reconstruction deep network is trained and optimized through a loss function to obtain the optimized super-resolution reconstruction deep network. S4: Input the UAV polarization image into the optimized super-resolution reconstruction depth network to obtain a high-resolution polarization image.

2. The UAV polarization image reconstruction method based on deep learning according to claim 1, characterized in that, The UAV polarization images were acquired using a split-focus plane polarization camera.

3. The UAV polarization image reconstruction method based on deep learning according to claim 1, characterized in that, The formula corresponding to the true degradation model of the polarization image is: ； in, This represents a true degradation model of polarization images. Represents a high-resolution polarization image. Represents a fuzzy matrix. This indicates a downsampling operation. This indicates the number of downsampling attempts, and N represents the noise level. Indicates the acquisition angle of the polarization image. .

4. The UAV polarization image reconstruction method based on deep learning according to claim 1, characterized in that, The formula corresponding to the loss function is: ； ； ； in, Represents the loss function. Represents the content loss function. Represents the polarization loss function. Represents the polarization degree function of a polarization image. Represents the polarization angle function of a polarized image. The balance parameter represents the weighting of each loss term. Indicates trainable parameters, This represents the convolution operation in a feature extraction network.

5. The UAV polarization image reconstruction method based on deep learning according to claim 1, characterized in that, Step S4 specifically involves: inputting a low-resolution polarization image into the network, extracting features in the low-resolution space, mapping the low-resolution features of the low-resolution image to the output of each layer through residual learning, and having each layer interact with data through dense connections. After upsampling the low-resolution features extracted from the feature image, the network performs convolution through a convolutional layer to obtain a high-resolution polarization image.

6. A reconstruction system employing the deep learning-based UAV polarization image reconstruction method according to any one of claims 1-5, characterized in that, include: The polarization image real degradation module is used to construct a polarization image real degradation model that integrates Gaussian blur, downsampling and Gaussian noise. The polarization image real degradation model is used to process a set number of UAV polarization images and generate a low-resolution polarization image dataset containing Stokes vectors, linear polarization degree and polarization angle. The super-resolution reconstruction deep network module is used to construct a super-resolution reconstruction deep network that integrates residual networks, dense networks, and sub-pixel convolutional layers. The network training optimization and image processing module is used to train and optimize the super-resolution reconstruction depth network using a loss function on a low-resolution polarization image dataset to obtain an optimized super-resolution reconstruction depth network; the UAV polarization image is input into the optimized super-resolution reconstruction depth network to obtain a high-resolution polarization image.

Images