Dark light field remote sensing image enhancement method and device based on data driving and medium

Abstract

The invention discloses a dim light field remote sensing image enhancement method and device based on data driving and a medium, and the method comprises the steps: S1, constructing an optical remote sensing image enhancement data set in a dim light field, and obtaining a collected to-be-enhanced remote sensing image; s2, constructing a decomposition network based on a Retinex theory, and training the decomposition network by using the enhanced data set to obtain an image decomposition model; inputting the to-be-enhanced remote sensing image into a trained image decomposition model, and performing decomposition to obtain a reflection feature map and an illumination feature map; s3, constructing a feature enhancement module based on frequency domain Fourier convolution, inputting the reflection feature map and the illumination feature map into the feature enhancement module, and performing frequency domain and space domain combined feature extraction and enhancement to obtain an enhanced feature map; and S4, outputting the enhanced remote sensing image. The method can improve the brightness, contrast and details of the image, effectively suppress the noise, keep the color natural, and improve the downstream visual task performance.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically to a data-driven method, device, and medium for enhancing remote sensing images in low-light fields. Background Technology

[0002] Optical remote sensing imagery is a crucial data carrier for acquiring surface information through Earth observation platforms such as satellites and drones. Its quality directly affects the accuracy and reliability of subsequent applications such as feature identification, environmental monitoring, and resource surveys. However, under weak light field imaging conditions such as at night, on cloudy days, in shaded areas, or at low solar altitude angles, the light signal received by the sensor is significantly weakened, resulting in images that generally suffer from low brightness, prominent noise, and poor contrast. The overall images are dark, with blurred details and distorted colors, making it difficult to distinguish features from the background and introducing a large amount of noise interference. This severely reduces the interpretability and analytical value of the images, limiting the practical application of remote sensing technology in complex lighting scenarios.

[0003] Currently, enhancement methods for remote sensing images with low light fields are mainly divided into enhancement methods based on traditional algorithms and intelligent enhancement methods based on deep learning. Traditional enhancement methods mostly rely on image grayscale transformation, frequency domain filtering, or physical model derivation. For example, histogram equalization enhances contrast by redistributing pixel grayscale, but it easily amplifies noise; Retinex theory improves visual perception by separating illumination and reflection components, but it is sensitive to noise and prone to color shift; gamma correction can adjust overall brightness, but its improvement in local contrast is limited; wavelet transform enhances details through multi-scale analysis, but its effect depends on manually designed filters and thresholds, resulting in insufficient adaptability. These methods usually do not rely on a large amount of training data, have good interpretability and low computational cost, but they have significant limitations in handling extreme low light, complex noise, and detail restoration, often struggling to balance brightness enhancement, color fidelity, and structure preservation.

[0004] With the development of deep learning technology, data-driven intelligent enhancement methods have gradually become a research hotspot. These methods treat image enhancement as an end-to-end mapping learning problem, automatically extracting features from samples and establishing complex nonlinear transformations from low-quality to high-quality images. Common technical approaches include convolutional neural networks (CNNs), which have strong feature extraction capabilities but are prone to overfitting or unnatural output under extreme lighting conditions; generative adversarial networks (GANs) can generate visually realistic images, but training is unstable, prone to artifacts, and requires high data quality; autoencoders have simple structures and are suitable for unsupervised learning, but are weak in recovering high-frequency details; models based on self-attention mechanisms can capture long-range dependencies and improve enhancement effects, but are accompanied by high computational complexity and large memory consumption, making them difficult to apply to real-time processing of high-resolution remote sensing images. Although deep learning methods excel in feature modeling, they are still limited by their high dependence on large-scale labeled data, insufficient generalization ability under different imaging conditions, and the balance between model computational efficiency and deployment feasibility.

[0005] In summary, existing enhancement methods often fail to effectively improve brightness and contrast while maintaining natural color consistency, suppressing noise, enhancing detail and texture, and balancing algorithm generalization ability and execution efficiency when dealing with remote sensing images in low-light conditions. Therefore, there is an urgent need to develop a remote sensing image enhancement technique that can adapt to complex lighting variations, achieve a good balance between detail enhancement and noise suppression, and possesses high practicality and adaptability, in order to promote the development of remote sensing image analysis and applications in low-light environments. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies, one of the objectives of this invention is to provide a data-driven method for enhancing remote sensing images in low-light fields, which can improve image brightness, contrast and detail while effectively suppressing noise and maintaining natural colors, thereby improving the performance of downstream visual tasks.

[0007] The second objective of this invention is to provide an electronic device that can improve image brightness, contrast and detail while effectively suppressing noise and maintaining natural colors, thereby improving the performance of downstream visual tasks.

[0008] The third objective of this invention is to provide a storage medium that can improve image brightness, contrast and detail while effectively suppressing noise and maintaining natural colors, thereby improving the performance of downstream visual tasks.

[0009] To achieve one of the objectives of this invention, the following solution is adopted: Data-driven methods for enhancing remote sensing images with weak light fields include: Step S1: Based on the remote sensing application scenario, set the light source angle and brightness conditions, construct an optical remote sensing image enhancement dataset under dim light field, and acquire the collected remote sensing images to be enhanced. Step S2: Construct a decomposition network based on Retinex theory, train the decomposition network using the augmentation dataset to obtain an image decomposition model; input the remote sensing image to be augmented into the trained image decomposition model to decompose it into a reflectance feature map and an illumination feature map; Step S3: Construct a feature enhancement module based on frequency domain Fourier convolution. Input the reflection feature map and the illumination feature map into the feature enhancement module to perform joint feature extraction and enhancement in the frequency domain and spatial domain to obtain the enhanced feature map. Step S4: Construct a reconstruction network model based on an encoder-decoder architecture, input the enhanced post-feature map into the reconstruction network model, extract global features through the encoder, and restore spatial resolution by combining multi-scale features through the decoder, and output the enhanced remote sensing image.

[0010] Furthermore, the construction of the optical remote sensing image enhancement dataset under low light field in step S1 specifically includes: Step S1.1: Acquire image data under different lighting conditions using a multi-platform remote sensing imaging device, and adjust the sensitivity to control the imaging brightness range within a preset range; Step S1.2: Collect sample images of various land cover types and record the lighting conditions and shooting parameters during imaging; Step S1.3: Perform at least one data augmentation operation on the acquired image, including rotation, flipping, brightness perturbation, and noise simulation.

[0011] Furthermore, the decomposition network based on Retinex theory in step S2 has a structure that includes an input preprocessing layer, an initialization convolutional layer, a feature extraction layer, an output convolutional layer, a Sigmoid activation layer, and a separation adjustment layer connected in sequence.

[0012] Furthermore, the input preprocessing layer is used to concatenate the image tensor of shape (B,H,W,3) and the illumination guidance tensor of shape (B,H,W,1) in the channel dimension, and adjust the dimension to (B,4,H,W). And / or, the kernel size of the initial convolutional layer is 9×9, the padding is 4, the input channel is 4, and the output channel is 64; And / or, the feature extraction layer consists of multiple consecutive convolutional layers with a kernel size of 3×3; And / or, the output convolutional layer has a 3×3 kernel, 64 input channels, and 4 output channels; And / or, the Sigmoid activation layer is used to normalize the output feature values ​​to the range [0,1]; And / or, the separation adjustment layer is used to separate the output features of the 4 channels into a reflection feature map of shape (B,H,W,3) and an illumination feature map of shape (B,H,W,1).

[0013] Furthermore, the feature enhancement module based on frequency domain Fourier convolution in step S3 includes an input segmentation layer, a local feature extraction layer, a feature concatenation layer, a frequency domain fusion unit, and an activation and normalization layer connected in sequence.

[0014] Furthermore, the frequency domain fusion unit is used to perform the following processing on the input features: Perform a Fourier transform on the feature map and separate the transform result into the real part and the imaginary part; The real and imaginary parts are subjected to dimensionality adjustment and convolution operations respectively to extract frequency domain features; Perform an inverse Fourier transform on the processed frequency domain features to restore them to the spatial domain; The recovered spatial features are jointly enhanced with the original spatial input features through a fusion module, and then output through residual connections.

[0015] Furthermore, the reconstruction network model based on the encoder-decoder architecture in step S4 includes: The encoder is used to downsample the input features layer by layer to generate multi-scale encoded features; The decoder, corresponding to each level of the encoder, is used to gradually restore spatial resolution through upsampling operations and to fuse encoded features from the corresponding encoder levels using skip connections, ultimately outputting an enhanced remote sensing image.

[0016] Furthermore, the method also includes a network training step, specifically: The augmented dataset is divided into a training set and a test set; The first stage of training the decomposition network uses the loss function as the difference between the decomposed reflection feature map and illumination feature map and the corresponding image feature map under normal illumination. The second stage trains the reconstruction network model, and the loss function is the difference between the enhanced image output by the reconstruction network model and the corresponding normal illumination image.

[0017] To achieve the second objective of this invention, the following solution is adopted: An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. The computer program is a low-light-field remote sensing image enhancement program. When the processor executes the low-light-field remote sensing image enhancement program, it implements the steps of the data-driven low-light-field remote sensing image enhancement method as described in one of the objectives of this invention.

[0018] To achieve the third objective of this invention, the following solution is adopted: A storage medium, which is a computer-readable storage medium, stores a computer program thereon, the computer program being a low-light-field remote sensing image enhancement program, which, when executed by a processor, implements the steps of the data-driven low-light-field remote sensing image enhancement method as described in one of the objectives of this invention.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention combines data-driven deep learning mechanisms with Retinex theory, achieving systematic enhancement of low-light remote sensing images through a cascaded "decomposition-enhancement-reconstruction" processing flow. First, a pre-trained decomposition network accurately separates the image into illumination and reflection components, overcoming the shortcomings of traditional Retinex methods such as blurred decomposition under extreme low light and sensitivity to noise. Then, a feature enhancement module based on frequency-domain Fourier convolution performs joint spatial and frequency-domain modeling and enhancement of the separated features, effectively preserving and enhancing high-frequency details such as object edges and textures while suppressing noise and illumination disturbances. Finally, a reconstruction network with an encoder-decoder architecture fuses multi-scale features to output a visually high-quality, naturally colored enhanced image. This invention significantly improves the brightness, contrast, and detail representation of low-light images, demonstrating excellent enhancement effects.

[0020] 2. This invention possesses excellent robustness and generalization ability. Through data-driven training based on the constructed dataset, this invention can adapt to the enhancement needs in various low-light scenarios. Even when faced with complex imaging conditions such as illumination attenuation and noise interference, this invention can stably output high-quality enhancement results.

[0021] 3. This invention improves image quality while also considering processing efficiency and practicality. The network structure adopted, especially the feature enhancement module and encoder-decoder framework, enables real-time or near-real-time processing of high-resolution remote sensing images (e.g., ≥30 FPS for 512×512 resolution) on hardware platforms such as GPUs. This meets the application requirements of remote sensing platforms such as satellites and UAVs for low-latency, high-throughput image processing and supports end-to-end deployment and application.

[0022] 4. This invention provides a higher-quality data foundation for downstream remote sensing vision tasks (such as land cover recognition, change detection, and target classification) through high-quality image enhancement. The improvements in brightness consistency, color fidelity, and detail clarity of the enhanced images can effectively reduce the false detection rate and false negative rate in subsequent high-level vision tasks, thereby significantly improving the accuracy and reliability of remote sensing information extraction. Attached Figure Description

[0023] Figure 1 This is a flowchart of a data-driven remote sensing image enhancement method for low-light fields according to an embodiment of the present invention; Figure 2 This is a diagram of a data-driven optical remote sensing image enhancement network structure in one embodiment of the present invention; Figure 3 This is a structural diagram of a frequency domain fusion unit in one embodiment of the present invention; Figure 4 This is a structural diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0024] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments. It should be noted that, without conflict, the various embodiments or technical features described below can be arbitrarily combined to form new embodiments.

[0025] Example 1 This invention provides a data-driven method for enhancing remote sensing images in low-light conditions. Combining Retinex theory with a data-driven deep network learning mechanism, this invention innovates in network structure design, enabling efficient brightness restoration and detail enhancement of optical remote sensing images in low-light environments. This significantly improves the accuracy and stability of downstream visual tasks (such as ground feature recognition and change detection). Compared to traditional image enhancement methods, this invention demonstrates superior performance in brightness restoration, color consistency, and noise suppression for low-light remote sensing images, while also possessing high computational efficiency and model generalization ability, meeting the practical needs of remote sensing image enhancement and analysis under complex lighting conditions.

[0026] like Figures 1 to 3 As shown, the data-driven remote sensing image enhancement method for weak light fields according to an embodiment of the present invention includes: Step S1: Based on the remote sensing application scenario, set the light source angle and brightness conditions, construct an optical remote sensing image enhancement dataset under low light field, and acquire the collected remote sensing images to be enhanced.

[0027] In this embodiment, step S1 specifically includes: Step S1.1: Acquire remote sensing images under low-light conditions using multi-source remote sensing imaging equipment (such as satellites or UAVs equipped with optical cameras) to ensure clear images can still be obtained in dim light environments. By adjusting parameters such as ISO and aperture, the imaging brightness range is controlled within 50–80 lux to cover light variations under different meteorological conditions.

[0028] Step S1.2: Collect samples from typical landform areas such as urban building areas, roads, water bodies, farmland, and woodlands to ensure the diversity and representativeness of landform types. Collect no fewer than 1000 samples for each type of landform and record illumination and sensor parameters.

[0029] Step S1.3: Perform data augmentation operations such as rotation, flipping, and brightness perturbation on the acquired images to expand the dataset and improve the robustness of training.

[0030] Step S2: Construct a decomposition network based on Retinex theory, train the decomposition network using the augmentation dataset to obtain an image decomposition model; input the remote sensing image to be augmented into the trained image decomposition model to decompose it into a reflection feature map and an illumination feature map.

[0031] Specifically, the decomposition network based on Retinex theory consists of cascaded convolutional layers and channel stitching modules, achieving end-to-end image decomposition. This network decomposes the input low-light remote sensing image into reflectance and illumination feature maps through spatial-channel feature transformation.

[0032] In this embodiment, step S2 specifically includes: Step S2.1: Decompose the network into an input preprocessing layer, an initialization convolutional layer, a feature extraction convolutional layer, an output convolutional layer, a sigmoid activation layer, and a separation adjustment layer.

[0033] Step S2.2: The input preprocessing layer concatenates the image tensor of shape (B,H,W,3) with the illumination guidance tensor of shape (B,H,W,1), and adjusts the dimensions to (B,4,H,W).

[0034] Step S2.3: Initialize the convolutional layer kernel size to 9×9, padding to 4, input channels to 4, and output channels to 64.

[0035] Step S2.4: The feature extraction layer consists of 5 identical convolutional layers with 3×3 kernels, 1 padding layer, and the number of channels remains unchanged.

[0036] Step S2.5: The output convolutional layer has a 3×3 kernel, 64 input channels, and 4 output channels.

[0037] Step S2.6: The Sigmoid activation layer normalizes the output features to the range [0,1].

[0038] Step S2.7: Separate adjustment layer separates the 4-channel output into a reflection feature map of (B,H,W,3) and an illumination feature map of (B,H,W,1).

[0039] Step S3: Construct a feature enhancement module based on frequency domain Fourier convolution. Input the reflection feature map and the illumination feature map into the feature enhancement module to perform joint feature extraction and enhancement in the frequency domain and spatial domain, and obtain the enhanced feature map.

[0040] Specifically, the real and imaginary parts of the feature map are separated, and then dimensionality adjustment and convolution operations are performed to extract frequency domain features. The feature map is then restored to the spatial domain using an inverse Fourier transform. Subsequently, the frequency domain features and spatial domain features are jointly enhanced using a feature enhancement module. Finally, the enhanced feature map is output through residual linking.

[0041] The frequency domain Fourier convolution module (i.e., the feature enhancement module based on frequency domain Fourier convolution) has the functions of frequency-spatial joint modeling and dynamic channel adaptation. It consists of an input segmentation layer, a local feature extraction layer, a feature concatenation layer, a frequency domain fusion unit, and an activation normalization layer.

[0042] In this embodiment, step S3 specifically includes: Step S3.1: The input segmentation layer dynamically adjusts the number of channels, divides the input feature map evenly along the channel dimension, and performs parallel local feature extraction.

[0043] Step S3.2: The local feature extraction layer uses 1×1 convolutions to capture feature associations in the channel direction.

[0044] Step S3.3: The feature splicing layer re-splices the two parts of features along the channel dimension, fusing local feature information.

[0045] Step S3.4: The frequency domain fusion unit decomposes the feature map into real and imaginary parts through Fourier transform, and performs dimensionality adjustment and convolution processing to extract frequency domain features, and then performs inverse transform to restore the spatial domain. The fusion module jointly enhances the frequency and spatial domain features, and outputs the residual link to retain low-frequency background information and alleviate gradient vanishing.

[0046] Step S3.5: The activation and normalization layer uses a combination of batch normalization and nonlinear activation functions to stabilize training and enhance feature representation capabilities.

[0047] Step S4: Construct a reconstruction network model based on an encoder-decoder architecture, input the enhanced post-feature map into the reconstruction network model, extract global features through the encoder, and restore spatial resolution by combining multi-scale features through the decoder, and output the enhanced remote sensing image.

[0048] The reconstruction network model is based on an encoder-decoder architecture. This network uses an encoder-decoder structure combined with a multi-scale feature fusion module to achieve feature extraction and reconstruction through layer-by-layer downsampling and skip connections.

[0049] Specifically, the encoder stage uses convolutional downsampling to progressively expand the receptive field to extract global features; the decoder stage uses interpolation upsampling and skip connections to gradually restore spatial resolution, and combines a multi-scale feature fusion module to output an enhanced remote sensing image.

[0050] In this embodiment, step S4 specifically includes: Step S4.1: The encoder consists of 4 convolutional layers with 3×3 kernels and a stride of 2, and downsamples step by step to extract multi-scale features.

[0051] Step S4.2: The decoder restores spatial resolution through interpolation upsampling and skip connections, and fuses encoder features.

[0052] Step S4.3: The feature fusion output layer performs multi-scale feature stitching and channel unification to output an enhanced remote sensing image.

[0053] In this embodiment, a network training method is required, the details of which are as follows: Step 1: Divide the dataset from step S1 into training set D1 and test set D2 in an 8:2 ratio, each containing two classes of images: low light and normal light.

[0054] Step Two: The training is divided into two stages: T1 decomposition stage: The training network decomposes the input image into reflection and illumination feature maps; T2 Reconstruction Phase: The network is trained to reconstruct the normally lit image based on the decomposition results.

[0055] Step 3: Design of T1 loss function: The reflection loss E1 is the L1 norm difference between the low-light and normal-light reflection feature maps; The illumination loss E2 is the L1 norm difference between the low-light and normal-light illumination feature maps; Total loss ET1 = E1 + E2.

[0056] Step 4: T2 loss function design: reconstruct the L1 norm difference between the output image and the target image (D1_normal / D2_normal).

[0057] Step 5: Hyperparameter settings. Use the Adam optimizer, learning rate 0.001, train for 100 epochs in T1 and T2 phases, and set the batch size to 32.

[0058] This invention presents a data-driven image enhancement method for low-light-field remote sensing images. This method employs an image enhancement algorithm that integrates data-driven learning and Retinex theory, effectively enhancing optical remote sensing images under low-light conditions. Based on Retinex theory, this method designs a cascaded backbone network structure consisting of a decomposition sub-network and a reconstruction sub-network. The network first decomposes the input low-light-field remote sensing image into a reflectance feature map and an illumination feature map. Then, a frequency-domain Fourier convolution module performs global feature modeling and local detail enhancement in both the spatial and frequency domains, significantly improving the quality of image feature reconstruction. Finally, an encoder-decoder architecture reconstruction network fuses and restores the decomposed and enhanced features, generating enhanced remote sensing images with higher visual quality. Compared to traditional low-light enhancement methods, this invention's data-driven image enhancement method for low-light conditions exhibits superior image enhancement performance, effectively improving the brightness, contrast, and detail quality of remote sensing images, meeting the enhancement requirements of optical remote sensing images under complex imaging conditions. The data-driven remote sensing image enhancement method based on low light field in this invention is particularly suitable for nighttime observation, shadow area feature identification, and remote sensing monitoring tasks in low light environments. It can provide a higher quality image foundation for subsequent advanced visual tasks such as target detection and change analysis.

[0059] The data-driven remote sensing image enhancement method for weak light fields according to embodiments of the present invention has the following technical effects: 1. The data-driven remote sensing image enhancement method for low-light fields according to embodiments of the present invention employs an image enhancement model that integrates data-driven learning and Retinex theory. By constructing a cascaded backbone structure consisting of a decomposition network and a reconstruction network, the input low-light remote sensing image is first decomposed into a reflectance feature map and an illumination feature map. Then, a frequency domain Fourier convolution module is used to achieve joint modeling in the spatial and frequency domains, enhancing local details while ensuring global illumination consistency. This design can fully utilize frequency domain information to extract and enhance features of the illumination and reflectance components, thereby effectively improving the brightness, contrast, and detail of remote sensing images under low-light fields, improving overall image quality, enhancing the feature extraction capability of ground objects, and significantly improving the accuracy and stability of downstream remote sensing visual tasks (such as target recognition, change detection, and ground object classification).

[0060] 2. The low-light field optical remote sensing image dataset constructed in this embodiment of the invention covers a variety of land cover types and imaging conditions, with an illumination range of 10-80 lux. It has good diversity and representativeness, and can provide a high-quality training data foundation and standardized sample source for related research such as remote sensing image enhancement, target detection, and change analysis.

[0061] 3. This invention presents a precise method for separating illumination and reflection components under low light conditions. This method combines Retinex theory with a data-driven optimization mechanism to effectively separate the illumination component (low-frequency information) and reflection component (high-frequency texture details) in remote sensing images. It overcomes the decomposition blurring problem that easily occurs in traditional Retinex methods under strong noise or extremely low illumination conditions, thereby significantly improving the visibility and distinguishability of typical ground features such as buildings, roads, and water edges.

[0062] 4. This invention proposes a global-local collaborative enhancement mechanism for low-light conditions. Based on a frequency-domain Fourier convolution module, this mechanism achieves collaborative optimization of global modeling in the frequency domain and local detail enhancement in the spatial domain: In frequency domain processing, Fourier transform is used to capture long-range dependencies in the image, suppressing random noise and illumination disturbances, and avoiding edge blurring that may be caused by traditional spatial filtering; in spatial domain processing, a multi-scale convolution structure is used to preserve high-frequency detail information such as ground texture, building edges, and road patterns, ultimately achieving an image enhancement effect consistent with visual perception and physical characteristics.

[0063] 5. This invention presents a method for dynamic range expansion and color fidelity preservation in low-light conditions. By optimizing the illumination and reflection components in stages through the cascaded decomposition-reconstruction network, the dynamic range of low-light images can be effectively expanded, avoiding overexposure and color distortion during the enhancement process. This ensures that the enhanced remote sensing image maintains natural harmony in terms of brightness distribution, color saturation, and consistency with the reflectivity of ground features, thereby significantly improving the color difference recognition between different ground feature types.

[0064] 6. The data-driven remote sensing image enhancement method for weak light fields in this invention has excellent real-time processing performance and broad hardware platform compatibility. This method employs an encoder-decoder structure combined with a lightweight frequency-domain convolution design, enabling the algorithm to achieve real-time enhancement of 512×512 resolution images (frame rate ≥ 30 FPS) on computing platforms such as GPUs. This meets the performance requirements of remote sensing data processing and airborne image systems in low-latency, high-throughput application scenarios, and supports edge computing deployment at satellite, UAV, and ground stations.

[0065] 7. The data-driven remote sensing image enhancement method for low-light fields in this invention possesses strong algorithmic robustness and scene generalization ability. Based on large-scale low-light remote sensing image data for data-driven learning, the method can adapt to different levels of illumination attenuation (10–50 lux) and noise levels, and can still stably output high-quality enhancement results under complex meteorological conditions such as cloudy days, rain, fog, and dust storms. This method can effectively reduce the false detection and false negative rates in subsequent ground feature detection and recognition tasks, significantly improving the reliability and practical value of the entire remote sensing information processing system.

[0066] Example 2 This invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. The computer program is a low-light-field remote sensing image enhancement program, such as... Figure 1 As shown, when the processor executes the low-light-field remote sensing image enhancement program, it performs the following steps: Step S1: Based on the remote sensing application scenario, set the light source angle and brightness conditions, construct an optical remote sensing image enhancement dataset under dim light field, and acquire the collected remote sensing images to be enhanced. Step S2: Construct a decomposition network based on Retinex theory, train the decomposition network using the augmentation dataset to obtain an image decomposition model; input the remote sensing image to be augmented into the trained image decomposition model to decompose it into a reflectance feature map and an illumination feature map; Step S3: Construct a feature enhancement module based on frequency domain Fourier convolution. Input the reflection feature map and the illumination feature map into the feature enhancement module to perform joint feature extraction and enhancement in the frequency domain and spatial domain to obtain the enhanced feature map. Step S4: Construct a reconstruction network model based on an encoder-decoder architecture, input the enhanced post-feature map into the reconstruction network model, extract global features through the encoder, and restore spatial resolution by combining multi-scale features through the decoder, and output the enhanced remote sensing image.

[0067] Furthermore, the construction of the optical remote sensing image enhancement dataset under low light field in step S1 specifically includes: Step S1.1: Acquire image data under different lighting conditions using a multi-platform remote sensing imaging device, and adjust the sensitivity to control the imaging brightness range within a preset range; Step S1.2: Collect sample images of various land cover types and record the lighting conditions and shooting parameters during imaging; Step S1.3: Perform at least one data augmentation operation on the acquired image, including rotation, flipping, brightness perturbation, and noise simulation.

[0068] Furthermore, the decomposition network based on Retinex theory in step S2 has a structure that includes an input preprocessing layer, an initialization convolutional layer, a feature extraction layer, an output convolutional layer, a Sigmoid activation layer, and a separation adjustment layer connected in sequence.

[0069] Furthermore, the input preprocessing layer is used to concatenate the image tensor of shape (B,H,W,3) and the illumination guidance tensor of shape (B,H,W,1) in the channel dimension, and adjust the dimension to (B,4,H,W). And / or, the kernel size of the initial convolutional layer is 9×9, the padding is 4, the input channel is 4, and the output channel is 64; And / or, the feature extraction layer consists of multiple consecutive convolutional layers with a kernel size of 3×3; And / or, the output convolutional layer has a 3×3 kernel, 64 input channels, and 4 output channels; And / or, the Sigmoid activation layer is used to normalize the output feature values ​​to the range [0,1]; And / or, the separation adjustment layer is used to separate the output features of the 4 channels into a reflection feature map of shape (B,H,W,3) and an illumination feature map of shape (B,H,W,1).

[0070] Furthermore, the feature enhancement module based on frequency domain Fourier convolution in step S3 includes an input segmentation layer, a local feature extraction layer, a feature concatenation layer, a frequency domain fusion unit, and an activation and normalization layer connected in sequence.

[0071] Furthermore, the frequency domain fusion unit is used to perform the following processing on the input features: Perform a Fourier transform on the feature map and separate the transform result into the real part and the imaginary part; The real and imaginary parts are subjected to dimensionality adjustment and convolution operations respectively to extract frequency domain features; Perform an inverse Fourier transform on the processed frequency domain features to restore them to the spatial domain; The recovered spatial features are jointly enhanced with the original spatial input features through a fusion module, and then output through residual connections.

[0072] Furthermore, the reconstruction network model based on the encoder-decoder architecture in step S4 includes: The encoder is used to downsample the input features layer by layer to generate multi-scale encoded features; The decoder, corresponding to each level of the encoder, is used to gradually restore spatial resolution through upsampling operations and to fuse encoded features from the corresponding encoder levels using skip connections, ultimately outputting an enhanced remote sensing image.

[0073] Furthermore, the method also includes a network training step, specifically: The augmented dataset is divided into a training set and a test set; The first stage of training the decomposition network uses the loss function as the difference between the decomposed reflection feature map and illumination feature map and the corresponding image feature map under normal illumination. The second stage trains the reconstruction network model, and the loss function is the difference between the enhanced image output by the reconstruction network model and the corresponding normal illumination image.

[0074] like Figure 4 The diagram shown is a structural diagram of an electronic device according to an exemplary embodiment. The contents shown in the diagram are merely examples and should not be construed as any limitation on the scope of the present invention.

[0075] The electronic device in this embodiment may specifically include: at least one processor, at least one memory, a power supply, a communication interface, an input / output interface, and a communication bus. Specifically, the memory stores a computer program, which is loaded and executed by the processor to implement the relevant steps in the data-driven low-light field remote sensing image enhancement method of the present invention. Alternatively, the electronic device in this embodiment may be a computer or other suitable device.

[0076] The power supply provides operating voltage for the various hardware devices on the electronic device; the communication interface can establish a data transmission channel between the electronic device and external devices, and the communication protocol it follows can be any protocol that can adapt to the technical solution of this application, and the specific protocol type is not limited; the input and output interface is used to receive external input data or output data to external devices, and the specific interface type can be selected according to the actual application requirements, and is not limited in detail here.

[0077] In addition, memory, as a carrier for resource storage, can be read-only memory (ROM), random access memory (RAM), disk, or optical disk, etc. The resources stored on it can include operating systems, computer programs, and data. Data may include various application data or model parameters, and the storage method can be short-term or long-term storage.

[0078] The operating system is used to manage and control the various hardware devices and computer programs on the electronic device. The operating system can be Windows Server, Netware, Unix, Linux, etc. In addition to programs for implementing the data-driven low-light field remote sensing image enhancement method of this invention, the computer programs may also include other programs for performing specific tasks.

[0079] Example 3 This invention also provides a storage medium, which is a computer-readable storage medium storing a computer program thereon. The computer program is a low-light-field remote sensing image enhancement program, such as... Figure 1 As shown, when the dim light field remote sensing image enhancement program is executed by the processor, it performs the following steps: Step S1: Based on the remote sensing application scenario, set the light source angle and brightness conditions, construct an optical remote sensing image enhancement dataset under dim light field, and acquire the collected remote sensing images to be enhanced. Step S2: Construct a decomposition network based on Retinex theory, train the decomposition network using the augmentation dataset to obtain an image decomposition model; input the remote sensing image to be augmented into the trained image decomposition model to decompose it into a reflectance feature map and an illumination feature map; Step S3: Construct a feature enhancement module based on frequency domain Fourier convolution. Input the reflection feature map and the illumination feature map into the feature enhancement module to perform joint feature extraction and enhancement in the frequency domain and spatial domain to obtain the enhanced feature map. Step S4: Construct a reconstruction network model based on an encoder-decoder architecture, input the enhanced post-feature map into the reconstruction network model, extract global features through the encoder, and restore spatial resolution by combining multi-scale features through the decoder, and output the enhanced remote sensing image.

[0080] Furthermore, the construction of the optical remote sensing image enhancement dataset under low light field in step S1 specifically includes: Step S1.1: Acquire image data under different lighting conditions using a multi-platform remote sensing imaging device, and adjust the sensitivity to control the imaging brightness range within a preset range; Step S1.2: Collect sample images of various land cover types and record the lighting conditions and shooting parameters during imaging; Step S1.3: Perform at least one data augmentation operation on the acquired image, including rotation, flipping, brightness perturbation, and noise simulation.

[0081] Furthermore, the decomposition network based on Retinex theory in step S2 has a structure that includes an input preprocessing layer, an initialization convolutional layer, a feature extraction layer, an output convolutional layer, a Sigmoid activation layer, and a separation adjustment layer connected in sequence.

[0082] Furthermore, the input preprocessing layer is used to concatenate the image tensor of shape (B,H,W,3) and the illumination guidance tensor of shape (B,H,W,1) in the channel dimension, and adjust the dimension to (B,4,H,W). And / or, the kernel size of the initial convolutional layer is 9×9, the padding is 4, the input channel is 4, and the output channel is 64; And / or, the feature extraction layer consists of multiple consecutive convolutional layers with a kernel size of 3×3; And / or, the output convolutional layer has a 3×3 kernel, 64 input channels, and 4 output channels; And / or, the Sigmoid activation layer is used to normalize the output feature values ​​to the range [0,1]; And / or, the separation adjustment layer is used to separate the output features of the 4 channels into a reflection feature map of shape (B,H,W,3) and an illumination feature map of shape (B,H,W,1).

[0083] Furthermore, the feature enhancement module based on frequency domain Fourier convolution in step S3 includes an input segmentation layer, a local feature extraction layer, a feature concatenation layer, a frequency domain fusion unit, and an activation and normalization layer connected in sequence.

[0084] Furthermore, the frequency domain fusion unit is used to perform the following processing on the input features: Perform a Fourier transform on the feature map and separate the transform result into the real part and the imaginary part; The real and imaginary parts are subjected to dimensionality adjustment and convolution operations respectively to extract frequency domain features; Perform an inverse Fourier transform on the processed frequency domain features to restore them to the spatial domain; The recovered spatial features are jointly enhanced with the original spatial input features through a fusion module, and then output through residual connections.

[0085] Furthermore, the reconstruction network model based on the encoder-decoder architecture in step S4 includes: The encoder is used to downsample the input features layer by layer to generate multi-scale encoded features; The decoder, corresponding to each level of the encoder, is used to gradually restore spatial resolution through upsampling operations and to fuse encoded features from the corresponding encoder levels using skip connections, ultimately outputting an enhanced remote sensing image.

[0086] Furthermore, the method also includes a network training step, specifically: The augmented dataset is divided into a training set and a test set; The first stage of training the decomposition network uses the loss function as the difference between the decomposed reflection feature map and illumination feature map and the corresponding image feature map under normal illumination. The second stage trains the reconstruction network model, and the loss function is the difference between the enhanced image output by the reconstruction network model and the corresponding normal illumination image.

[0087] The storage media of the present invention include any other form of storage media known in the art, such as random access memory (RAM), memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, register, hard disk, magnetic disk or optical disk.

[0088] The steps of the data-driven remote sensing image enhancement method for weak light fields of the present invention can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, register, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0089] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A data-driven method for enhancing remote sensing images with weak light fields, characterized in that, include: Step S1: Based on the remote sensing application scenario, set the light source angle and brightness conditions, construct an optical remote sensing image enhancement dataset under dim light field, and acquire the collected remote sensing images to be enhanced. Step S2: Construct a decomposition network based on Retinex theory, train the decomposition network using the augmentation dataset to obtain an image decomposition model; input the remote sensing image to be augmented into the trained image decomposition model to decompose it into a reflectance feature map and an illumination feature map; Step S3: Construct a feature enhancement module based on frequency domain Fourier convolution. Input the reflection feature map and the illumination feature map into the feature enhancement module to perform joint feature extraction and enhancement in the frequency domain and spatial domain to obtain the enhanced feature map. Step S4: Construct a reconstruction network model based on an encoder-decoder architecture, input the enhanced post-feature map into the reconstruction network model, extract global features through the encoder, and restore spatial resolution by combining multi-scale features through the decoder, and output the enhanced remote sensing image.

2. The data-driven remote sensing image enhancement method for weak light fields according to claim 1, characterized in that, The specific steps in step S1 of constructing the optical remote sensing image enhancement dataset under low light field conditions include: Step S1.1: Acquire image data under different lighting conditions using a multi-platform remote sensing imaging device, and adjust the sensitivity to control the imaging brightness range within a preset range; Step S1.2: Collect sample images of various land cover types and record the lighting conditions and shooting parameters during imaging; Step S1.3: Perform at least one data augmentation operation on the acquired image, including rotation, flipping, brightness perturbation, and noise simulation.

3. The data-driven remote sensing image enhancement method for weak light fields according to claim 1 or 2, characterized in that, The decomposition network based on Retinex theory in step S2 has a structure that includes an input preprocessing layer, an initialization convolutional layer, a feature extraction layer, an output convolutional layer, a Sigmoid activation layer, and a separation adjustment layer connected in sequence.

4. The data-driven remote sensing image enhancement method for weak light fields according to claim 3, characterized in that, The input preprocessing layer is used to concatenate an image tensor of shape (B,H,W,3) and an illumination guiding tensor of shape (B,H,W,1) in the channel dimension, and adjust the dimension to (B,4,H,W). And / or, the kernel size of the initial convolutional layer is 9×9, the padding is 4, the input channel is 4, and the output channel is 64; And / or, the feature extraction layer consists of multiple consecutive convolutional layers with a kernel size of 3×3; And / or, the output convolutional layer has a 3×3 kernel, 64 input channels, and 4 output channels; And / or, the Sigmoid activation layer is used to normalize the output feature values ​​to the range [0,1]; And / or, the separation adjustment layer is used to separate the output features of the 4 channels into a reflection feature map of shape (B,H,W,3) and an illumination feature map of shape (B,H,W,1).

5. The data-driven remote sensing image enhancement method for weak light fields according to claim 1, characterized in that, The feature enhancement module based on frequency domain Fourier convolution in step S3 includes an input segmentation layer, a local feature extraction layer, a feature splicing layer, a frequency domain fusion unit, and an activation and normalization layer connected in sequence.

6. The data-driven remote sensing image enhancement method for weak light fields according to claim 5, characterized in that, The frequency domain fusion unit is used to perform the following processing on the input features: Perform a Fourier transform on the feature map and separate the transform result into real and imaginary parts; The real and imaginary parts are subjected to dimensionality adjustment and convolution operations respectively to extract frequency domain features; Perform an inverse Fourier transform on the processed frequency domain features to restore them to the spatial domain; The recovered spatial features are jointly enhanced with the original spatial input features through a fusion module, and then output through residual connections.

7. The data-driven method for enhancing remote sensing images with weak light fields according to claim 1, characterized in that, The reconstruction network model based on the encoder-decoder architecture in step S4 includes: The encoder is used to downsample the input features layer by layer to generate multi-scale encoded features; The decoder, corresponding to each level of the encoder, is used to gradually restore spatial resolution through upsampling operations and to fuse encoded features from the corresponding encoder levels using skip connections, ultimately outputting an enhanced remote sensing image.

8. The data-driven remote sensing image enhancement method for weak light fields according to claim 1, characterized in that, The method also includes a network training step, specifically: The augmented dataset is divided into a training set and a test set; The first stage of training the decomposition network uses the loss function as the difference between the decomposed reflection feature map and illumination feature map and the corresponding image feature map under normal illumination. The second stage trains the reconstruction network model, and the loss function is the difference between the enhanced image output by the reconstruction network model and the corresponding normal illumination image.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is a low-light-field remote sensing image enhancement program, characterized in that... When the processor executes the low-light-field remote sensing image enhancement program, it implements the steps of the data-driven low-light-field remote sensing image enhancement method as described in any one of claims 1-8.

10. A storage medium, said storage medium being a computer-readable storage medium, having stored thereon a computer program, said computer program being a low-light-field remote sensing image enhancement program, characterized in that, When the dim light field remote sensing image enhancement program is executed by the processor, it implements the steps of the data-driven dim light field remote sensing image enhancement method as described in any one of claims 1-8.

Images