Discrete laser spot restoration method based on atmospheric parameter adaptive fusion
By constructing a deep learning-based dual-branch neural network and combining it with adaptive fusion of atmospheric parameters, the problem of insufficient generalization ability of laser spot restoration methods in complex atmospheric environments was solved, and spot restoration with high precision and physical consistency was achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CENT CHINA OPTOELECTRONICS TECH RES INST (CHINA STATE SHIPBUILDING CORP 717TH RES INST)
- Filing Date
- 2025-12-22
- Publication Date
- 2026-04-28
AI Technical Summary
Existing laser spot restoration methods lack generalization ability in complex atmospheric environments, fail to comprehensively consider the combined degradation effects of atmospheric turbulence and discrete sampling effects of detector arrays, and lack prior knowledge of the physical morphology of laser spots.
A deep learning-based dual-branch neural network is constructed, which is combined with adaptive fusion of atmospheric parameters. By acquiring atmospheric parameters in real time and introducing an adaptive feature fusion module, high-precision reconstruction of discrete laser spots is achieved. The U-Net structure and multilayer perceptron are used for feature extraction and reconstruction, and the physical consistency loss function is used for training.
It significantly improves the accuracy and fidelity of laser spot reconstruction, enhances the model's generalization ability under different atmospheric environments, and ensures the physical consistency of the reconstruction results.
Smart Images
Figure CN121937331A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of laser detection and photoelectric signal processing technology, specifically to a discrete laser spot restoration method based on adaptive fusion of atmospheric parameters. Background Technology
[0002] In fields such as lidar, free-space optical communication, adaptive optics, and high-energy laser testing, laser light is transmitted through the atmosphere and then received by a detector array. The received light spot image is a discrete sampled signal. This signal is not only affected by waveform distortion caused by atmospheric turbulence and thermal corona effects, but is also inevitably constrained by discretization effects such as the spatial sampling frequency, fill factor, and noise of the detector array.
[0003] Traditional restoration methods typically address atmospheric turbulence correction or detector array super-resolution reconstruction independently, failing to comprehensively consider the combined degradation effects caused by the coupling of physical distortion and discrete sampling.
[0004] While deep learning-based image super-resolution methods can effectively improve image resolution, they are mostly designed for general natural images and lack prior knowledge of the physical morphology of laser spots. Furthermore, they do not incorporate key atmospheric parameters that affect the morphology of the spots as guidance, resulting in insufficient generalization ability and poor physical consistency in complex and variable atmospheric environments. Summary of the Invention
[0005] To address the problems of poor generalization ability and sensitivity to changes in atmospheric parameters in existing laser spot restoration algorithms under complex atmospheric environments, this invention proposes a laser spot restoration method based on deep learning and adaptive fusion of atmospheric parameters.
[0006] The technical solution adopted by this invention to solve its technical problem is: a discrete laser spot restoration method based on adaptive fusion of atmospheric parameters, comprising the following steps: Step 1: Construct a real-time atmospheric parameter inversion module to obtain the turbulence intensity under the current atmospheric environment. C n 2 Coherence length r 0 and thermal aberration number N D Key atmospheric parameters; Step two: Under different atmospheric conditions, use an array-type laser spot detection module to simultaneously acquire discrete laser spot images under the current atmospheric conditions and record the corresponding turbulence intensities. C n 2 Coherence length r 0 and thermal aberration number N DBased on atmospheric parameters, synchronous acquisition data of discrete sampled light spot images under real environment is constructed; the random phase screen method based on Fourier transform is used to simulate atmospheric turbulence phase disturbance under different turbulence intensities to obtain continuous light spot distortion images. The continuous light spot distortion images are discretized to obtain simulation data; training and test datasets are constructed by selecting synchronous acquisition data and simulation data at a ratio of 1:1. Step 3: Construct a deep learning network with an encoding-fusion-encoding structure. This network is designed as a dual-branch deep neural network. One image encoding branch is an encoder-encoder structure used to extract features from light spot images, and the other atmospheric parameter encoding branch is a multilayer perceptron used to encode atmospheric parameters. Step four: Introduce an adaptive feature fusion module into the deep neural network. Using methods such as channel attention or conditional instance normalization, the atmospheric parameter features output by the multilayer perceptron are adaptively injected into the decoding and reconstruction branches of image reconstruction, so as to guide the restoration process with physical priors. Step 5: The neural network outputs a restored high-sampling-rate, high-precision laser spot image; Step 6: Based on the training and testing datasets, jointly guide the neural network training with image reconstruction loss (L1 loss, multi-scale SSIM loss) and physical consistency loss (spot moment constraint, energy conservation constraint); Step 7: Deploy the trained neural network to achieve real-time high-precision restoration of sampled discrete laser spots and output the restored high-resolution spot image.
[0007] Furthermore, the image encoding branch takes a low-resolution discrete sampled spot image as input and outputs a high-resolution restored spot image. The network backbone adopts a lightweight U-Net structure. The encoder includes a downsampling module with two repeated 3×3 convolutional layers and a LeakyReLU activation function. After the downsampling module, a 2×2 max pooling layer with a stride of 2 is connected for downsampling. After the repeated downsampling module, a global average pooling layer is connected to extract the global image feature vector of the image.
[0008] Furthermore, the atmospheric parameter encoding branch is composed of fully connected layers. The number of neurons in the input layer is the same as the dimension of the atmospheric parameters. The network backbone is composed of a three-layer multilayer perceptron, which consists of a fully connected layer (FC), a ReLU activation function, and another fully connected layer (FC). The output layer outputs a 128-dimensional atmospheric parameter feature vector to characterize the atmospheric state.
[0009] Furthermore, the adaptive feature fusion module uses conditional instance normalization for deep fusion, and generates a set of scaling parameters and translation parameters respectively by passing the atmospheric parameter feature vector through two different fully connected layers.
[0010] Furthermore, the decoding and reconstruction branch includes an upsampling module and a feature stitching module following the upsampling module. The feature stitching module takes the global image feature vector as input and outputs it to the conditional instance normalization module. The conditional instance normalization module, along with the input scaling and translation parameters, sends them to the downsampling module with a 3×3 convolutional layer and a LeakyReLU activation function. The repeated upsampling module outputs the restored spot image through the final convolutional layer.
[0011] Furthermore, step five, to ensure the physical reality of the restored light spot composite, has the following formula for physical consistency loss: ,in L radius To restore the second moment between the original light spot and the ideal light spot, that is, the deviation of the light spot radius; L centroid The Euclidean distance between the centroid of the restored light spot and the ideal light spot.
[0012] Furthermore, in step six, the AdamW optimizer is used for training with an initial learning rate of 1e-4, a cosine annealing strategy, a batch size of 16, and a total of 500 epochs.
[0013] The beneficial effects of this invention are as follows: This invention adaptively integrates atmospheric physical parameters as prior knowledge into the deep learning model, overcoming the limitations of traditional methods in separating discrete sampling and atmospheric interference, as well as the poor physical consistency of pure data-driven methods. It has the following advantages: it achieves joint correction and restoration of the discrete sampling effect of the detector array and the atmospheric turbulence and thermal halo effect, significantly improving the accuracy and fidelity of the restored light spot; it introduces dynamic atmospheric parameters as conditional inputs to the network, greatly enhancing the model's generalization ability and adaptability under different atmospheric environments; by introducing physical consistency loss, it ensures the physical characteristics of the composite light spot in the restored result, improving the reliability and interpretability of the method; this invention can be widely used in systems such as lidar and optical remote sensing that require high-precision analysis of discrete sampled light spots. Attached Figure Description
[0014] Figure 1 This is a detailed network structure diagram of the present invention. Detailed Implementation
[0015] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0016] The technical problem this invention aims to solve is to design a deep learning network capable of simultaneously sensing discrete sampling effects and atmospheric turbulence and thermal halo, and dynamically fusing atmospheric physical parameters, to achieve high-precision, physically consistent restoration of discrete distorted light spots. (Refer to...) Figure 1 As shown, the present invention provides a method for high-precision laser spot restoration by fusing atmospheric physical parameters with discrete sampled signals from an array detector. The implementation steps of the discrete laser spot restoration method based on deep learning algorithms are as follows.
[0017] Step 1: Construct a real-time atmospheric parameter inversion module to obtain the turbulence intensity under the current atmospheric environment. C n 2 Coherence length r 0 and thermal aberration number N D Key atmospheric parameters, etc.
[0018] Step two: Under different atmospheric conditions, use an array-type laser spot detection module to simultaneously acquire discrete laser spot images under the current atmospheric conditions and record the corresponding turbulence intensities. C n 2 Coherence length r 0 and thermal aberration number N D Based on atmospheric parameters, synchronously acquired data of discrete sampled light spot images under real-world conditions was constructed. The stochastic phase screen method based on Fourier transform was used to simulate atmospheric turbulence phase disturbances under different turbulence intensities, obtaining continuous light spot distortion images. These continuous light spot distortion images were then discretized to obtain simulation data. Training and testing datasets were constructed by selecting synchronously acquired data and simulation data at a 1:1 ratio.
[0019] Step 3: Construct a deep learning network with an encoding-fusion-encoding structure. This network is designed as a two-branch deep neural network. One image encoding branch is an encoder-encoder structure used to extract light spot image features, and the other atmospheric parameter encoding branch is a multilayer perceptron used to encode atmospheric parameters.
[0020] The image encoding branch takes a low-resolution discrete sampled spot image as input and outputs a high-resolution restored spot image. The network backbone uses a lightweight U-Net structure. The encoder consists of four downsampling modules, each containing two repeated 3×3 convolutional layers followed by a LeakyReLU activation function (with a negative slope of 0.2), and then a 2×2 max-pooling layer with a stride of 2 for downsampling. The initial number of channels is set to 64, doubling at each downsampling stage. At the end of the encoder, a global average pooling layer (GAP) is used to extract the global image feature vector.
[0021] The atmospheric parameter encoding branch consists of fully connected layers, and the network backbone consists of a three-layer multilayer perceptron. The number of neurons in the input layer is the same as the dimension of the atmospheric parameters. The number of neurons in the two hidden layers are 64 and 128, respectively, and the ReLU activation function is used. The output layer outputs a 128-dimensional atmospheric parameter feature vector to represent the atmospheric state.
[0022] Step four involves introducing an adaptive feature fusion module into the deep neural network. By employing methods such as channel attention or conditional instance normalization, the atmospheric parameter features output by the multilayer perceptron are adaptively injected into the decoding and reconstruction branches used for image reconstruction, thereby guiding the restoration process with physical priors.
[0023] In this embodiment, the adaptive feature fusion module uses conditional instance normalization for deep fusion, and generates a set of scaling parameters and translation parameters respectively by passing the atmospheric parameter feature vector through two different fully connected layers.
[0024] Before each upsampling module in the decoding and reconstruction branch, an affine transformation is performed on the feature map from the skip connections to dynamically and adaptively inject atmospheric environment / condition parameters into the image features, thereby guiding the decoding process. This step decodes through the decoding and reconstruction branch and outputs the restored spot image.
[0025] Step 5: The neural network outputs a restored high-sampling-rate, high-precision laser spot image.
[0026] Design loss function ,in l 1 and l 2 is a constant, set to 1.0 and 0.5 respectively; L recon To reconstruct the lost components; L physi This represents the physical consistency loss component. To facilitate the restoration of details, the reconstruction loss component uses Charbonnier Loss, calculated using the following formula: ,in I restored To restore the light spot image, I ideal For an ideal spot image, e Set to 0.001.
[0027] To ensure the physical reality of the restored light spot composite, the physical consistency loss component is designed as a loss based on the light spot moment, and its calculation formula is as follows: ,in L radius To restore the second moment between the original light spot and the ideal light spot, that is, the deviation of the light spot radius; L centroidThe Euclidean distance between the centroid of the restored light spot and the ideal light spot.
[0028] Step 6: Based on the training and testing datasets, the image reconstruction loss (L1 loss, multi-scale SSIM loss) and physical consistency loss (spot moment constraint, energy conservation constraint) are combined to guide the training of the neural network.
[0029] The AdamW optimizer was used for training with an initial learning rate of 1e-4, a cosine annealing strategy, and a batch size of 16. Training lasted for 500 epochs. During training, performance was tested on the validation set after each epoch, and the model parameters with the lowest validation loss were saved.
[0030] Step 7: Deploy the trained neural network. After training, convert the network model to ONNX format and deploy it on an embedded computing platform to achieve real-time high-precision restoration of sampled discrete laser spots and output the restored high-resolution spot image.
[0031] Those skilled in the art will readily understand that the above description is merely a preferred use case of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A discrete laser spot restoration method based on adaptive fusion of atmospheric parameters, characterized in that: Includes the following steps Step 1: Construct a real-time atmospheric parameter inversion module to obtain the current atmospheric parameters, including turbulence intensity C. n 2 Coherence length r0 and thermal distortion number N D Key atmospheric parameters, including; Step 2: Use an array-type laser spot detection module to synchronously acquire discrete laser spot images under the current atmospheric environment and record the corresponding key atmospheric parameters to construct synchronous acquisition data of discrete sampled spot images under real environment. The random phase screen method was used to simulate atmospheric turbulence phase disturbances, and continuous spot distortion images were obtained. These images were then discretized to obtain simulation data. Training and test datasets were constructed by selecting synchronously acquired data and simulation data at a ratio of 1:
1. Step 3: Design a dual-branch deep neural network. One image coding branch is an encoder-encoder structure that extracts features from discrete sampled light spot images, and the other atmospheric parameter coding branch is a multilayer perceptron that encodes atmospheric parameters. Step four: Introduce an adaptive feature fusion module into the deep layers of the neural network. Using channel attention or conditional instance normalization, the atmospheric parameter features output by the multilayer perceptron are adaptively injected into the decoding and reconstruction branches to guide the restoration process with physical priors. Step 5: The neural network outputs the restored laser spot image; Step 6: Based on the training and testing datasets, jointly guide the training of the neural network using image reconstruction loss and physical consistency loss. Step 7: Deploy the trained neural network. After training, convert the network model to ONNX format to restore the sampled discrete laser spot and output the restored high-resolution spot image.
2. The discrete laser spot restoration method based on adaptive fusion of atmospheric parameters according to claim 1, characterized in that, The image coding branch network backbone adopts a lightweight U-Net structure, including a downsampling module with two repeated 3×3 convolutional layers and a LeakyReLU activation function. After the downsampling module, a 2×2 max pooling layer with a stride of 2 is connected to perform downsampling. After the repeated downsampling module, a global average pooling layer is connected to extract the global image feature vector of the image.
3. The discrete laser spot restoration method based on adaptive fusion of atmospheric parameters according to claim 2, characterized in that, The atmospheric parameter encoding branch network backbone consists of a fully connected layer, a ReLU activation function, and another fully connected layer, and outputs an atmospheric parameter feature vector.
4. A discrete laser spot restoration method based on adaptive fusion of atmospheric parameters according to claim 1, 2, or 3, characterized in that, The adaptive feature fusion module uses conditional instance normalization for deep fusion, and generates a set of scaling and translation parameters by passing the atmospheric parameter feature vector through two different fully connected layers.
5. The discrete laser spot restoration method based on adaptive fusion of atmospheric parameters according to claim 4, characterized in that, The decoding and reconstruction branch includes an upsampling module and a feature stitching module following the upsampling module. The feature stitching module takes the global image feature vector as input and outputs it to the conditional instance normalization module. The conditional instance normalization module, along with the input scaling and translation parameters, sends them to the downsampling module, which has a 3×3 convolutional layer and a LeakyReLU activation function. The repeated upsampling module outputs the restored spot image through the final convolutional layer.
6. The discrete laser spot restoration method based on adaptive fusion of atmospheric parameters according to claim 5, characterized in that, The formula for physical consistency loss in step five is as follows: L radius To restore the second moment of the light spot and the ideal light spot, L centroid The Euclidean distance between the centroid of the restored light spot and the ideal light spot.
7. The discrete laser spot restoration method based on adaptive fusion of atmospheric parameters according to claim 6, characterized in that, Step six uses the AdamW optimizer for training, with an initial learning rate of 1e-4, a cosine annealing strategy, a batch size of 16, and a total of 500 epochs.