High frequency guided light weight scramjet flow field schlieren image reconstruction method

Abstract

This invention discloses a high-frequency guided method for reconstructing flow field schlieren images of a lightweight scramjet engine, belonging to the technical field of image reconstruction. The method includes acquiring one-dimensional sparse pressure sequence data of the scramjet engine wall; generating an initial two-dimensional feature tensor by projecting the one-dimensional sparse pressure sequence data; inputting the initial two-dimensional feature tensor into a multi-scale feature encoder to obtain intermediate features and deep global semantic features at each scale; inputting the deep global semantic features into a spatial decoder for multi-stage cascaded decoding; and then completing cross-scale feature aggregation. After multi-stage cascaded decoding by the spatial decoder, a high-resolution scramjet engine flow field schlieren image is reconstructed. This invention eliminates redundant feature extraction branches and can directly and accurately map extremely sparse pressure sequence data into a high-resolution flow field schlieren image, achieving accurate capture and dynamic compensation of high-frequency detail features with low computational power consumption.

Description

Technical Field

[0001] This invention belongs to the technical field of image reconstruction, specifically relating to a high-frequency guided method for reconstructing flow field schlieren images of a lightweight scramjet engine. Background Technology

[0002] The scramjet engine is a core power unit in the field of hypersonic propulsion, suitable for air-breathing flight at Mach 5 and above. Its structure is extremely simple, with no rotating parts, consisting of an air intake, isolator, combustion chamber, and exhaust nozzle. It relies on the ramjet effect generated by the high-speed flight of the aircraft to compress air, and the airflow directly enters the combustion chamber at supersonic speed for combustion, avoiding the high temperature and high total pressure losses caused by the strong deceleration of traditional subsonic ramjet engines. Compared to rockets, scramjet engines do not require oxidizer, have higher specific impulse, and longer range, making them the preferred power source for hypersonic missiles, near-space vehicles, and space-to-ground transportation vehicles.

[0003] Real-time monitoring of the supersonic combustion flow field inside a scramjet engine is crucial for the efficient design and active control of aircraft. In recent years, deep learning techniques for visual flow field reconstruction using wall pressure have made significant progress. However, existing high-performance reconstruction networks generally suffer from large parameter counts and high computational costs, making it difficult to meet the real-time online inference requirements of edge computing hardware under harsh environments. Summary of the Invention

[0004] The purpose of this invention is to address the aforementioned shortcomings in the prior art by providing a high-frequency guided lightweight scramjet engine flow field schlieren image reconstruction method. This method solves the problems of existing high-performance reconstruction networks having a large number of parameters and high computational overhead, making it difficult to meet the real-time online inference requirements of edge computing hardware under harsh environments.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A high-frequency guided method for reconstructing the schlieren image of a lightweight scramjet engine flow field includes the following steps: S1. Obtain one-dimensional sparse pressure sequence data of the scramjet engine wall; S2. The one-dimensional sparse pressure sequence data is transformed by projection to generate an initial two-dimensional feature tensor; S3. Input the initial two-dimensional feature tensor into a multi-scale feature encoder, and after multi-stage cascaded residual feature extraction and layer-by-layer downsampling, obtain intermediate features and deep global semantic features at each scale. S4. Input the deep global semantic features into the spatial decoder and perform multi-stage cascaded decoding. In each decoding stage, the deep decoding features output by the previous decoding stage are upsampled through a high-frequency guided convolution module. The upsampled features are then aligned with the intermediate features of the corresponding scale in the encoder and concatenated at the channel level to complete cross-scale feature aggregation. S5. Based on the cross-scale feature aggregation results, a high-resolution scramjet engine flow field schlieren image is reconstructed through multi-stage cascaded decoding of the spatial decoder.

[0006] Furthermore, S2 specifically includes: One-dimensional sparse pressure sequence data is input into the projection layer for projection linear transformation to obtain the initial two-dimensional feature tensor.

[0007] Furthermore, in S3, residual features are extracted through a residual feature extraction module; the residual feature extraction module includes a lightweight dual-branch residual topology, in which the main branch of the dual-branch residual topology is connected in series with two CBS feature extraction modules, and the side branches of the dual-branch residual topology are single-layer convolutions. The CBS feature extraction module performs nonlinear semantic mapping on the input features. The side branches are dimension-aligned by a single-layer convolution and then added element-wise with the output of the main branch to obtain the output feature map.

[0008] Furthermore, in S3, downsampling includes: The output feature map of the residual feature extraction module is input into the max pooling layer. The max pooling layer performs spatial dimension downsampling through a 2×2 non-overlapping sliding window to extract the strongest activation features in the local receptive field, so that the spatial resolution of the output feature map is reduced proportionally, thus obtaining intermediate features. After multi-stage cascaded downsampling, the multi-scale feature encoder finally outputs deep global semantic features.

[0009] Furthermore, in S4, the high-frequency guided convolution module is a dual-branch parallel dynamic upsampling topology, including a dynamic kernel prediction backbone branch and a feature recombination auxiliary branch. The dynamic kernel prediction backbone branch extracts high-frequency edge priors of the flow field, and the feature recombination auxiliary branch generates a context tensor containing local receptive field information. Finally, a dynamic weighted aggregation mechanism is used to fuse the output of the dynamic kernel prediction backbone branch and the output of the feature recombination auxiliary branch to obtain the features after upsampling by the high-frequency guided convolution module, which is the dynamic weighted fusion result.

[0010] Furthermore, the dynamic kernel prediction backbone branch extracts high-frequency edge priors of the flow field using the discrete Laplace operator, which is expressed as: In the formula, For the high-frequency edge features of the c-th feature channel, For the spatial semantic features of the c-th feature channel, It is an isotropic 8-adjacent Laplace filter kernel; This is a two-dimensional discrete convolution operation; Subsequently, the original semantic features High-frequency edge features Tight stitching is performed along the channel dimension, and dynamic recombination weights are generated through sub-pixel convolution and spatial normalization. .

[0011] Furthermore, in the feature reconstruction auxiliary branch, the input features of the feature reconstruction auxiliary branch are sequentially subjected to nearest neighbor interpolation upsampling and spatial expansion operations, and reshaped into a context tensor containing local receptive field information. .

[0012] Furthermore, a dynamic weighted aggregation mechanism is adopted to dynamically reorganize the weights of the predicted backbone branches by the dynamic kernel. Context tensor of feature reorganization auxiliary branch The fusion process yields a dynamically weighted fusion result, which is represented as follows: In the formula, The value of the final output feature tensor at a specific channel c and a specific pixel position (i, j); This represents the side length of the local aggregation window; It is then traversing this K 2 A loop variable of pixels; The coordinates of the spatial dimensions for traversing the entire feature map.

[0013] Furthermore, in step S4, the upsampled features are aligned with the intermediate features of the corresponding scale in the encoder and concatenated at the channel level to complete cross-scale feature aggregation, which is expressed as: In the formula, For the first Deep decoding features of hierarchical output, express Deep decoding features of hierarchical output; The intermediate feature represents the shallow feature response of the same resolution level of the multi-scale feature encoder. Its internal structure has not undergone excessive spatial compression, thus preserving extremely rich fine-grained coordinate and physical boundary information. This represents the upsampling operator in the process of reconstructing spatial resolution. This upsampling operator corresponds to the high-frequency guided convolution module mentioned below, which implements the upsampling operation. It is a cascaded nonlinear feature transformation network.

[0014] The high-frequency guided schlieren image reconstruction method for a lightweight scramjet engine provided by this invention has the following beneficial effects: This invention eliminates redundant feature extraction branches, enabling the direct and accurate mapping of extremely sparse pressure sequence data into high-resolution flow field schlieren images. To fundamentally eliminate upsampling artifacts, a High-Frequency Guided Convolution (HFGC) module is designed. This module achieves accurate capture and dynamic compensation of high-frequency detailed features with low computational cost through the synergistic effect of dynamic kernel prediction main branches and feature recombination auxiliary branches. Attached Figure Description

[0015] Figure 1 This is a flowchart of the flow field schlieren image reconstruction method for a lightweight scramjet engine guided by high frequency in the embodiment.

[0016] Figure 2 The example shows the network architecture of the lightweight deep neural network HGL-Net. Detailed Implementation

[0017] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0018] This embodiment provides a high-frequency guided lightweight schlieren image reconstruction method for the flow field of a scramjet engine. This embodiment designs an end-to-end high-frequency guided lightweight deep neural network HGL-Net. To break down the barrier between model lightweighting and reconstruction accuracy, this network makes core innovations based on a U-Net-like encoding and decoding framework: the encoder extracts global high-order semantics with extremely low parameter overhead through simplified feature projection and layer-by-layer downsampling; the decoder abandons conventional upsampling operators and innovatively introduces a high-frequency guided composite upsampling mechanism. This composite upsampling mechanism, in conjunction with skip connections, not only efficiently compensates for information loss during dimensionality reduction but also accurately and efficiently restores the local high-frequency details and spatial resolution of the schlieren image with relatively small parameter consumption. Figure 1 Specifically, it includes the following: S1. Obtain one-dimensional sparse pressure sequence data of the scramjet engine wall; S2. Input the one-dimensional sparse compressed sequence data into the projection layer for projection linear transformation to obtain the initial two-dimensional feature tensor, thereby establishing the spatial topological basis required for subsequent convolution operations. The projection layer is represented as follows: In the formula, This represents the input feature vector. for 3D real space, The input dimension for one-dimensional sparse stress sequence data; The weight matrix represents the mapping relationship from the input space to the output space. The output feature dimension after linear transformation of the projection layer; This represents the bias vector. for 3D real space; This represents the output vector after projection.

[0019] S3. Input the initial two-dimensional feature tensor into the multi-scale feature encoder of the lightweight deep neural network HGL-Net. After multi-stage cascaded residual feature extraction and layer-by-layer downsampling, intermediate features and deep global semantic features at each scale are obtained. In some embodiments, the multi-scale feature encoder employs a multi-stage cascaded downsampling topology, aiming to progressively filter out redundant local noise and aggregate deep, high-order macroscopic semantics through hierarchical feature space compression. The intermediate features extracted at each scale in this stage are synchronously fed forward to the decoding network (spatial decoder) via skip connections to maintain feature scale consistency across stages and provide rich underlying structural references for the spatial decoder's reconstruction process.

[0020] In one specific embodiment, to achieve an optimal balance between feature extraction accuracy and model lightweighting, the multi-scale feature encoder alternately stacks the residual feature extraction module (ResDConv) and the max pooling operator. For example... Figure 2 As shown, the ResDConv module introduces a lightweight two-branch residual topology: its main branch connects two basic CBS (Conv-BatchNorm-SiLU) feature extraction modules to perform nonlinear semantic mapping. This module can be specifically represented as follows: In the formula, The feature tensor is the final output of this module. SiLUIntroducing nonlinearity into the network enables the model to fit complex function mappings. BatchNorm To standardize the results of convolution and make their distribution more stable; Conv This represents convolution, used to extract local spatial features from an image; This is the intermediate feature tensor output by the convolution operation. Feature data within the current batch b The mean, Feature data within the current batch b variance To ensure numerical stability, the smallest positive integers with a denominator of 0 are placed. Learnable scaling parameters for batch normalization operations. Learnable translation parameters for batch normalization.

[0021] Side branches are dimension-aligned using single-layer convolutions and element-wise added to the outputs of the main branches. This local residual design not only effectively mitigates gradient vanishing during forward propagation in deep networks but also significantly enhances the expressive richness of flow field features with extremely low parameter increments. ResDConv does not alter the scale of the data; it only aims to enrich the feature maps and broaden the receptive field.

[0022] Downsampling; the subsequent max-pooling layer performs spatial downsampling through a 2×2 non-overlapping sliding window. Given the input feature map. ,in, The number of pixels or feature grids in the vertical direction of the input feature map. The number of pixels or feature grids in the horizontal direction of the input feature map; in channel c and output space coordinates The pooling operation at a certain point can be formalized as follows: In the formula, This is the feature value of the output feature map at the c-th channel and spatial coordinates (m, n) after the pooling operation; This represents the local offset along the vertical direction within the pooling window. This represents the local offset along the horizontal direction within the pooling window. The input feature map is located at the coordinates (2m+p, 2n+q) in the c-th channel of the corresponding pooling window. This operation extracts the strongest activation features within the local receptive field, proportionally reducing the spatial resolution of the output feature map. This mechanism not only significantly reduces the computational overhead of subsequent networks but also robustly expands the global receptive field of deep feature maps, thus laying a solid semantic prior for the accurate reconstruction and mapping of the macroscopic morphology of physical fields.

[0023] After multi-stage cascaded downsampling, the multi-scale feature encoder finally outputs deep global semantic features. The input data sequentially passes through four sets of cascaded "feature extraction and downsampling" structures. With the spatial resolution decreasing layer by layer and the channel dimension increasing proportionally, the model gradually strips away redundant details at the bottom layer, ultimately compressing the original information into a deep feature tensor (deep global semantic features) with a shape of 1024×4×16. This tensor is located in the bottleneck layer of the network architecture, containing rich global structural information and abstract features. Subsequently, the deep global semantic features are used as the core carrier for schlieren image reconstruction and input into the subsequent decoder network, providing crucial global semantic guidance for the accurate layer-by-layer reconstruction of the physical field.

[0024] S4. Input the deep global semantic features into the spatial decoder of the lightweight deep neural network HGL-Net and perform multi-stage cascaded decoding. In each decoding stage, the deep decoding features output by the previous decoding stage are upsampled through a high-frequency guided convolution module. The upsampled features are then aligned with the intermediate features of the corresponding scale in the encoder and concatenated at the channel level to complete cross-scale feature aggregation. In some embodiments, the spatial decoder performs an exact inverse mapping from a low-dimensional space containing global priors to a high-resolution two-dimensional physical field. When dealing with physical scenes possessing multi-scale features and complex boundary conditions, traditional dense predictive image generation paradigms often inevitably fall into bottlenecks of computational overload and a surge in the number of parameters. To overcome this limitation, this embodiment proposes and constructs a lightweight cascaded decoding architecture. This architecture deeply integrates symmetrical cross-layer skip connections, aiming to progressively recover spatial resolution while achieving refined fusion of multi-order semantics.

[0025] In one specific embodiment, in the first t During the spatial reconstruction process in each decoding stage, the network not only expands the spatial dimension of the deep abstract features from the previous layer, but also strictly aligns and concatenates them with the shallow features at the corresponding perceptual scale at the multi-scale feature encoder. This cross-scale semantic alignment and feature aggregation mechanism can be represented as: In the formula, For the first Deep decoding features of hierarchical output, express Deep decoding features of hierarchical output; The intermediate feature represents the shallow feature response of the same resolution level of the multi-scale feature encoder. Its internal structure has not undergone excessive spatial compression, thus preserving extremely rich fine-grained coordinate and physical boundary information. This represents the upsampling operator in the process of reconstructing spatial resolution. This upsampling operator corresponds to the high-frequency guided convolution module mentioned below, which implements the upsampling operation. This is a cascaded nonlinear feature transformation network. After feature concatenation is completed, the cascaded nonlinear feature transformation network... It will perform deep feature extraction on the aggregated tensor to eliminate the aliasing effect that may be introduced by the upsampling operation and promote the deep fusion of features at different levels within the local receptive field.

[0026] This cross-scale feature fusion mechanism demonstrates key advantages in both the forward inference and backward optimization processes of the network. In the forward propagation stage, skip connections construct a physical path from shallow features to deeper layers, effectively compensating for the irreversible loss of local high-frequency topological information caused by continuous downsampling. In the backward propagation stage, this mechanism provides an additional backpropagation path for the error gradient, significantly alleviating the gradient vanishing problem that is prone to occur in deep networks and improving the stability of model training. In summary, this spatial decoder, by enhancing the reuse of multi-scale features, successfully achieves high-fidelity reconstruction of macroscopic physical fields under strict control of parameter quantity and computational cost.

[0027] In one specific embodiment, a high-frequency guided convolution module is provided; To address the spatial dimensionality expansion during decoding, this embodiment abandons the computationally intensive deconvolution operator, which is prone to chessboard effects, and proposes and integrates a lightweight High-Frequency Guided Convolution (HFGC) module to perform the upsampling task. Based on a content-aware feature reconstruction paradigm, this module innovatively constructs a dual-branch parallel dynamic upsampling topology, as detailed below. Figure 2 As shown: In the dynamic kernel prediction backbone, a high-frequency prior extraction mechanism based on the discrete Laplacian operator is introduced to capture and amplify microscopic gradient mutations and structural boundaries in the physical field. This is given by the feature tensor compressed via 1×1 convolution. The network employs parameter-free, depthwise separable convolutions to approximate the spatial second derivative channel-by-channel. For the spatial semantic features of the c-th feature channel... The high-frequency edge features of its c-th feature channel The extraction process can be represented as: In the formula, It is an isotropic 8-adjacent Laplace filter kernel; This is a two-dimensional discrete convolution operation.

[0028] in, For an isotropic 8-adjacent Laplace filter kernel, its weight matrix is ​​defined as: This Laplace filter kernel, with its extremely strong second-order differential sensitivity to high-frequency spatial signals, can efficiently extract pure topological edge priors without introducing any additional learnable parameters. Subsequently, the feature tensor... With this high-frequency prior Tight stitching is performed along the channel dimension, and dynamic reassembly weights that are highly sensitive to local topology are generated through sub-pixel convolution and spatial normalization. .

[0029] In the feature reconstruction auxiliary branch, the input features are sequentially subjected to nearest neighbor interpolation upsampling and spatial expansion operations, reshaping them into a context tensor containing local receptive field information. Ultimately, instead of using the traditional element-by-element addition, the two feature tables are deeply fused through a dynamic weighted aggregation mechanism. In the formula, The value of the final output feature tensor at a specific channel c and a specific pixel position (i, j); This represents the side length of the local aggregation window; It is then traversing this K 2 A loop variable of pixels; The coordinates of the spatial dimensions for traversing the entire feature map.

[0030] This upsampling paradigm not only seamlessly injects high-frequency physical boundary priors into the feature reconstruction process, but also significantly enhances the network's ability to resolve complex evolutionary details and preserve topology with extremely low parameter redundancy.

[0031] S5. Based on the cross-scale feature aggregation results, a high-resolution scramjet engine flow field schlieren image is reconstructed through multi-stage cascaded decoding of the spatial decoder.

[0032] The experimental results and analysis include the following: Experimental Environment: The model of this invention (lightweight deep neural network HGL-Net) was trained and evaluated on the same GPU (NVIDIA RTX 4090 24GB). The model was built using PyTorch 2.2 + cu118 framework on an Ubuntu 20.04 system equipped with an i9-14900KF CPU and 64GB of memory, and the corresponding code was run using a Python 3.10.4 interpreter. The experiment used the Adam optimizer to update the model parameters, with an initial learning rate set to... The weight decay coefficient is set to The learning rate employs a stepped decay strategy, multiplying by a decay factor of 0.9 every 20 training epochs. The batch size is set to 128, and training lasts for 100 epochs. Evaluation metrics: To comprehensively and rigorously evaluate the model's reconstruction performance, four core benchmark metrics are introduced for multi-dimensional assessment: Peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) focus on measuring the signal clarity of the reconstructed field at the global pixel level and the consistency of local structure and texture; while root mean square error (RMSE) and mean absolute error (MAE) are used to accurately quantify the absolute numerical deviation between the predicted value and the actual physical quantity.

[0033] The higher the PSNR value, the better the model suppresses reconstruction distortion and the higher the signal fidelity.

[0034] The theoretical range of SSIM is [0,1]. The closer its value is to 1, the more perfect the network's reproduction of complex local spatial structures such as microscopic flow field boundaries and vortices is.

[0035] The smaller the RMSE value, the higher the accuracy of the model's overall numerical fit to the target physical field.

[0036] MAE has the same dimensions as the measured physical quantity, and can directly and objectively reflect the average reliability of the model's predicted values ​​in the macroscopic domain.

[0037] The experiment specifically includes the following models: Heuristic Guided Learning Multi-Scale Fusion Neural Network (HGL-MSFNN). Data-driven hybrid neural network model (DDHNN); Multipath flow field prediction convolutional neural network (MFFPCNN); Lightweight Multi-View Domain Adaptive Generative Network (Light MV-DAGN); Multi-scale local-global feature grouping network (MSCNN_LGFG); A multi-scale attention algorithm with deblurring properties (DB-MSAN); U-Net; Quantitative analysis, as detailed below: To comprehensively and objectively evaluate the overall performance of each network model in the flow field reconstruction task, this invention uses four indicators to evaluate image quality and regression accuracy: peak signal-to-noise ratio (PSNR), structural similarity (SSIM), root mean square error (RMSE), and mean absolute error (MAE), as well as the number of model parameters to evaluate network complexity. Quantitative tests were conducted on all comparison models, as shown in Table 1.

[0038] Table 1 Comparison of quantitative analysis results of different models Models such as MFFPCNN, DDHNN, and HGL-MSFNN exhibit low accuracy across various metrics. Their PSNR hovers around 15, SSIM is less than 0.4, and RMSE and MAE error terms are high. This indicates that these networks are severely lacking in feature representation capabilities when facing complex nonlinear mapping tasks across dimensions, failing to effectively recover the spatial structure and pixel-level intensity of high-dimensional images. In contrast, models like U-Net, Light MV-DAGN, and DB-MSAN achieve significant leaps in reconstruction accuracy. The DB-MSAN model, in particular, which performed second best, achieved a PSNR of 21.49 and an SSIM of 0.674. However, this improvement in accuracy often comes at the cost of computational resources. DB-MSAN has a parameter count as high as 147.48 MB, while U-Net and LightMV-DAGN have parameter counts of 76.5 MB and 110 MB, respectively. This presents significant computational and storage bottlenecks in practical engineering deployments and high-frequency real-time flow field monitoring applications.

[0039] The model proposed in this invention achieved the best results across all five evaluation metrics. Its PSNR (21.957) and SSIM (0.687) were the highest in the field, representing average improvements of 18.16% and 33.4%, respectively. Simultaneously, RMSE (20.643) and MAE (13.421) were compressed to the lowest levels, with average reductions of 34.86% and 36.77%, respectively. More notably, while achieving the highest reconstruction accuracy, the model's parameter size was significantly optimized to only 31.58 MB, an average reduction of 86.73%. Through efficient feature extraction, this model not only perfectly handles complex reconstruction tasks ranging from sparse sensor signals to high-dimensional flow field images but also meets the requirements for extremely simple and lightweight deployment.

[0040] Although specific embodiments of the invention have been described in detail with reference to the accompanying drawings, this should not be construed as limiting the scope of protection of this patent. Various modifications and variations that can be made by a person skilled in the art without inventive effort within the scope described in the claims still fall within the scope of protection of this patent.

Claims

1. A high-frequency guided method for reconstructing the schlieren image of a lightweight scramjet engine flow field, characterized in that, Includes the following steps: S1. Obtain one-dimensional sparse pressure sequence data of the scramjet engine wall; S2. The one-dimensional sparse pressure sequence data is transformed by projection to generate an initial two-dimensional feature tensor; S3. Input the initial two-dimensional feature tensor into a multi-scale feature encoder, and after multi-stage cascaded residual feature extraction and layer-by-layer downsampling, obtain intermediate features and deep global semantic features at each scale. S4. Input the deep global semantic features into the spatial decoder and perform multi-stage cascaded decoding. In each decoding stage, the deep decoding features output by the previous decoding stage are upsampled through a high-frequency guided convolution module. The upsampled features are then aligned with the intermediate features of the corresponding scale in the encoder and concatenated at the channel level to complete cross-scale feature aggregation. S5. Based on the cross-scale feature aggregation results, a high-resolution scramjet engine flow field schlieren image is reconstructed through multi-stage cascaded decoding of the spatial decoder.

2. The high-frequency guided schlieren image reconstruction method for a lightweight scramjet engine according to claim 1, characterized in that, S2 specifically includes: One-dimensional sparse pressure sequence data is input into the projection layer for projection linear transformation to obtain the initial two-dimensional feature tensor.

3. The high-frequency guided schlieren image reconstruction method for a lightweight scramjet engine according to claim 1, characterized in that, In S3, residual features are extracted through a residual feature extraction module. The residual feature extraction module includes a lightweight dual-branch residual topology. The main branch of the dual-branch residual topology is connected in series with two CBS feature extraction modules, and the side branches of the dual-branch residual topology are single-layer convolutions. The CBS feature extraction module performs nonlinear semantic mapping on the input features. The side branches are dimension-aligned by a single-layer convolution and then added element-wise with the output of the main branch to obtain the output feature map.

4. The high-frequency guided schlieren image reconstruction method for a lightweight scramjet engine according to claim 3, characterized in that, In S3, downsampling includes: The output feature map of the residual feature extraction module is input into the max pooling layer. The max pooling layer performs spatial dimension downsampling through a 2×2 non-overlapping sliding window to extract the strongest activation features in the local receptive field, so that the spatial resolution of the output feature map is reduced proportionally, thus obtaining intermediate features. After multi-stage cascaded downsampling, the multi-scale feature encoder finally outputs deep global semantic features.

5. The high-frequency guided schlieren image reconstruction method for a lightweight scramjet engine according to claim 1, characterized in that, In S4, the high-frequency guided convolution module is a dual-branch parallel dynamic upsampling topology, including a dynamic kernel prediction backbone branch and a feature recombination auxiliary branch. The dynamic kernel prediction backbone branch extracts high-frequency edge priors of the flow field, and the feature recombination auxiliary branch generates a context tensor containing local receptive field information. Finally, a dynamic weighted aggregation mechanism is used to fuse the output of the dynamic kernel prediction backbone branch and the output of the feature recombination auxiliary branch to obtain the features after upsampling by the high-frequency guided convolution module, which is the dynamic weighted fusion result.

6. The high-frequency guided schlieren image reconstruction method for a lightweight scramjet engine according to claim 5, characterized in that, The dynamic kernel prediction backbone branch extracts high-frequency edge priors of the flow field using the discrete Laplace operator, which is expressed as: In the formula, For the high-frequency edge features of the c-th feature channel, For the spatial semantic features of the c-th feature channel, It is an isotropic 8-adjacent Laplace filter kernel; This is a two-dimensional discrete convolution operation; Then, the feature tensor High-frequency edge features Tight stitching is performed along the channel dimension, and dynamic recombination weights are generated through sub-pixel convolution and spatial normalization. .

7. The high-frequency guided schlieren image reconstruction method for a lightweight scramjet engine according to claim 6, characterized in that, In the feature reconstruction auxiliary branch, the input features are sequentially subjected to nearest neighbor interpolation upsampling and spatial expansion operations to reshape them into a context tensor containing local receptive field information. .

8. The high-frequency guided schlieren image reconstruction method for a lightweight scramjet engine according to claim 7, characterized in that, A dynamic weighted aggregation mechanism is adopted to dynamically reorganize the weights of the predicted backbone branches by the dynamic kernel. Context tensor of feature reorganization auxiliary branch The fusion process yields a dynamically weighted fusion result, which is represented as follows: In the formula, The value of the final output feature tensor at a specific channel c and pixel position (i, j); This represents the side length of the local aggregation window; It is traversing K 2 A loop variable of pixels.

9. The high-frequency guided schlieren image reconstruction method for a lightweight scramjet engine according to claim 8, characterized in that, In step S4, the upsampled features are aligned with the intermediate features of the corresponding scale in the encoder and concatenated at the channel level to complete cross-scale feature aggregation, which is expressed as follows: In the formula, For the first Deep decoding features of hierarchical output, express Deep decoding features of hierarchical output; The intermediate feature represents the shallow feature response at the same resolution level of the multi-scale feature encoder; This represents the upsampling operator in the process of reconstructing spatial resolution. This upsampling operator corresponds to the high-frequency guided convolution module, which implements the upsampling operation. It is a cascaded nonlinear feature transformation network.

Images