A container-based visible light and infrared image super-resolution reconstruction method

Abstract

This invention discloses a container-based super-resolution reconstruction method for visible light and infrared images, relating to the field of remote sensing technology. The method includes: a user submitting a task containing image and magnification parameters via a front-end; a back-end generating a task number and pushing it into a scheduling queue; a scheduling module allocating the task to an idle container; the container parsing the parameters and loading the corresponding super-resolution model, which includes a Local Texture Estimator (LTE) model and a Dual Regression Learning Network (DRN) model; reconstruction using mixed-precision inference; and uploading the results to a storage module, updating the task status, and logging the results for user front-end querying. This method achieves fully automated, containerized, and parallel processing from task submission to result return. This invention can achieve high-resolution reconstruction at arbitrary magnification for multi-source images such as visible light and infrared, and supports high-concurrency task scheduling and model inference in a containerized environment.

Description

Technical Field

[0001] This invention relates to the field of remote sensing technology, and in particular to a container-based super-resolution reconstruction method for visible light and infrared images. Background Technology

[0002] The rapid development of remote sensing technology has significantly improved the spatiotemporal resolution and coverage of Earth observation, providing crucial support for refined decision-making in fields such as ecological monitoring and disaster assessment. However, limited by sensor performance and imaging conditions, remote sensing images generally suffer from insufficient spatial resolution and blurred details, hindering high-precision analysis and applications.

[0003] Improving image quality primarily relies on two paths: hardware optimization and software enhancement. Hardware methods directly improve the quality of raw data by modifying optical systems and detectors, but they are costly, difficult to upgrade, and limited in large-scale deployment. Software enhancement methods, represented by deep learning, reconstruct high-resolution images through algorithms and have become the mainstream research direction. Based on models such as convolutional neural networks and generative adversarial networks, they can learn the mapping relationship from low resolution to high resolution from a large number of samples, significantly improving image sharpness and detail without modifying hardware, and have advantages such as low cost and strong adaptability.

[0004] While existing deep learning super-resolution methods have made progress, they still face significant limitations. At the algorithmic level, most models are designed only for fixed integer magnification, failing to flexibly support reconstruction needs at arbitrary or fractional magnifications. Furthermore, training is typically based on idealized bicubic downsampling simulations, which differ from the complex real-world imaging degradation process, resulting in insufficient generalization ability and susceptibility to texture artifacts and edge blurring in practical applications. At the system deployment level, existing solutions largely rely on traditional remote sensing processing software or single-machine inference frameworks, lacking effective support for GPU parallel computing and large-scale concurrent tasks. Although some containerization technologies have been introduced, lightweight scheduling and caching optimizations for deep learning inference tasks have not yet been implemented, leading to low resource utilization and insufficient response efficiency. Summary of the Invention

[0005] The purpose of this invention is to provide a container-based method for super-resolution reconstruction of visible light and infrared images, aiming to solve or improve at least one of the aforementioned technical problems.

[0006] To achieve the above objectives, the present invention provides the following solution: A container-based super-resolution reconstruction method for visible light and infrared images, comprising: Users log in to the Web platform through the front-end interactive module, select the visible light or infrared image to be processed, set the super-resolution parameters, and submit a task request. When the task management module receives a request, it generates a unique task number, writes the task information into the database, and pushes it to the task scheduling queue. When the task scheduling module detects an idle inference container, it assigns the task information to the container instance. In the containerized inference module, task parameters are parsed, the corresponding super-resolution model is loaded according to the task requirements, and a high-resolution image is constructed using a hybrid precision inference strategy; the super-resolution model includes a local texture estimator LTE model and a dual regression learning network DRN model. The reconstructed high-resolution images are uploaded to the data storage module, and the task management module updates the task status and records logs, allowing users to query the reconstruction results through the front-end module.

[0007] Optionally, the super-resolution model employs an LRU caching mechanism and a global thread lock mechanism during loading, and the specific loading process includes: First, the task parameters are parsed to determine the type of model to be loaded. Then, the LRU model caching mechanism is used to check whether the weights have been loaded. If the model exists in the cache, it is called directly. If the cache is empty, the model is loaded from the weight file and cached in memory.

[0008] Optionally, the process of constructing the high-resolution image includes: The input image is preprocessed using the loaded super-resolution model, converted into standardized tensor data, and coordinates and cell grid input are generated according to the set magnification to update the current super-resolution model. The updated model is run in a GPU environment for super-resolution reconstruction, and a mixed-precision inference strategy is used to construct high-resolution images.

[0009] Optionally, the local texture estimator LTE model includes an encoder, a local texture estimation module, and a decoder; wherein the local texture estimation module includes a frequency domain feature extraction layer, a multi-branch residual convolutional unit, and a channel attention mechanism.

[0010] Optionally, the training of the local texture estimator LTE model employs a multi-loss joint optimization strategy that combines pixel reconstruction loss, perception loss, and edge preservation loss.

[0011] Optionally, the dual regression learning network (DRN) model includes a generator G, a degenerate mapping network R, and a discriminator D, forming a bidirectional mapping structure from low resolution to high resolution and then back to low resolution.

[0012] Optionally, the training of the dual regression learning network (DRN) model employs a joint optimization strategy that combines content loss, perceptual loss, adversarial loss, and dual loss.

[0013] Optionally, the discriminator D adopts a PatchGAN structure for adversarial discrimination in local regions.

[0014] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: This invention discloses a container-based super-resolution reconstruction method for visible light and infrared images. The method includes: a user submitting a task containing image and magnification parameters via a front-end; a back-end generating a task number and pushing it into a scheduling queue; a scheduling module allocating the task to an idle container; the container parsing the parameters and loading the corresponding super-resolution model, which includes a Local Texture Estimator (LTE) model and a Dual Regression Learning Network (DRN) model; reconstruction using mixed-precision inference; and the result being uploaded to a storage module, with the task status updated and logs recorded for user front-end querying. This method achieves fully automated, containerized, and parallel processing from task submission to result return. This invention can achieve high-resolution reconstruction at arbitrary magnification for multi-source images such as visible light and infrared, and supports high-concurrency task scheduling and model inference in a containerized environment. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of the overall process of the present invention; Figure 2 This is the task interaction diagram in this embodiment; Figure 3 This is a flowchart of image super-resolution reconstruction based on a local texture estimator in this embodiment; Figure 4 This is a flowchart of image super-resolution reconstruction based on generative adversarial networks and dual learning in this embodiment. Figure 5 These are the visible light image and infrared image corresponding to the original image in this embodiment; Figure 6 These are the visible light image and infrared image corresponding to the 2x super-resolution image in this embodiment; Figure 7 These are the visible light image and infrared image corresponding to the 4x super-resolution image in this embodiment. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] The purpose of this invention is to provide a container-based method for super-resolution reconstruction of visible light and infrared images, aiming to solve or improve at least one of the aforementioned technical problems.

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0020] like Figure 1 As shown, this invention provides a container-based super-resolution reconstruction method for visible light and infrared images, comprising: S1. Users log in to the Web platform through the front-end interactive module, select the visible light or infrared image to be processed, set the super-resolution parameters, and submit a task request.

[0021] S2. When the task management module receives the request, it generates a unique task number, writes the task information into the database, and pushes it to the task scheduling queue.

[0022] S3. When the task scheduling module detects an idle inference container, it assigns the task information to the container instance.

[0023] S4. In the containerized inference module, parse the task parameters, load the corresponding super-resolution model according to the task requirements, and construct the high-resolution image using a hybrid precision inference strategy; the super-resolution model includes the Local Texture Estimator (LTE) model and the Dual Regression Learning Network (DRN) model.

[0024] As a specific implementation method, the super-resolution model employs an LRU caching mechanism and a global thread lock mechanism during loading.

[0025] The specific loading process includes: First, the task parameters are parsed to determine the type of model to be loaded. Then, the LRU model caching mechanism is used to check whether the weights have been loaded. If the model exists in the cache, it is called directly. If the cache is empty, the model is loaded from the weight file and cached in memory.

[0026] As one specific implementation method, the process of constructing the high-resolution image includes: The input image is preprocessed using the loaded super-resolution model, converted into standardized tensor data, and coordinates and cell grid input are generated according to the set magnification to update the current super-resolution model. Then, the updated model is run in the GPU environment to perform super-resolution reconstruction, and a mixed precision inference strategy is used to construct high-resolution images.

[0027] As a specific implementation, the local texture estimator LTE model includes an encoder, a local texture estimation module, and a decoder; wherein, the local texture estimation module includes a frequency domain feature extraction layer, a multi-branch residual convolutional unit, and a channel attention mechanism. The training of the local texture estimator LTE model employs a multi-loss joint optimization strategy that combines pixel reconstruction loss, perceptual loss, and edge preservation loss.

[0028] As a specific implementation, the dual regression learning network (DRN) model includes a generator G, a degenerate mapping network R, and a discriminator D, forming a bidirectional mapping structure from low resolution to high resolution and back to low resolution. The training of the DRN model employs a joint optimization strategy that combines content loss, perceptual loss, adversarial loss, and dual loss. Specifically, the discriminator D uses a PatchGAN structure for adversarial discrimination in local regions.

[0029] S5. Upload the reconstructed high-resolution image to the data storage module, update the task status and record logs using the task management module, and allow users to query the reconstruction results through the front-end module.

[0030] Based on the above technical solution, the following embodiments are provided.

[0031] In this embodiment, we address the issue from three levels: model structure, degradation constraints, and system deployment. We construct an arbitrary-rate reconstruction network, a dual-learning constraint network, and a containerized inference platform, achieving a comprehensive breakthrough from algorithm performance to system usability. Specific steps include: (1) Container-based online web service for super-resolution remote sensing images This invention provides a container-based method and system for super-resolution reconstruction of visible light and infrared images, belonging to the field of intelligent image processing and artificial intelligence applications. This method can achieve high-resolution reconstruction at arbitrary magnification for multi-source images such as visible light and infrared, and supports high-concurrency task scheduling and model inference in a containerized environment.

[0032] The overall system architecture of this invention consists of a front-end interaction module, a task management module, a containerized inference module, a data storage module, a task scheduling and caching module, and a logging and monitoring module. The system uses container services as its core and automates the entire process from image selection, parameter setting, model inference to result display through a task flow mechanism.

[0033] Users first log in to the web platform through the front-end interaction module, select visible light or infrared images to be processed from the system's image database, set the corresponding super-resolution parameters (such as ×2, ×4, ×8, or any decimal magnification), and submit a task request. The front-end module is responsible for collecting task information, including task name, image type, and magnification, and then packaging the request information and sending it to the task management module. The task management module is deployed on a Java backend and is used for task initialization and scheduling control. After receiving the request, the system automatically generates a unique task ID, writes the task parameters and status information to the database, and pushes it to the task scheduling queue to await distribution. When the Redis queue detects an idle inference container, the task scheduling module immediately assigns the task to the corresponding container instance to achieve dynamic load balancing of parallel tasks.

[0034] like Figure 2 As shown, after task distribution, the containerized inference module enters the working state. This module, built on Python and the Flask framework, is the core execution unit of the system, responsible for model loading, inference computation, and result feedback. When the container receives a task, it first parses the task parameters, determines the required model type (integrated local texture estimator LTE model or dual regression learning network model DRN), and checks whether the weights have been loaded using the LRU model caching mechanism. If the model exists in the cache, it is directly called; if the cache is empty, the model is loaded from the weight file and cached in memory. To prevent memory conflicts caused by multiple threads loading models simultaneously, the system introduces a global thread lock mechanism to achieve resource mutual exclusion protection during model loading and release.

[0035] After the model is loaded, the container performs image preprocessing. The input images are uniformly converted into standardized tensor data, and corresponding coordinates and cell grid inputs are generated according to the user-defined magnification. During inference, the system runs the deep learning model in a GPU environment, employing a mixed-precision (FP16) inference strategy to improve GPU memory utilization and inference speed. After the model completes reconstruction, it outputs high-resolution image results, which are saved in PNG or TIFF format to the container's local results directory.

[0036] Subsequently, the container automatically uploads the result files to the data storage module. This module includes a relational database and an object storage system (MinIO). The former records task metadata, status, and index paths, while the latter stores the reconstructed image files. Upon receiving the JSON response from the container, the task management module updates the task status and records inference time and system logs. Simultaneously, the logging and monitoring module records and monitors the model loading status, memory usage, and container operation in real time throughout the inference process, ensuring system traceability and high reliability.

[0037] Users can query task execution results through the front-end interactive module. The system supports viewing the task list, previewing results, and comparing them. The front-end can retrieve the result index from the database based on the task ID and display the reconstructed images in real time on the page. It supports comparison between the original image and the super-resolution results, zoom switching, and quality assessment.

[0038] Through the above process, this invention achieves full automation from task creation, distribution, model inference to result display, constructing a scalable and reusable intelligent remote sensing image processing platform. Compared with existing offline algorithm execution methods, this invention implements a containerized structure design at the system level that separates model loading from task execution, allowing multiple tasks to be executed independently in different container instances; through model weight caching and thread lock mechanisms, it significantly improves GPU resource utilization and task response speed; at the same time, the system adopts a service-oriented deployment approach, which can be quickly expanded to a multi-node distributed environment, supporting unified reconstruction of images from different sensors and multiple sources.

[0039] (2) Super-resolution reconstruction of visible light and infrared images based on local texture estimator like Figure 3 As shown, this invention introduces a local texture estimator network model into a containerized system architecture to achieve arbitrary-magnification super-resolution reconstruction of visible light and infrared images. This model can learn the nonlinear mapping relationship from low-resolution to high-resolution images in a continuous spatial coordinate domain, thereby recovering details lost due to imaging degradation and achieving high-fidelity restoration of texture information. Compared with traditional convolutional neural networks with fixed magnification, this invention, based on a local texture estimator model, achieves arbitrary-magnification mapping in continuous space through implicit coordinate representation mechanisms and frequency domain feature modeling, effectively overcoming the limitation of traditional super-resolution models that only support integer-magnification reconstruction.

[0040] In the data preparation phase, visible light or infrared images are selected from the remote sensing image database as high-resolution samples (HR), and corresponding low-resolution images (LR) are generated using Gaussian blur, noise overlay, and downsampling operations to construct a low-to-high resolution matching dataset. The data samples cover various scales and different noise conditions to enhance the model's generalization ability. The model takes the low-resolution images as input and first extracts multi-scale feature tensors through the encoder module. Simultaneously, the system generates a continuous coordinate grid (coord) and pixel unit size (cell) matrix based on the spatial distribution of the input images, enabling the network to achieve continuous prediction at any coordinate point. This constructed "feature-coordinate-cell" triplet input format provides a continuous spatial foundation for subsequent implicit feature mapping.

[0041] During the feature modeling phase, a local texture estimation module (LTE Module) is embedded in the middle of the network. This module comprises three sub-components: a frequency domain feature extraction layer, a multi-branch residual convolutional unit, and a channel attention mechanism. The frequency domain feature transformation layer captures high-frequency detail features of the image; the multi-branch residual convolutional unit fuses and interacts features at different scales to improve the structural consistency of local textures; and the channel attention mechanism adaptively adjusts the weights of each channel based on the feature response intensity, thereby highlighting significant texture regions and suppressing redundant features. Through the collaborative modeling of these three parts, the model can perform frequency domain estimation and reconstruction of local textures in implicit space, achieving detail enhancement and accurate edge restoration.

[0042] During the decoding and reconstruction stages, the model upsamples the fused features using a combination of deconvolution and linear interpolation, outputting a high-resolution image corresponding to the target magnification. Due to the introduction of an implicit coordinate input mechanism, the model can perform interpolation inference in continuous space, outputting results with arbitrary magnification ratios (including non-integer magnifications such as ×2.5, ×3.7, etc.) simply by adjusting the magnification parameter. During training, the model employs a joint optimization strategy of pixel reconstruction loss (L1), perceptual loss (LPIPS), and edge preservation loss. L1 loss maintains overall structural consistency, perceptual loss improves subjective visual quality, and edge loss suppresses blur and artifacts.

[0043] During the inference phase, the model loads the training weight file epoch-best.pth using the PyTorch framework and deploys it as a Flask service within the container. The system provides a unified REST interface / GetLocationMC / lte_sr, with parameters including image type (type: visible light / infrared), magnification (scale), and input file path (files[]). The Flask service automatically detects the computing environment and prioritizes GPUs for inference, while enabling mixed-precision (FP16) computation to improve memory utilization and inference efficiency. The generated results are saved in PNG format to the specified directory within the container, and the result path and runtime are returned in JSON format. After the container execution is complete, the system automatically uploads the output results to the object storage module (MinIO) and records the task information in the database to achieve task status tracking and result reuse.

[0044] Through the above structural design, the super-resolution reconstruction method based on a local texture estimator in this invention has significant advantages in model accuracy, reconstruction speed, and system adaptability. This model can support image reconstruction at arbitrary and fractional magnifications, adapting to various types of image data. With the help of the frequency domain feature estimation mechanism, the model excels in high-frequency detail recovery and texture fidelity. Simultaneously, the model structure is lightweight, computationally inefficient, and has a fast response speed during the inference phase, making it suitable for deployment in containerized environments for online concurrent inference. This method can be further extended to the joint reconstruction of multispectral and infrared fused images, providing strong technical support for high-precision image reconstruction and subsequent intelligent analysis.

[0045] (3) Super-resolution reconstruction of visible light and infrared images based on generative adversarial networks and dual learning like Figure 4 As shown, this invention proposes a super-resolution reconstruction method for remote sensing images based on generative adversarial networks and dual learning mechanisms. This method significantly improves the subjective visual quality of the reconstructed images while maintaining structural accuracy. By introducing bidirectional mapping constraints and adversarial optimization mechanisms, the model achieves significant improvements in realism, texture consistency, and stability. This method introduces degradation modeling and dual learning concepts into the traditional unidirectional super-resolution network, constructing a bidirectional mapping relationship from low resolution to high resolution and back to low resolution (Low→High→Low). This forms a closed-loop consistency constraint during the training phase, enabling the generator to maintain low-resolution spatial consistency while recovering high-frequency details.

[0046] In the data preparation phase, this invention selects visible light or infrared remote sensing images as high-resolution samples (HR) and generates low-resolution samples (LR) through various degradation kernel functions (including Gaussian blur, noise superposition, and random downsampling) to enhance the robustness of the model under complex degradation conditions. The training dataset contains sample pairs with multiple scales and noise levels to support the model's generalization ability under different resolutions and light sensitivity conditions.

[0047] In terms of network architecture design, the DRN model consists of three sub-modules: a generator G, a degenerate mapping network R, and a discriminator D. The generator G is responsible for mapping the input low-resolution image to a high-resolution space. Internally, it employs an architecture combining multi-layer residual convolutional blocks and upsampling modules, effectively capturing local texture information and reconstructing lost high-frequency details. The reconstructed image output by the generator not only preserves the original structure but also possesses high visual realism. The degenerate network R simulates the inverse mapping process from high resolution to low resolution, constructing a bidirectional constraint relationship between high and low resolution spaces by modeling the degradation of the generated image. This process enables the generator to be reversible while learning the reconstruction mapping function, ensuring that the network maintains structural consistency and numerical stability under complex degradation conditions. The discriminator D distinguishes the generated image from the real high-resolution image, employing a 70×70 PatchGAN architecture. It captures artifacts and texture inconsistencies through a local discrimination mechanism. The true / false probability signal output by the discriminator serves as adversarial feedback for the generator, guiding it to continuously improve the visual realism of the image.

[0048] During training, the model employs a joint loss optimization strategy. The overall loss function comprises three parts: content loss, dual loss, and adversarial loss. The content loss constrains the structural consistency between the generated image and the real image; the dual loss measures the difference between the degraded mapped image and the original low-resolution image, thus forming a two-way closed-loop consistency; and the adversarial loss, provided by the discriminator, enhances the naturalness and high-frequency detail quality of the generated image. When paired data is available, the generator performs joint optimization using both content loss and dual loss; when paired data is unavailable, the generator relies solely on the dual loss for learning, thus maintaining high reconstruction performance even with incomplete data. This multi-loss joint optimization strategy effectively prevents mode collapse and improves the model's convergence speed and training stability.

[0049] During the inference phase, only the generator G module is retained for actual reconstruction. The system loads the weight file drn_best.pth through a containerized Flask service and provides a unified interface / GetLocationMC / super_result, which can call the DRN model by specifying model parameters. The inference phase performs mixed-precision (FP16) computation in a GPU environment to improve memory utilization and runtime efficiency. The reconstructed images output by the generator are saved in PNG format and automatically uploaded to the object storage system (MinIO) by the container. Simultaneously, the task ID, inference time, and performance metrics (including PSNR and SSIM) are recorded in the database for system performance monitoring and subsequent analysis.

[0050] The DRN super-resolution reconstruction system of this invention constructs a unified architecture that integrates dual learning and adversarial training through the collaborative optimization of the generator, degenerate network, and discriminator, achieving simultaneous improvement in structural accuracy and subjective quality. Compared with existing unidirectional super-resolution methods, this invention has significant advantages in visual detail recovery, texture naturalness, and model generalization. Its bidirectional mapping mechanism ensures the stability and reversibility of the high-low resolution spatial mapping relationship, the adversarial optimization strategy enhances the visual realism of the generated results, and the multi-loss joint optimization achieves a balance between structure and perception. Relying on containerized deployment, this invention can be integrated with LTE models and deployed on the same online inference platform, enabling rapid switching and concurrent operation of multiple models, and possesses good engineering scalability and application promotion value.

[0051] Running the above process will yield the following results: Figures 5-7 The results of super-resolution reconstruction of visible light and infrared images are shown in Table 1. Super-resolution technical parameters (PSNR and SSIM) are shown in Table 1.

[0052] Table 1 Super-resolution PSNR and SSIM

[0053] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0054] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A container-based super-resolution reconstruction method for visible light and infrared images, characterized in that, include: Users log in to the Web platform through the front-end interactive module, select the visible light or infrared image to be processed, set the super-resolution parameters, and submit a task request. When the task management module receives a request, it generates a unique task number, writes the task information into the database, and pushes it to the task scheduling queue. When the task scheduling module detects an idle inference container, it assigns the task information to the container instance. In the containerized inference module, task parameters are parsed, the corresponding super-resolution model is loaded according to the task requirements, and a high-resolution image is constructed using a hybrid precision inference strategy; the super-resolution model includes a local texture estimator LTE model and a dual regression learning network DRN model. The reconstructed high-resolution images are uploaded to the data storage module, and the task management module updates the task status and records logs, allowing users to query the reconstruction results through the front-end module.

2. The container-based visible light and infrared image super-resolution reconstruction method according to claim 1, characterized in that, The super-resolution model employs an LRU caching mechanism and a global thread lock mechanism during loading. The specific loading process includes: First, the task parameters are parsed to determine the type of model to be loaded. Then, the LRU model caching mechanism is used to check whether the weights have been loaded. If the model exists in the cache, it is called directly. If the cache is empty, the model is loaded from the weight file and cached in memory.

3. The container-based visible light and infrared image super-resolution reconstruction method according to claim 1, characterized in that, The process of constructing the high-resolution image includes: The input image is preprocessed using the loaded super-resolution model, converted into standardized tensor data, and coordinates and cell grid input are generated according to the set magnification to update the current super-resolution model. The updated model is run in a GPU environment for super-resolution reconstruction, and a mixed-precision inference strategy is used to construct high-resolution images.

4. The container-based visible light and infrared image super-resolution reconstruction method according to claim 1, characterized in that, The local texture estimator LTE model includes an encoder, a local texture estimation module, and a decoder; wherein, the local texture estimation module includes a frequency domain feature extraction layer, a multi-branch residual convolutional unit, and a channel attention mechanism.

5. The container-based visible light and infrared image super-resolution reconstruction method according to claim 1, characterized in that, The training of the local texture estimator LTE model employs a multi-loss joint optimization strategy that combines pixel reconstruction loss, perceptual loss, and edge preservation loss.

6. The container-based visible light and infrared image super-resolution reconstruction method according to claim 1, characterized in that, The dual regression learning network (DRN) model includes a generator G, a degenerate mapping network R, and a discriminator D, forming a bidirectional mapping structure from low resolution to high resolution and then back to low resolution.

7. The container-based visible light and infrared image super-resolution reconstruction method according to claim 1, characterized in that, The training of the dual regression learning network (DRN) model employs a joint optimization strategy that combines content loss, perceptual loss, adversarial loss, and dual loss.

8. The container-based visible light and infrared image super-resolution reconstruction method according to claim 6, characterized in that, The discriminator D adopts the PatchGAN structure for adversarial discrimination in local regions.

Images