CN-121998823-A - Remote sensing image super-resolution reconstruction method based on multi-process conditional attention

CN121998823ACN 121998823 ACN121998823 ACN 121998823ACN-121998823-A

Abstract

The invention discloses a remote sensing image super-resolution reconstruction method based on multi-process condition attention, which relates to the technical field of remote sensing image processing, and comprises the steps of obtaining a data set according to high-resolution remote sensing image data; the method comprises the steps of carrying out multistage degradation treatment on a data set to obtain a low-resolution image, forming a training data pair by the low-resolution image and a high-resolution image corresponding to the low-resolution image in the data set, carrying out countermeasure training on a countermeasure network according to the training data pair to obtain a super-resolution reconstruction model, obtaining a low-resolution remote sensing image to be reconstructed, and reconstructing the low-resolution remote sensing image through the super-resolution reconstruction model to obtain a high-resolution remote sensing image. According to the invention, the real degradation process is simulated through multistage degradation and the frequency domain characteristic modulation is combined, so that the high-quality super-resolution reconstruction of the remote sensing image is realized through countermeasure training.

Inventors

HAN XIAOQI
GAO ENYU
ZHOU XIN
SUN PENG

Assignees

北京微纳星空科技股份有限公司

Dates

Publication Date: 20260508
Application Date: 20251204

Claims (10)

1. A remote sensing image super-resolution reconstruction method based on multi-process condition attention is characterized by comprising the following steps: s1, obtaining a data set according to high-resolution remote sensing image data; S2, forming a training data pair by the low-resolution image and a high-resolution image corresponding to the low-resolution image in the data set, and performing countermeasure training on a countermeasure network according to the training data pair to obtain a super-resolution reconstruction model; s3, obtaining a low-resolution remote sensing image to be reconstructed, and reconstructing the low-resolution remote sensing image through the super-resolution reconstruction model to obtain a high-resolution remote sensing image.
2. The method of claim 1, wherein the multi-stage degradation process comprises at least two stages of degradation processes, wherein each stage of degradation process comprises at least a blurring process, a scaling process, a noise process, and a compression process.
3. The remote sensing image super-resolution reconstruction method based on multi-process condition attention as set forth in claim 2, wherein the performing of the multi-stage degradation process on the data set to obtain a low-resolution image specifically comprises performing the two-stage degradation process on the data set to obtain a low-resolution image; The second stage degradation processing comprises coupling the sinc filter and the compression processing with preset probability, and is used for simulating ringing artifacts and blocky artifacts generated in the satellite imaging link and satellite compression transmission process.
4. A method of super-resolution reconstruction of a remote sensing image based on multi-process conditional attention as in any one of claims 1-3 wherein said multi-stage degradation process further comprises: Encoding degradation nuclear parameters adopted in the degradation process, and generating a degradation condition vector through a degradation encoder; Generating modulation parameters according to the degradation condition vector, and injecting the modulation parameters into a residual error dense block of the countermeasure network for realizing channel-by-channel affine modulation.
5. The method for reconstructing the super-resolution of the remote sensing image based on the multi-process conditional attention as set forth in claim 4, wherein the implementation of the degradation encoder comprises: encoding the degradation kernel parameters through a convolution layer, a ReLU activation function, a self-adaptive average pooling layer and a full connection layer to generate a degradation condition vector; and when the batch size of the degradation condition vector is not matched with the batch size of the residual error dense block output characteristic, expanding the degradation condition vector to a target batch size, so as to realize adaptive affine modulation based on degradation processing.
6. The method for reconstructing a super-resolution image of a remote sensing image based on multi-process conditional attention as recited in claim 1, wherein, The discriminator in the countermeasure network adopts a U-Net structure with spectrum normalization and is used for inhibiting artifacts in the countermeasure training process and stabilizing gradient propagation; The generator in the countermeasure network is constructed based on residual dense blocks and comprises a feature extraction module, wherein a frequency domain attention mechanism is embedded in each residual dense block, modulation parameters are received, and channel-by-channel affine transformation is carried out on a feature map output by the residual dense blocks, so that degenerate self-adaptive feature reconstruction is realized; The feature extraction module is used for carrying out feature extraction processing on any image in the training data pair to obtain an input feature map for the frequency domain attention mechanism; the frequency domain attention mechanism is used for performing frequency domain conversion processing on the input feature map.
7. The method for reconstructing a super-resolution of a remote sensing image based on multi-process conditional attention as set forth in claim 6, wherein said frequency domain conversion process specifically comprises: Converting the input feature map from a space domain to a frequency domain through two-dimensional fast Fourier transform to obtain a frequency spectrum feature map; Determining a frequency weight map of the input feature map based on the low frequency mask and the high frequency mask; Converting the frequency weight graph back to a space domain through two-dimensional inverse fast Fourier transform, extracting a real number part as a modulated space feature graph, and carrying out weighted fusion on the space feature graph and the input feature graph.
8. The remote sensing image super-resolution reconstruction system based on the multi-process condition attention is characterized by comprising a data degradation module, an countermeasure training module and a model reasoning module; The data degradation module is used for obtaining a data set according to the high-resolution remote sensing image data; The countermeasure training module is used for forming a training data pair by the low-resolution image and a high-resolution image corresponding to the low-resolution image in the data set, and performing countermeasure training on a countermeasure network according to the training data pair to obtain a super-resolution reconstruction model; the model reasoning module is used for acquiring a low-resolution remote sensing image to be reconstructed, and reconstructing the low-resolution remote sensing image through the super-resolution reconstruction model to obtain a high-resolution remote sensing image.
9. A computer device comprising a processor coupled to a memory, the memory having stored therein at least one computer program that is loaded and executed by the processor to cause the computer device to implement the method of any of claims 1 to 7.
10. A computer readable storage medium having stored therein at least one computer program that is loaded and executed by a processor to cause a computer to implement the method of any one of claims 1 to 7.

Description

Remote sensing image super-resolution reconstruction method based on multi-process conditional attention Technical Field The invention relates to the technical field of remote sensing image processing, in particular to a remote sensing image super-resolution reconstruction method based on multi-process condition attention. Background With the rapid development of remote sensing satellite technology, the acquired remote sensing images face higher requirements in terms of spatial resolution, spectral resolution and timeliness. However, complex degradation inevitably occurs during satellite imaging, including systematic blurring caused by imaging sensors, quantization bit depth limitations, overlapping effects caused by resampling, atmospheric noise, and compression artifacts resulting from satellite compression transmissions. These degradation factors tend to complex and spatially variable, resulting in limited accuracy in interpretation of conventional single images. Meanwhile, the paired high-resolution and low-resolution remote sensing data are difficult to acquire in practical application, the quantity is rare, and the general super-resolution model is difficult to adapt to the characteristics of abundant details, huge breadth and high quantization bit number of the remote sensing image, so that the fine application requirements of the remote sensing image in the fields of resource monitoring, disaster evaluation, national defense safety and the like are severely restricted. In recent years, super-resolution technology based on deep learning has become a mainstream direction, in which generation of a countermeasure network (GAN) is widely used in the field of remote sensing due to its advantages in image texture restoration and visual reality. The prior art generally adopts the following scheme that low-resolution training data are firstly constructed in a plurality of downsampling modes or existing paired remote sensing images are directly utilized, then a GAN network is constructed, the low-resolution data are received by a generating network in a training stage, the mapping relation from low resolution to high resolution is extracted and learned through multi-layer convolution characteristics, and optimized parameters are fed back on the basis of a discriminator, the discriminator compares a generated image with a real high-resolution image, the reliability of the generated image is evaluated through multi-layer convolution and pooling extraction characteristics, loss functions such as perception loss, counterloss and the like are calculated and counter-propagated, and a stable super-resolution model is obtained after multi-round countertraining. However, the prior art solutions described above have significant drawbacks. Firstly, a data construction mode based on single-stage downsampling is difficult to simulate a complex degradation process of a satellite in a full link from imaging and transmission to preprocessing, so that the input data and a real remote sensing scene have larger deviation, and the performance boundary of a model is directly influenced. Secondly, the general super-resolution model lacks a flexible adjustment mechanism when processing the feature features of the features with different frequency components, and the phenomena that the edges of the features are over-sharpened, the low-frequency region is smooth and the like and do not accord with the physical features of the remote sensing image often appear, so that the reconstruction effect on the complex features tends to be averaged, and the interpretation usability is insufficient. Third, the traditional model generally does not have the capability of processing high-quantization bit images (such as 12 bits and 16 bits) commonly used in the remote sensing field, and a great deal of detail information is lost in direct re-quantization processing. In addition, although the diffusion model has better reconstruction quality, the model is huge and has low inference speed, and the real-time requirement of on-board edge calculation is difficult to meet, while the existing GAN model is light and lacks an explicit modeling and self-adaptive processing mechanism for remote sensing degradation. In summary, the core contradiction faced by the existing super-resolution technology in the remote sensing field is that the reality of the data degradation simulation is insufficient, the self-adaptability of the model to the remote sensing characteristics is poor, and the algorithm efficiency and accuracy are unbalanced. Therefore, there is an urgent need for a remote sensing-specific super-resolution method that can explicitly utilize degradation information, adaptively allocate computing resources, and is compatible with high-bit quantization characteristics, so as to improve high-frequency information recovery accuracy and suppress artifacts in a real complex scene, and form a high-quality image product that meets the requirements of remote se