CN-122023121-A - Small satellite image high-efficiency super-resolution method and system based on key features and dynamic priori guidance

CN122023121ACN 122023121 ACN122023121 ACN 122023121ACN-122023121-A

Abstract

The invention relates to the technical field of remote sensing image processing and deep learning intersection, and discloses a small satellite image high-efficiency super-resolution method and system based on key features and dynamic priori guidance. The invention greatly reduces the occupation of platform resources while ensuring the quality of the reconstructed image, effectively realizes the balance of performance and efficiency, breaks through the bottleneck of small satellite imaging, and can be applied to remote sensing scenes such as environment monitoring, resource investigation and the like.

Inventors

ZHAO SHENGRONG
ZHAO ZHIBO
LIANG HU

Assignees

齐鲁工业大学(山东省科学院)

Dates

Publication Date: 20260512
Application Date: 20260213

Claims (10)

1. A small satellite image high-efficiency super-resolution method based on key features and dynamic priori guidance is characterized by comprising the following steps: s1, acquiring a low-resolution small satellite image to be processed, preprocessing the low-resolution small satellite image, and constructing an LR-HR image pair; s2, shallow feature extraction is carried out on the preprocessed low-resolution small satellite image, and a coarse feature map is obtained; S3, deep feature extraction is carried out on the crude feature map through a distillation module KDDB containing N key features connected in series and dynamically guided in prior, each KDDB is integrated with a partial channel cascade and distillation unit PCDM, a dynamically guided channel attention unit DPCA and a high-efficiency key feature perception self-attention unit KeySA, and a fine feature map is output after multi-scale feature extraction and fusion; s4, carrying out residual connection on the fine feature image and the coarse feature image, and outputting a super-resolution small satellite image with preset multiple after convolution processing and up-sampling processing.
2. The method for efficiently super-resolution of a small satellite image based on key features and dynamic prior guidance according to claim 1, wherein the preprocessing in step S1 comprises the steps of generating a low-resolution small satellite image from an original high-resolution small satellite image through downsampling, and constructing an LR-HR image pair; in step S2, shallow layer feature extraction is performed by using a 3×3 convolution layer, and the calculation formula of the coarse feature map is as follows: , Wherein, the In order to make the characteristic diagram thick, A3 x 3 convolution operation is shown, Is a preprocessed low-resolution small satellite image.
3. The method for efficiently performing super-resolution on a small satellite image based on key features and dynamic priori guidance according to claim 1, wherein in step S3, N is a positive integer, and the KDDB output feature images with a value of 6;N are processed by channel splicing, and sequentially pass through a1×1 convolution layer, a GELU activation function and a bottleneck convolution BSConv to obtain the fine feature image : , , , Wherein, the Representing the feature after activation by the GELU function, Representing deep features sufficiently fused by BSConv, N KDDB output characteristic diagrams respectively, For the channel-splicing operation, The function is activated for the purpose of GELU, In order for the intermediate feature to be activated, Representing the bottleneck convolution BSConv operation, C, H, W represents the number of channels, height, width of the feature map, respectively.
4. The method for high-efficiency super-resolution of small satellite images based on key features and dynamic prior guidance according to claim 1, wherein the partial channel cascade and distillation unit PCDM comprises at least two cascade partial channel distillation blocks PCDB, and a cascade strategy and a distillation strategy are adopted; The cascade strategy is formed by the PCDB cascades, and each PCDB adopts depth separable convolution DWConv with different parameter configurations to extract characteristics; And respectively distilling the output characteristics of each PCDB by the distillation strategy to obtain at least one of local surface details, multi-scale contour characteristics and global structure information, splicing the characteristics obtained by distillation with the input characteristics of the PCDM, and then fusing and outputting the characteristics through a convolution layer.
5. The method for high-efficiency super-resolution of small satellite images based on key features and dynamic prior guidance according to claim 1, wherein the step of performing the dynamic prior guidance channel attention unit DPCA comprises: s301, calculating the mean prior of each channel through global average pooling: , and calculating the contrast prior of each channel through standard deviation: , Wherein, the C-channel characteristics, Is that Is used for the average value of (a), A priori feature map is calculated for the current to-be-calculated; S302, the mean prior and the contrast prior are spliced into a combined tensor after being activated : , Convolution can be separated by kernel= (2, 1) depth Generating dynamic weights : , Two priori independent weighting fusion of each channel is obtained : , Wherein, the For the number of channels of the current input feature, The combined prior tensor weighted by the dynamic weight for the ith channel, The dynamic weight value of the ith channel; S303, optimizing the residual multi-layer perceptron RRMLP which is input with the fused prior and can be re-parameterized, and calculating through a Sigmoid activation function to obtain a channel attention map: , weighted fusion with the output features of the PCDM: , Wherein, the The representation RRMLP is optimized and, Representing the Sigmoid activation function, For the output of the PCDM, Is the output of DPCA.
6. The method for efficient super resolution of small satellite images based on key features and dynamic prior guidance according to claim 1, wherein the step of performing the efficient key feature aware self-attention unit KeySA comprises: s311, generating a query matrix Q, a key matrix K and a value matrix V through linear transformation and depth separable convolution on input features; s312, performing downsampling and channel number compression processing on K, V to obtain an optimized key feature matrix 、 ; S313, remolding Q and then combining Calculating a similarity matrix based on the similarity matrix and And calculating the attention output, and outputting after the optimization of the gating deep convolution feedforward network GDFN.
7. The efficient super-resolution method of a small satellite image based on key features and dynamic prior guidance according to claim 1, wherein the up-sampling process in step S4 is implemented by pixel rearrangement PixelShuffle, the preset multiple is 4 times, the convolution process is performed by using a 3×3 convolution layer, and the calculation formula of the super-resolution small satellite image is as follows: , Wherein, the Is a super-resolution small satellite image, For the pixel rearrangement operation, As a result of the fine feature map, Is a rough characteristic diagram.
8. The method for efficiently super-resolution of small satellite images based on key features and dynamic priori guidance according to claim 1 is characterized by further comprising the steps of adopting a two-stage training strategy, wherein a first stage takes a low-resolution small satellite image of a first size as an input, a first learning rate and a first iteration number are set, a second stage takes a low-resolution small satellite image of a second size as an input, a second learning rate and a second iteration number are set, and a loss function is minimized through an optimizer, so that final model weights are saved.
9. The small satellite image high-efficiency super-resolution system based on key features and dynamic priori guidance is characterized by comprising a shallow feature extraction module, a deep feature extraction module and a reconstruction module which are connected in sequence; The shallow feature extraction module is used for carrying out preliminary feature extraction on the input low-resolution small satellite image and outputting a coarse feature map; The deep feature extraction module comprises N serially connected key features and dynamic priori guided distillation modules KDDB, each KDDB integrated with a partial channel cascade and distillation unit PCDM, a dynamic priori guided channel attention unit DPCA and a high-efficiency key feature perception self-attention unit KeySA, and the deep feature extraction module is used for outputting a fine feature graph after multi-scale feature extraction and fusion of the coarse feature graph; the reconstruction module is used for carrying out residual connection on the fine feature image and the coarse feature image, and outputting a super-resolution small satellite image after convolution processing and up-sampling processing.
10. The small-sized satellite image high-efficiency super-resolution system based on key features and dynamic priori guidance according to claim 9, wherein the dynamic priori guidance channel attention unit DPCA comprises a priori extraction sub-module, a dynamic fusion sub-module and a channel attention allocation sub-module; The prior extraction submodule calculates the prior of the mean value of each channel through global average pooling: The dynamic fusion submodule is used for splicing the mean prior and the contrast prior into a combined tensor after activation treatment, generating dynamic weights through depth separable convolution, and independently weighting and fusing the two prior of each channel; And the channel attention allocation submodule optimizes the fused prior input re-parameterizable residual multi-layer perceptron RRMLP, calculates the channel attention map through an activation function, and performs weighted fusion with the output characteristics of the PCDM.

Description

Small satellite image high-efficiency super-resolution method and system based on key features and dynamic priori guidance Technical Field The invention relates to the fields of remote sensing image processing, small satellite application and deep learning intersection, in particular to a small satellite image high-efficiency super-resolution method and system based on key features and dynamic priori guidance. Background In recent years, the development of high-capacity microelectronics technology has driven the widespread use of satellites. The small satellite has the advantages of low manufacturing cost, high deployment speed, strong emission adaptability, excellent anti-striking capability, wide coverage range after constellation deployment and the like, plays an important role in the fields of environmental monitoring, resource investigation, urban planning, weather prediction and national defense, and becomes a key data source for earth optical imaging. However, the small satellite is limited by physical size, and has inherent technical bottlenecks that firstly, effective load and power are limited, the aperture of a telescope is smaller, so that imaging resolution is low, secondly, the thermal stability of an imaging system and a satellite platform is difficult to maintain, image definition is further reduced, subsequent decision accuracy is influenced, thirdly, the calculation force and storage resources of the small satellite are constrained, and complex algorithms are difficult to deploy. The cost of traditional resolution improvement modes such as aperture increase is increased in a square level, and high-resolution image data transmission is time-consuming, so that the practical application requirements are difficult to meet. Image Super Resolution (SR) technology provides a way to solve the above problems, and the core is to restore a Low Resolution (LR) image to a High Resolution (HR) image. The early SR method relies on interpolation or fixed priori to be disjointed with an actual scene, a Convolutional Neural Network (CNN) becomes a mainstream after deep learning is raised, such as SRCNN, RCAN and the like, although accuracy is improved, the inherent structure of the CNN is difficult to effectively model global context information, while the architecture of a Transformer enhances the global modeling capability, the self-attention computation complexity of a window is increased in square level along with the size of the window, and the existing methods such as HAT, DAT and the like have redundant computation or rely on external priori to cause high memory overhead, so that the method is difficult to adapt to the resource constraint environment of a small satellite. Although the existing light super-resolution method reduces the calculated amount, the contradiction between light weight and high precision is difficult to balance at the cost of losing reconstruction details. Therefore, there is a need for an image super-resolution technique that adapts to the constraints of the satellites resources, while allowing for efficient reasoning and high quality reconstruction. Disclosure of Invention The invention aims to provide a small satellite image high-efficiency super-resolution method and system based on key features and dynamic priori guidance, which can realize the coordination of model weight reduction, reasoning high-efficiency and reconstruction high precision and meet the calculation power and storage requirements of a small satellite platform. The invention aims to achieve the above purpose by the following technical scheme. A small satellite image high-efficiency super-resolution method based on key features and dynamic priori guidance comprises the following steps: s1, acquiring a low-resolution small satellite image to be processed, preprocessing the low-resolution small satellite image, and constructing an LR-HR image pair; s2, shallow feature extraction is carried out on the preprocessed low-resolution small satellite image, and a coarse feature map is obtained; S3, deep feature extraction is carried out on the crude feature map through a distillation module KDDB containing N key features connected in series and dynamically guided in prior, each KDDB is integrated with a partial channel cascade and distillation unit PCDM, a dynamically guided channel attention unit DPCA and a high-efficiency key feature perception self-attention unit KeySA, and a fine feature map is output after multi-scale feature extraction and fusion; s4, carrying out residual connection on the fine feature image and the coarse feature image, and outputting a super-resolution small satellite image with preset multiple after convolution processing and up-sampling processing. Further, the preprocessing in the step S1 comprises the steps of generating a low-resolution small satellite image from an original high-resolution small satellite image through downsampling, and constructing an LR-HR image pair; in step S2, sh