A lightweight remote sensing image learning-based compression method and system for on-board applications

Abstract

The application provides a lightweight remote sensing image learning compression method and system for on-board applications, and belongs to the technical field of remote sensing image processing and data compression. The problem of lightweight learning of the existing learning compression method with high compression efficiency and low model complexity is solved. The method comprises the following steps: converting a remote sensing image into a latent representation through a lightweight encoder; the lightweight encoder is designed based on a size kernel convolution structure; processing the latent representation by using a hyper-encoder to obtain a hyper-latent representation; quantizing the latent representation and the hyper-latent representation; encoding the quantized representation into a bit stream through arithmetic coding based on a spatial channel context parameter; at the decoding end, processing the bit stream by using a decoder to reconstruct the image. The method provided by the application can be used for efficient encoding of remote sensing images in resource-limited environments such as satellite platforms, space loads and unmanned aerial vehicles.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image processing and data compression technology, and in particular relates to a lightweight remote sensing image learning compression method for satellite applications. Background Technology

[0002] With the rapid development of high-resolution remote sensing satellite technology, the amount of remote sensing image data acquired on-board has exploded, posing severe challenges to satellite downlink bandwidth, on-board storage space, and processor computing power. Due to limited satellite platform resources, massive amounts of remote sensing images are difficult to transmit back to ground processing centers in a timely and complete manner. Therefore, it is urgent to perform efficient image compression on-board to reduce data size and improve downlink efficiency. Currently, the mainstream image compression methods on-board mainly rely on traditional transform coding algorithms, such as JPEG2000. These methods are based on wavelet transform and entropy coding, and can maintain acceptable image quality at a certain compression ratio. However, under high compression ratios, image fidelity decreases significantly, manifesting as obvious blur artifacts and texture loss, making it difficult to meet the stringent requirements of high-resolution remote sensing images for detailed reconstruction of ground features.

[0003] In recent years, learning-based image compression methods have demonstrated the potential to outperform traditional algorithms in reconstruction quality by utilizing neural networks for end-to-end optimization. These methods automatically learn image feature distributions through deep network models, achieving higher peak signal-to-noise ratios (PSNR) and multi-scale structural similarity (MS-SSIM) at the same bit rate, as exemplified by advanced methods such as the ELIC model based on residual networks. However, learning-based compression methods generally suffer from complex network structures, large parameter counts, and high computational and storage resource requirements. For example, ELIC's encoder has up to 4.78M parameters and requires 7.78G FLOPs of computation, making it difficult to deploy directly on spacecraft platforms with limited computing power and strict power consumption constraints, such as small satellites or UAV payloads. This high complexity leads to deployment feasibility challenges for learning-based methods in space environments, making it difficult to balance compression performance with resource constraints.

[0004] Specifically, the shortcomings of existing learning-based compression methods are mainly reflected in the following aspects: First, their neural network models typically contain a large number of convolutional layers and residual connections, requiring high memory bandwidth and powerful computing capabilities, which on-board processors often cannot meet. Second, these models consume a lot of energy during training and inference, conflicting with the power consumption limitations of on-board platforms. Furthermore, while traditional methods such as JPEG2000 have low complexity, their reconstruction quality is insufficient at low bit rates, while learning-based methods, although offering superior quality, are highly complex, creating a dilemma. Therefore, current technology lacks a lightweight learning-based compression solution that can operate in resource-constrained on-board environments while simultaneously achieving high compression efficiency and low model complexity. This has become a key bottleneck restricting the real-time processing and transmission of remote sensing images. Summary of the Invention

[0005] In view of this, the present invention aims to propose a lightweight remote sensing image learning compression method and system for satellite applications, in order to solve the problem of lightweight learning in existing learning compression methods that combine high compression efficiency and low model complexity.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a lightweight remote sensing image learning-based compression method for space applications, the method comprising: Remotely sensed images are converted into latent representations using a lightweight encoder; the lightweight encoder is designed based on a big.me and small kernel convolutional structure. The latent representation is processed using a super encoder to obtain a hyperlatent representation; The latent representation and hyperlatent representation are quantized; The quantized representation is encoded into a bit stream using arithmetic coding based on spatial channel context parameters; At the decoding end, the decoder is used to process the bitstream and reconstruct the image.

[0007] Furthermore, a preferred embodiment is proposed, wherein the big and small kernel convolutional structure in the lightweight encoder includes a big kernel sensing part and a small kernel aggregation part; The large kernel perception part uses a large convolutional kernel to capture global contextual information of the input features and generates adaptive weights; The small kernel aggregation part uses the adaptive weights to guide the small convolutional kernels to extract local features; The large and small kernel convolutional structure is represented by the formula: Large-core sensing process:

[0008] in, For each point in the feature map, K L For the size of a large core, For Centered on and with a size of K L × K L The area The kernel size is K L × K L Depthwise convolution, PW is pointwise convolution; for The generated weights; Small nucleus polymerization process:

[0009] in, Let i be the weight of the i-th point in the g-th group. for The c-th channel, For Centered on and with a size of K S × K S The area K S This refers to the size of the small core.

[0010] Furthermore, a preferred approach is proposed, wherein the big kernel convolutional structure further includes a deep convolutional layer, an SE layer, and a feedforward network; the big kernel sensing part adopts a combination of deep convolution and pointwise convolution, and the small kernel aggregation part adopts a group convolution and weight sharing mechanism.

[0011] Furthermore, a preferred approach is proposed, wherein the method further includes a training step employing a knowledge distillation-based training strategy: Extracting multi-feature representations using a teacher model ; During student model training, minimize the student encoder output features. The loss function of the training strategy, based on the differences between teacher characteristics, is:

[0012] in, For rate-distortion optimization, For bitrate loss, To reconstruct the distortion loss, For knowledge distillation loss, This is the weighting coefficient for balancing distortion.

[0013] Furthermore, a preferred method is proposed, wherein the bit rate loss The calculation method is as follows:

[0014] in, for bitrate, for bitrate, for The probability distribution, for The probability distribution, This indicates the potential value of y after quantization. The value after z-quantization is represented by the hyperpotential representation.

[0015] Furthermore, a preferred method is proposed, wherein the quantization operation includes:

[0016] in, Indicates uniformly distributed noise. Indicates quantization operation. y For potential representation, z This represents a hyperpotential representation.

[0017] Furthermore, a preferred approach is proposed, in which the training strategy for knowledge distillation employs an ELIC model based on residual networks for the teacher model and a lightweight encoder for the student model, with alignment achieved at the feature level.

[0018] Based on the same inventive concept, this invention also proposes a lightweight remote sensing image learning compression system for space applications, the system comprising: A lightweight encoder module is used to convert remotely sensed images into latent representations using a lightweight encoder; the lightweight encoder is designed based on a big.me and small kernel convolutional structure. The super encoder module is used to process the latent representation using a super encoder to obtain a super latent representation; A quantization module is used to quantize the latent representation and the hyperlatent representation; The arithmetic coding module is used to encode the quantized representation into a bit stream using arithmetic coding based on the spatial channel context parameters; The decoding module is used at the decoding end to process the bitstream and reconstruct the image using the decoder.

[0019] Based on the same inventive concept, the present invention also proposes a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes a lightweight remote sensing image learning compression method for satellite applications according to any one of the above.

[0020] Based on the same inventive concept, the present invention also proposes a computer-readable storage medium storing a computer program that, when executed by a processor, performs the steps of a lightweight remote sensing image learning compression method for space applications as described in any of the above-mentioned embodiments.

[0021] Compared with the prior art, the beneficial effects of the present invention are: Unlike traditional methods that simply port terrestrial compression schemes, this invention pioneers a collaborative framework of lightweight on-board coding and high-performance ground-based decoding. This design fully leverages the asymmetry between satellite and ground computing capabilities, distributing computationally intensive tasks to the ground, while the onboard system only needs to run a lightweight encoder. This asymmetric architecture overcomes the limitations imposed by the onboard environment on algorithm complexity, achieving a fundamental shift from infeasibility to feasibility.

[0022] Unlike traditional convolutional neural networks that process data at a single scale, this invention creatively proposes a large-kernel-small-kernel collaborative working mechanism. The large-kernel convolution is responsible for capturing global contextual information, while the small-kernel convolution performs adaptive feature aggregation based on this. This design simulates the global-local collaborative perception mechanism of the human visual system, significantly improving the model's ability to express multi-scale features of remote sensing images without substantially increasing computational costs.

[0023] To address the inherent challenge of limited representational capabilities in lightweight models, this invention proposes a feature-level knowledge distillation strategy. Through multi-layered feature guidance from the teacher model, the student model learns richer feature representations, effectively compensating for the performance loss caused by model simplification. This training mechanism enables lightweight models to achieve compression capabilities approaching those of large models while maintaining low inference complexity, breaking through the performance bottleneck of lightweight models. Compared to traditional methods such as JPEG2000, this invention learns the essential feature distribution of images through neural networks, rather than relying on fixed transformation basis functions; compared to mainstream learning methods, this invention resolves the contradiction between complexity and performance through structural innovation and training optimization.

[0024] Compared to traditional learning-based compression methods, this invention reduces encoder computational complexity (FLOPs) to 2.07G and parameter count to 0.82M, achieving at least a 4.6-fold reduction in computational complexity compared to mainstream methods. This lightweight nature allows the model to be directly deployed on satellite platforms with limited computing resources, solving the long-standing deployment feasibility problem of learning-based methods. In DOTA dataset testing, this invention achieved a PSNR of 36.25dB at a bitrate of 0.6bpp, an improvement of approximately 7dB compared to JPEG2000; on the GF6 dataset, the PSNR improvement exceeded 5dB at the same bitrate. More importantly, through a knowledge distillation training strategy, the lightweight model maintains low complexity while achieving rate-distortion performance close to the teacher model ELIC, with a difference of less than 1dB compared to the teacher model at a bitrate of 0.8bpp.

[0025] The method proposed in this invention exhibits excellent performance in preserving structural texture and fine-grained edge details when reconstructing images. Compared to the obvious blur artifacts produced by JPEG2000, this invention can clearly preserve the outlines and texture features of ground objects, achieving a PSNR value of 32.06dB even at a low bit rate of 0.5bpp, with visual effects comparable to highly complex learning-based methods.

[0026] The method proposed in this invention can be used for efficient encoding of remote sensing images in resource-constrained environments such as satellite platforms, space payloads, and UAVs. Attached Figure Description

[0027] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a schematic diagram of the overall structure of the lightweight remote sensing image learning compression method for space applications described in this invention. Figure 2 This is a schematic diagram of the big and small kernel convolution module described in this invention; Figure 3 This is a schematic diagram of the feature knowledge distillation process described in this invention; Figure 4 This is a schematic diagram of the rate-distortion performance curves of the three models described in this invention; Figure 5 This is a comparison chart of the rate-distortion performance of different methods described in this invention on the DOTA dataset; Figure 6The image is a visual comparison of the different methods described in this invention for reconstructing images on the DOTA dataset. From left to right, the images represent the input image, the image reconstructed by the traditional JPEG2000 method, the image reconstructed by the learning compression method HL-RSCompNet, the image reconstructed by the learning compression method ELIC, and the image reconstructed by this invention RS-LLIC. Figure 7 This is a comparison chart of the rate-distortion performance of different methods described in this invention on the GF6 dataset; Figure 8 This is a visual comparison of the compression and reconstruction of different methods described in this invention on the GF6 dataset. From left to right, the images represent the input image, the image reconstructed using the traditional JPEG2000 method, the image reconstructed using the HL-RSCompNet learning compression method, the image reconstructed using the ELIC learning compression method, and the image reconstructed using the RS-LLIC method of this invention. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of the present invention can be combined with each other, and the described embodiments are only some embodiments of the present invention, not all embodiments.

[0029] Implementation Method 1: This implementation method addresses the challenge of lightweight learning-based compression methods that combine high compression efficiency and low model complexity in existing learning-based compression methods. It proposes a lightweight remote sensing image learning-based compression method for satellite applications, comprising: Remotely sensed images are converted into latent representations using a lightweight encoder; the lightweight encoder is designed based on a big.me and small kernel convolutional structure. The latent representation is processed using a super encoder to obtain a hyperlatent representation; The latent representation and hyperlatent representation are quantized; The quantized representation is encoded into a bit stream using arithmetic coding based on spatial channel context parameters; At the decoding end, the decoder is used to process the bitstream and reconstruct the image.

[0030] The core technical principle of this implementation lies in the targeted design and innovative neural network structure for the special on-board environment, which is fundamentally different from existing technologies. While existing learning-based compression methods such as ELIC and TCM offer excellent compression performance, their network structures are generally complex. Directly porting them to the on-board environment would lead to excessive computational load and power consumption. Traditional on-board compression methods such as JPEG2000, although low in complexity, have limited compression efficiency and cannot meet the quality requirements of high-resolution remote sensing images.

[0031] This implementation proposes a lightweight encoder structure based on big and small kernel convolutions, a design that overcomes the limitations of traditional convolutional neural networks. The big kernel convolution is responsible for capturing the global receptive field, while the small kernel convolution adaptively guides the feature aggregation process using global information. This collaborative mechanism enables efficient modeling of complex textures and multi-scale features in remote sensing images while maintaining low computational complexity. Unlike traditional networks that simply stack convolutional layers, the big and small kernel convolution modules in this implementation achieve an optimal balance between feature representation capability and computational efficiency through fine-grained dimensionality control and group weight sharing mechanisms.

[0032] Another core element lies in the satellite-ground collaborative training strategy based on knowledge distillation. This method cleverly utilizes the asymmetry in computing power between satellite and ground, guiding the lightweight student model on satellite at the feature level through a high-performance teacher model on the ground, enabling the lightweight model to learn feature representation capabilities approaching those of complex models. This training strategy effectively addresses the inherent problem of insufficient representational capabilities in lightweight models, significantly improving compression performance without increasing inference complexity.

[0033] Implementation Method 2, see below Figures 1 to 8 This embodiment describes a complete implementation of the lightweight remote sensing image learning compression method for space applications described in Embodiment 1, including: The method proposed in this embodiment is specifically designed for on-board applications and optimizes the remote sensing image compression model, particularly the encoder, which is the most computationally intensive component in the learning-based compression framework. This lightweight design achieves competitive compression performance while significantly reducing model complexity and computational cost. The overall structure of the method proposed in this embodiment is as follows: Figure 1 As shown, it includes: Remotely sensed images are converted into latent representations using a lightweight encoder; the lightweight encoder is designed based on a big.me and small kernel convolutional structure. The latent representation is processed using a super encoder to obtain a hyperlatent representation; The latent representation and hyperlatent representation are quantized; The quantized representation is encoded into a bit stream using arithmetic coding based on spatial channel context parameters; At the decoding end, the decoder processes the bitstream to reconstruct the image; The method is optimized using a knowledge distillation-based training strategy, which guides the student encoder’s feature learning through a teacher model to improve compression ratio distortion performance with low computational complexity.

[0034] Specifically, in the overall structure of the lightweight remote sensing image learning-based compression method for space applications, the main structure corresponds to the student network in the diagram, while the teacher model at the top assists in the training process of the main structure. The teacher model adopts an ELIC model based on residual networks. In the diagram, Conv k m s n This indicates that it has a kernel size of m Step size is n Convolutional layers, TConvk m s n Indicates the kernel size as m Step size is n The transposed convolution is used. The lightweight encoder in this embodiment is designed based on a hybrid big.LITTLE architecture, while the decoder, superencoder, and superdecoder employ structures based on residual networks. During training, a knowledge distillation strategy is used to guide the learning of the student model.

[0035] Given an input image x Lightweight encoder g a Transform it into a latent representation y The super encoder further processes the data to obtain the hyperlatent representation. z : (1) Then, use a quantizer Q right y and z Quantification: (2) It is worth noting that, to achieve end-to-end training, uniform noise is used for approximate quantization during training, while standard quantization is used during inference. The formula is as follows: (3) in, Indicates uniformly distributed noise. This indicates a quantization operation. Then, parameters obtained from the spatial channel context are used. Arithmetic coding is performed to encode the quantized latent representation into a bitstream. During decoding, parameters are used. Decode the bitstream to recover and through the decoder g s Obtain the reconstructed image: (4) The lightweight encoder proposed in this embodiment consists of standard convolutional layers and big.l. and small kernel convolutional blocks. Figure 1As shown on the right, it consists of deep convolutional layers, SE layers, a feedforward network, and big.LITTLE convolutional layers. The big.LITTLE convolutional layers are the core components of the big.LITTLE convolutional block. By employing a large convolutional kernel, it captures contextual information from the image, which helps to better model global feature statistics during compression. Furthermore, the big.LITTLE convolution generates small kernel weights based on the large kernel's content-awareness, enhancing context aggregation and relational modeling. The detailed structure of the big.LITTLE convolution is shown below. Figure 2 As shown.

[0036] The big-kernel and small-kernel convolutional layer consists of two parts: a large-kernel sensing layer and a small-kernel aggregation layer. Let the input feature map be... First, 1×1 pointwise convolutions are used to map the features to a lower dimension C / 2 to minimize computation. Then, for each point in the feature map... Use a depthwise convolution with a size of KL×KL to capture around In this implementation, KL=7 is used to expand the spatial context information of the large receptive field. Large kernel depthwise convolution can effectively expand the receptive field, enhancing context awareness at minimal cost. Then, pointwise convolution is used to model channel relationships, thereby generating adaptive context weights. Used for the small-kernel aggregation step. The entire large-kernel sensing process can be represented as: (5) in, This is for The generated weights.

[0037] For the small kernel aggregation process, for the input feature map as First, they are divided into G groups, each containing C / G channels, and channels within the same group share aggregation weights, thereby reducing memory and computing resource consumption. For each For the corresponding weights Dimensional reshaping yields Where KS×KS is the size of the small core. In this embodiment, KS=3. Then, using to gather around it A highly relevant context. Specifically, The c-th channel is defined as It belongs to the g-th group, and then, through and Obtain aggregated features In this way, fine-grained features can be well represented. The process of small-kernel aggregation can be represented as: (6) To improve the feature representation capability of the lightweight encoder without increasing model complexity, this implementation proposes a training strategy based on knowledge distillation. For example... Figure 3 As shown, a complete teacher model is first trained, and then the encoder part of the teacher model is extracted separately as a knowledge guidance module. Subsequently, during the training of the student model, the teacher encoder outputs multi-layer feature maps during the forward propagation phase to guide the feature learning of the student encoder.

[0038] To reduce model complexity while maintaining compression performance, this invention employs a loss function design that combines rate-distortion optimization and knowledge distillation guidance during model training. The overall training objective is defined as follows: (7) in, For rate-distortion optimization, For bitrate loss, To reconstruct the distortion loss, For knowledge distillation loss, This represents the weighting coefficient for balance distortion. Wherein, The calculation method is as follows: (8) Then the mean square error (MSE) is used: (9) To maintain the encoder's representational power under lightweight conditions, this implementation introduces a distillation mechanism between the teacher network and the student encoder. By allowing the student network to mimic the feature distribution in the latent space of the teacher network, the lightweight model achieves encoding representation capabilities close to those of a high-performance model on the satellite platform. The final total loss is: (10) This design simultaneously optimizes rate-distortion performance and feature representation capabilities, enhancing the model's applicability to achieve efficient computation and high-fidelity compression in on-board computing environments.

[0039] To verify the effectiveness of the proposed method, multiple datasets were used, including the DOTA dataset and the GF6 dataset. The DOTA dataset contains 2806 images, ranging in size from 800×800 to 4000×4000, and includes various typical target categories and scenes. The training set was used to train the model, and 100 images from the test set were used for testing. The GF6 dataset is private data sourced from the China High Resolution Earth Observation System GF-6 satellite, with an image size of 1024×1024. It includes 1616 training images and 405 test images.

[0040] The experimental platform used was an Intel i9-10850k CPU, 64GB of RAM, and an NVIDIA RTX A6000 GPU. The training phase consisted of 600 epochs, with learning rate decay implemented at epochs 400 and 550. The initial learning rate was set to 10. -4 The weight decay is set to 10. -4 The model parameters are optimized using the Adam optimizer. The batch size is set to 8. In formula (7), The bit rate is controlled by five different sets of values: 0.01, 0.02, 0.05, 0.001, and 0.005.

[0041] Image quality is evaluated using Peak Signal-to-Noise Ratio (PSNR) and Multiscale Structural Similarity Index (MS-SSIM), with bit rate measured in bits per pixel (bpp). PSNR is calculated as follows: (11) in, and For reference images and reconstructed images, This represents the total number of pixels. For pixel dynamic range. MS-SSIM is calculated as follows: (12) in. and These are the reference image and the reconstructed image, respectively. M For multi-scale quantities, and The first m Image mean at each scale , The first m Image standard deviation at each scale For the first m Covariance at each scale and To maintain stability and prevent the denominator from being zero, and For the first m The weighting coefficients for each scale. To clearly compare the differences in MS-SSIM values, they are converted to dB values. The calculation method is as follows: (13) Therefore, bpp-PSNR and bpp-MS-SSIM curves were used to evaluate the rate-distortion performance of different models.

[0042] First, to demonstrate the effectiveness of this invention, we compared the complexity of the compression portion of our proposed method with other excellent learning-based compression methods, including the encoder portion and the complexity of the entire compression process. The metrics used included computational FLOPs and parameter counts (Params). The comparison methods included TCM, ELIC, Checkboard, HL-RSCompNet, and Joint Autoregressive. The results are shown in Table 1.

[0043] Table 1: Comparison of encoder and compression process complexity between the method of this invention and several other methods

[0044] Compared to other methods, the encoder FLOPs of the method of this invention are reduced by at least 4.6 times, while the overall compression process is reduced by at least 2.3 times. In addition to FLOPs, the encoder of this invention also contains fewer parameters, which helps to reduce memory requirements.

[0045] In addition, to verify the effectiveness of the knowledge distillation-based training strategy proposed in this invention, ablation experiments were conducted. Specifically, the teacher model ELIC was first trained on the DOTA and GF6 datasets. Then, the teacher model was used to guide the training of the student model. The model trained using knowledge distillation is denoted as RS-LLIC, while the model without knowledge distillation is denoted as RS-LLIC w / o KD. The rate-distortion curves of the three models on the DOTA and GF6 datasets are shown below. Figure 4 As shown.

[0046] As shown in the figures, the rate-distortion performance of the model trained using knowledge distillation is significantly better than that of the model without knowledge distillation. On the DOTA dataset, at a bit rate of 0.8 bpp, the PSNR is improved by approximately 1 dB, and the rate-distortion performance is very close to that of the teacher model ELIC. On the GF6 dataset, at a bit rate of 0.7 bpp, the PSNR is improved by approximately 2 dB. Furthermore, in terms of MS-SSIM, the RS-LLIC model consistently outperforms RS-LLIC without knowledge distillation on both datasets. The rate-distortion performance of RS-LLIC is almost comparable to that of the teacher model ELIC, indicating that the proposed knowledge distillation-based training strategy effectively improves the rate-distortion performance of the lightweight RS-LLIC model.

[0047] Furthermore, the method of this invention was compared with other excellent methods on the DOTA and GF6 datasets. The comparison methods included the traditional JPEG2000 method and learned compression methods such as HL-RSCompNet, JointAutoregressive, ELIC, and Checkboard. The rate-distortion performance comparison of different methods on the DOTA dataset is as follows: Figure 5 As shown.

[0048] like Figure 5 As shown, the RS-LLIC method proposed in this invention significantly outperforms the commonly used on-board compression method JPEG2000. At a bit rate of 0.6 bpp, RS-LLIC achieves a PSNR of 36.25 dB, MS-SSIM 18.17 dB, while JPEG2000 only reaches 29.26 dB and 10.83 dB, respectively. PSNR is improved by approximately 7 dB, and MS-SSIM by 7.34 dB. Furthermore, compared to learned compression methods such as HL-RSCompNet and Checkboard, RS-LLIC exhibits significantly superior rate-distortion performance. Although its performance is slightly lower than ELIC, its computational complexity is greatly reduced. In addition, this embodiment provides a visual comparison of reconstructed images generated by different compression methods on the DOTA dataset, such as... Figure 6 As shown.

[0049] For the first image, RS-LLIC achieved a PSNR of 30.11 dB at 0.523 bpp, outperforming JPEG2000 (24.36 dB at 0.598 bpp) and HL-RSCompNet (28.88 dB at 0.623 bpp) while maintaining a lower bitrate. Compared to ELIC, RS-LLIC achieved a highly competitive PSNR with a 9% reduction in bitrate. A similar trend was observed in the second image, where RS-LLIC reached a PSNR of 32.06 dB at 0.499 bpp, again outperforming JPEG2000 (24.05 dB at 0.594 bpp) and HL-RSCompNet (30.96 dB at 0.586 bpp), approaching the performance of ELIC (32.46 dB at 0.523 bpp). Visually, all three learning methods significantly outperformed JPEG2000. JPEG2000 exhibits noticeable blur artifacts, while HL-RSCompNet shows only slight texture loss. In contrast, the RS-LLIC of this invention effectively preserves structural texture and fine-grained edge details even at low bit rates.

[0050] Experimental results on the GF6 dataset further validate the effectiveness of the method described in this invention. Figure 7As shown, RS-LLIC consistently outperforms most comparable methods in rate-distortion performance. Compared to JPEG2000, RS-LLIC achieves significantly higher reconstruction quality at the same bit rate, demonstrating its ability to maintain structural fidelity while maintaining high compression efficiency. Compared to other learned compression methods such as HL-RSCompNet and Checkboard, RS-LLIC also exhibits a clear performance advantage at different bit rates. Although its rate-distortion curve differs slightly from ELIC, RS-LLIC has a much lower coding complexity, making it more suitable for resource-constrained satellite-based scenarios.

[0051] Furthermore, the reconstruction results on the GF6 dataset also demonstrate the advantages of RS-LLIC, such as... Figure 8 As shown, even at high compression ratios, the RS-LLIC of this invention can preserve texture and edge structure with high fidelity, producing clearer hierarchical details and overall visual quality. In contrast, JPEG2000 exhibits more noticeable blurred textures. While ELIC achieves slightly higher reconstruction quality, it requires a higher bit rate. Overall, RS-LLIC achieves a better balance between subjective image quality, structural fidelity, and applicability to on-board computing environments.

[0052] Implementation Method 3: A lightweight remote sensing image learning compression system for space applications, as described in this implementation method, includes: A lightweight encoder module is used to convert remotely sensed images into latent representations using a lightweight encoder; the lightweight encoder is designed based on a big.me and small kernel convolutional structure. The super encoder module is used to process the latent representation using a super encoder to obtain a super latent representation; The quantization module is used to quantize the latent representation and the hyperlatent representation; The arithmetic coding module is used to encode the quantized representation into a bit stream using arithmetic coding based on the spatial channel context parameters; The decoding module is used at the decoding end to process the bitstream and reconstruct the image using the decoder.

[0053] Implementation Method 4: A computer device according to this implementation method includes a memory and a processor. The memory stores a computer program. When the processor runs the computer program stored in the memory, the processor executes a lightweight remote sensing image learning compression method for satellite applications according to any one of Implementation Methods 1 to 2.

[0054] Implementation Method 5: A computer-readable storage medium according to this implementation method stores a computer program, which, when executed by a processor, performs the steps of a lightweight remote sensing image learning compression method for satellite applications as described in any one of Implementation Methods 1 to 2.

[0055] Those skilled in the art will understand that embodiments of this disclosure can be provided as methods, systems, or computer program products. Therefore, this disclosure can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this disclosure can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0056] This disclosure is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0057] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0058] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this disclosure and not to limit its protection scope. Although this disclosure has been described in detail with reference to the above embodiments, those skilled in the art should understand that after reading this disclosure, they can still make various changes, modifications or equivalent substitutions to the specific implementation of the invention, but these changes, modifications or equivalent substitutions are all within the protection scope of the published pending claims.

Claims

1. A lightweight remote sensing image learning-based compression method for satellite applications, characterized in that, The method includes: Remotely sensed images are converted into latent representations using a lightweight encoder; the lightweight encoder is designed based on a big.me and small kernel convolutional structure. The latent representation is processed using a super encoder to obtain a hyperlatent representation; The latent representation and hyperlatent representation are quantized; The quantized representation is encoded into a bit stream using arithmetic coding based on spatial channel context parameters; At the decoding end, the decoder processes the bitstream to reconstruct the image; The lightweight encoder's big-small kernel convolutional structure includes a big-kernel sensing part and a small-kernel aggregation part; The large kernel perception part uses a large convolutional kernel to capture global contextual information of the input features and generates adaptive weights; The small kernel aggregation part uses the adaptive weights to guide the small convolutional kernels to extract local features; The large and small kernel convolutional structure is represented by the formula: Large-core sensing process: in, For each point in the feature map, K L For the size of a large core, For Centered on and with a size of K L × K L The area The kernel size is K L × K L Depthwise convolution, PW is pointwise convolution; This is for The generated weights; For the small kernel aggregation process, the input feature map is first divided into G groups, each group contains multiple channels, and the channels in the same group share the aggregation weights. Small nucleus polymerization process: in, Let i be the weight of the i-th point in the g-th group. for The c-th channel, For Centered on and with a size of K S × K S The area K S For the size of the small core; This is an aggregation feature; For each For the corresponding weights Dimensional reshaping yields ; The c-th channel is defined as It belongs to the g-th group; The big-kernel and small-kernel convolutional structure also includes a deep convolutional layer, an SE layer, and a feedforward network; the big-kernel sensing part adopts a combination of deep convolution and pointwise convolution, and the small-kernel aggregation part adopts a group convolution and weight sharing mechanism. The method also includes a training step, employing a knowledge distillation-based training strategy: Extracting multi-feature representations using a teacher model ; During student model training, minimize the student encoder output features. The loss function of the training strategy, based on the differences between teacher characteristics, is: in, For rate-distortion optimization, For bitrate loss, To reconstruct the distortion loss, For knowledge distillation loss, This is the weighting coefficient for balancing distortion.

2. The lightweight remote sensing image learning-based compression method for space applications according to claim 1, characterized in that, The bit rate loss The calculation method is as follows: in, for bitrate, for bitrate, for The probability distribution, for The probability distribution, This indicates the potential value of y after quantization. The value after z-quantization is represented by the hyperpotential representation.

3. The lightweight remote sensing image learning-based compression method for space applications according to claim 1, characterized in that, The quantization operation includes: in, Indicates uniformly distributed noise. Indicates quantization operation. y For potential representation, z This represents a hyperpotential representation.

4. The lightweight remote sensing image learning-based compression method for space applications according to claim 1, characterized in that, In the training strategy for knowledge distillation, the teacher model adopts the ELIC model based on residual networks, and the student model adopts a lightweight encoder, which is aligned at the feature level.

5. A lightweight remote sensing image learning compression system for space applications, characterized in that, The system is implemented based on the method of claim 1, and the system includes: A lightweight encoder module is used to convert remotely sensed images into latent representations using a lightweight encoder; the lightweight encoder is designed based on a big.me and small kernel convolutional structure. The super encoder module is used to process the latent representation using a super encoder to obtain a super latent representation; The quantization module is used to quantize the latent representation and the hyperlatent representation; The arithmetic coding module is used to encode the quantized representation into a bit stream using arithmetic coding based on the spatial channel context parameters; The decoding module is used at the decoding end to process the bitstream and reconstruct the image using the decoder.

6. A computer device, characterized in that: It includes a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes a lightweight remote sensing image learning compression method for satellite applications according to any one of claims 1-4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of a lightweight remote sensing image learning compression method for space applications as described in any one of claims 1-4.

Images