A channel adaptive coding and decoding method, system, terminal device and medium for unmanned aerial vehicle image transmission

Abstract

The application discloses a channel adaptive coding and decoding method and system for unmanned aerial vehicle image transmission, a terminal device and a medium, and relates to the technical field of communication and remote sensing image processing. The method comprises the following steps: acquiring an air-ground image collected by an unmanned aerial vehicle; performing lightweight feature extraction on the air-ground image to obtain spatial-spectral compact features; performing channel enhancement based on a real-time channel signal-to-noise ratio to obtain channel enhanced features; performing transmission rate adaptive adjustment based on a real-time channel bandwidth ratio to obtain transmission features; transmitting the transmission features after power normalization through an air-ground wireless channel; and performing corresponding inverse recovery and reconstruction processing on the damaged features to obtain a reconstructed air-ground image. The application improves the robustness and bandwidth adaptation capability of image transmission under a complex channel, and realizes high-fidelity and high-efficiency unmanned aerial vehicle air-ground image transmission.

Description

Technical Field

[0001] This invention relates to the field of communication and remote sensing image processing technology, and in particular to a channel adaptive encoding and decoding method, system, terminal equipment and medium for UAV image transmission. Background Technology

[0002] With the widespread application of drones, the demand for channel adaptability, bandwidth utilization, and transmission robustness in air-to-ground image transmission is becoming increasingly urgent, but existing technologies have significant shortcomings. Specifically, traditional two-stage compression and channel coding schemes mainly use JPEG combined with LDPC and Turbo coding, which are mature in engineering but lack end-to-end global optimization. They suffer from severe loss of image details under low signal-to-noise ratio and are difficult to flexibly adapt to dynamic bandwidth. Deep learning end-to-end methods based on DeepJSCC, VAE, etc., although achieving integrated compression and transmission, suffer from insufficient channel adaptability, limited bandwidth adaptation capability, and homogenized feature channel processing, resulting in large performance fluctuations in complex scenarios. Long dependency modeling schemes based on Transformer or state-space models have excellent global modeling capabilities, but their high computational complexity makes them difficult to deploy on drone edge devices, and they lack explicit signal-to-noise ratio and bandwidth condition modeling, resulting in insufficient robustness.

[0003] Existing technologies cannot simultaneously meet the requirements of channel adaptability, dynamic bandwidth adaptation, lightweight deployment, and high-fidelity transmission. Therefore, there is an urgent need for a lightweight, robust, and bandwidth-adaptive channel-adaptive end-to-end encoding and decoding method for UAV image transmission to fill the gaps in existing technologies. Summary of the Invention

[0004] The technical problem this invention aims to solve is that, in the field of image joint source and channel coding and decoding technology for UAV air-to-ground links, existing technologies cannot simultaneously adapt to time-varying channel signal-to-noise ratio and dynamic bandwidth. These technologies suffer from drawbacks such as a sharp drop in image reconstruction quality under low signal-to-noise ratio environments, low transmission efficiency in bandwidth-constrained scenarios, high model complexity making deployment on UAV edge devices difficult, and insufficient robustness in transmission under complex channels. Therefore, an effective solution is urgently needed to address these technical problems.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: In a first aspect, the present invention provides a channel adaptive encoding and decoding method for image transmission in unmanned aerial vehicles (UAVs), applied to image transmission in UAV air-to-ground links, the method comprising: Acquire aerial images of the ground captured by the drone; Lightweight feature extraction processing is performed on the open ground image to extract local detail features of the open ground image. Based on the local detail features, long dependency features are extracted to obtain the spatial-spectral compact features of the open ground image. Based on the real-time channel signal-to-noise ratio of the UAV air-to-ground link, channel enhancement is performed on the spatial-spectral compact features, and different channel features are adaptively enhanced or suppressed to obtain channel-enhanced features; Based on the real-time channel bandwidth ratio of the UAV air-to-ground link, the transmission rate of the channel enhancement features is adaptively adjusted, and the features are dynamically filtered and compressed to obtain the transmission features. The transmission characteristics are power normalized and transmitted via an air-to-ground wireless channel; The damaged features transmitted through the air-to-ground wireless channel are received, and the damaged features are sequentially subjected to inverse recovery and reconstruction processing corresponding to transmission rate adaptation, channel enhancement, and lightweight feature extraction to obtain the reconstructed air-to-ground image.

[0006] In one implementation, the lightweight feature extraction process on the open ground image to extract local detail features of the open ground image, and the extraction of long-dependency features based on the local detail features to obtain the spatial-spectral compact features of the open ground image, includes: The empty ground image is subjected to convolution and activation processing to obtain the embedding features of the empty ground image; The embedded features are sequentially input into the first residual block and the second residual block to obtain the local detail features of the open ground image. The first residual block and the second residual block have the same structure, and the first residual block and the second residual block include convolution, activation and skip connection operations. A horizontal forward token sequence is generated from the local detail features through the first token generation block, and the horizontal forward token sequence is modeled with long dependency features through the first long dependency feature module to obtain the first long dependency feature. The first long-dependency feature is processed by a self-attention layer to perform feature interaction processing on global context and long-range spatial dependency, resulting in self-attention processed features. A horizontal backward token sequence is generated from the self-attention processing features through the second token generation block, and the horizontal backward token sequence is modeled with long dependency features through the second long dependency feature module to obtain the second long dependency feature. The second long-dependency feature is used as the spatial-spectral compact feature of the empty ground image.

[0007] In one implementation, both the first long dependency feature and the second long dependency feature are Mamba features, and the first long dependency feature module and the second long dependency feature module are structurally identical Mobile Mamba modules. The process of modeling the long dependency features of the token sequence using the Mobile Mamba module includes: The horizontal forward token sequence or the horizontal backward token sequence is normalized by a normalization layer to obtain normalized features. The normalized features are convolved by a one-dimensional convolutional layer to obtain the convolution result; Intermediate features are obtained by activating the convolution results using a non-linear activation function. The intermediate features are linearly transformed using a linear layer to obtain the linear transformation result; The extraction results are obtained by extracting long-dependency features from the linear transformation results using a selective state-space model. The extraction results are normalized by a normalization layer to obtain the first or second longest dependency feature.

[0008] In one implementation, the real-time channel signal-to-noise ratio based on the UAV air-to-ground link is used to enhance the spatial-spectral compact features, and to adaptively enhance or suppress different channel features to obtain channel-enhanced features, including: The real-time channel signal-to-noise ratio based on the UAV air-to-ground link is used to generate a channel weight vector through a signal-to-noise ratio adjustment block; The spatial-spectral compact features are subjected to three involution operations in sequence, and after each involution operation, the output features of the involution operation are multiplied by the channel weight vector broadcast channel by channel to obtain the involution channel features; The involute channel features are sequentially subjected to linear transformation and softmax operation to generate global channel weight parameters. The global channel weight parameters are then multiplied by the involute channel features channel by channel broadcast to obtain the channel enhancement features. The involution operations are all the same channel-dimensional adaptive feature aggregation operations, and the signal-to-noise ratio adjustment block includes multi-level fully connected layers and non-linear activation functions to map the real-time signal-to-noise ratio into a weight vector that matches the number of spatial-spectral compact feature channels.

[0009] In one implementation, the transmission rate of the channel enhancement features is adaptively adjusted based on the real-time channel bandwidth ratio of the UAV air-to-ground link, and the features are dynamically filtered and compressed to obtain transmission features, including: The real-time channel bandwidth ratio based on the UAV air-to-ground link generates a rate vector through a rate adjustment block. The channel enhancement features are subjected to three convolution operations in sequence, and after each convolution operation, the output feature of the convolution operation is multiplied by the rate vector broadcast channel by channel to obtain the convolution weighted intermediate features; Perform a convolution operation on the convolution-weighted intermediate features to obtain rate-adjusted intermediate features; The rate adjustment intermediate feature is used to generate a mask vector through a rate mask block. The mask vector is then multiplied by the channel enhancement feature broadcast channel by channel to obtain the transmission feature. The rate adjustment block includes multi-level fully connected layers and nonlinear activation functions, which are used to map the real-time channel bandwidth ratio into a rate vector that matches the number of channels of the channel enhancement feature. The rate mask block generates a mask vector through average pooling, linear transformation and nonlinear activation, and the operation mode of each convolution operation is the same.

[0010] In one implementation, the step of power normalizing the transmission characteristics and transmitting them via an air-to-ground wireless channel includes: Obtain the global element expectation of the second-order Euclidean norm of the preset transmit power and the transmission characteristics; The power normalization coefficient is calculated based on the preset transmission power and the expected global element power. The power normalization coefficient is multiplied element-wise by the transmission characteristic to obtain the normalized transmission characteristic; The normalized transmission characteristics are transmitted via the UAV air-to-ground link.

[0011] In one implementation, the end-to-end training process of the method includes: Construct a channel-adaptive coding / decoding network that matches the method described above; The training images are acquired and input into the encoding model of the encoding and decoding network. After lightweight feature extraction, channel enhancement, and adaptive transmission rate adjustment, the training transmission features are obtained. The training transmission features are subjected to power normalization to obtain training normalized transmission features; By simulating a real UAV air-to-ground link channel environment, additive white Gaussian noise interference and Rayleigh fading interference are introduced into the training normalized transmission characteristics to obtain training impaired characteristics. The training-damaged features are input into the decoding model of the encoder-decoder network, and the training-reconstructed image is obtained after inverse recovery and reconstruction processing. Based on the training image and the training reconstructed image, calculate the reconstruction loss in the pixel domain; Based on the training normalized transmission features and the training impaired features, the perceptual loss of the feature domain is calculated. Based on the reconstruction loss and perception loss, a composite loss function is constructed, and the parameters of the encoding and decoding network are iteratively optimized with the goal of minimizing the composite loss function, to obtain the trained channel adaptive encoding and decoding network.

[0012] Secondly, embodiments of the present invention also provide a channel adaptive encoding and decoding system for UAV image transmission, the system comprising: The data acquisition module is used to acquire aerial and ground images collected by the drone; The lightweight feature extraction module is used to perform lightweight feature extraction processing on the ground image, extract local detail features of the ground image, extract long dependency features based on the local detail features, and obtain the spatial-spectral compact features of the ground image. The channel enhancement module is used to enhance the spatial-spectral compact features based on the real-time channel signal-to-noise ratio of the UAV air-to-ground link, and to adaptively enhance or suppress different channel features to obtain channel enhancement features; The transmission rate adaptive module is used to adaptively adjust the transmission rate of the channel enhancement features based on the real-time channel bandwidth ratio of the UAV air-to-ground link, and to dynamically filter and compress the features to obtain the transmission features. A transmission module is used to normalize the power of the transmission characteristics and transmit them through an air-to-ground wireless channel; The receiving and reconstruction module is used to receive the damaged features transmitted through the air-to-ground wireless channel, and sequentially perform reverse recovery and reconstruction processing on the damaged features corresponding to transmission rate adaptation, channel enhancement, and lightweight feature extraction to obtain the reconstructed air-to-ground image.

[0013] Thirdly, embodiments of the present invention also provide a terminal device, the terminal device including a memory, a processor, and a channel adaptive encoding and decoding program for UAV image transmission stored in the memory and executable on the processor, wherein when the processor executes the channel adaptive encoding and decoding program for UAV image transmission, it implements the steps of the channel adaptive encoding and decoding method for UAV image transmission described in any of the above schemes.

[0014] Fourthly, embodiments of the present invention also provide a computer-readable storage medium storing a channel adaptive encoding / decoding program for UAV image transmission. When the channel adaptive encoding / decoding program for UAV image transmission is executed by a processor, it implements the steps of the channel adaptive encoding / decoding method for UAV image transmission described in any of the above schemes.

[0015] Beneficial Effects: This invention discloses a channel adaptive encoding / decoding method, system, terminal device, and medium for UAV image transmission, relating to the fields of communication and remote sensing image processing technology. The method is applied to image transmission via an UAV air-to-ground link. First, it acquires air-to-ground images collected by the UAV. Lightweight feature extraction processing is then performed on the air-to-ground images to extract local detail features. Based on these local detail features, long-dependency features are extracted to obtain the spatial-spectral compact features of the air-to-ground images. Subsequently, based on the real-time channel signal-to-noise ratio of the UAV air-to-ground link, channel enhancement is performed on the spatial-spectral compact features. Different channel features are adaptively enhanced or suppressed to obtain channel-enhanced features. Then, based on the real-time channel bandwidth ratio of the UAV air-to-ground link, the transmission rate of the channel-enhanced features is adaptively adjusted, and features are dynamically filtered and compressed to obtain transmission features. Finally, the transmission features are power normalized and transmitted via an air-to-ground wireless channel. Damaged features transmitted via the air-to-ground wireless channel are received, and the damaged features are sequentially subjected to inverse recovery and reconstruction processing corresponding to the transmission rate adaptation, channel enhancement, and lightweight feature extraction to obtain the reconstructed air-to-ground image. This invention adapts to the deployment constraints of UAV edge devices through a lightweight feature extraction architecture. Based on real-time channel signal-to-noise ratio and bandwidth ratio, it achieves dual adaptive adjustment of feature channels and transmission rate, significantly improving transmission robustness and bandwidth utilization in complex channel environments. Furthermore, through end-to-end joint optimization, it balances transmission efficiency and image reconstruction accuracy, and can still achieve high-fidelity and high-reliability transmission of air-to-ground images in low signal-to-noise ratio and bandwidth-limited scenarios. Attached Figure Description

[0016] Figure 1 A flowchart illustrating a specific implementation of the channel adaptive encoding / decoding method for UAV image transmission provided in this invention.

[0017] Figure 2 This is a schematic diagram of a channel-adaptive end-to-end image transmission network for a channel-adaptive encoding / decoding method for UAV image transmission provided in an embodiment of the present invention.

[0018] Figure 3 This is a schematic diagram of a lightweight feature extraction operation for a channel adaptive encoding and decoding method for UAV image transmission provided in an embodiment of the present invention.

[0019] Figure 4 This diagram illustrates the horizontal forward and horizontal backward token generation of the channel adaptive encoding / decoding method for UAV image transmission provided in this embodiment of the invention.

[0020] Figure 5 This is a schematic diagram of channel enhancement operation for a channel adaptive encoding and decoding method for UAV image transmission provided in an embodiment of the present invention.

[0021] Figure 6 This is a schematic diagram illustrating the transmission rate adaptive operation of the channel adaptive encoding and decoding method for UAV image transmission provided in an embodiment of the present invention.

[0022] Figure 7 This is a typical scenario transmission recovery diagram of the channel adaptive encoding and decoding method for UAV image transmission provided in the embodiments of the present invention.

[0023] Figure 8 Peak signal-to-noise ratio (SNR) performance curves of the channel adaptive encoding / decoding method for UAV image transmission provided in this embodiment of the invention under different SNR conditions.

[0024] Figure 9 The peak signal-to-noise ratio performance curves of the channel adaptive encoding and decoding method for UAV image transmission provided in this embodiment of the invention are shown.

[0025] Figure 10 This is a schematic diagram of a channel adaptive encoding and decoding device for UAV image transmission provided in an embodiment of the present invention.

[0026] Figure 11 This is a block diagram illustrating the internal structure of the terminal device provided in an embodiment of the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and effects of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0028] The flowchart shown in the attached diagram is for illustrative purposes only and does not necessarily include all content, operations, or steps, nor does it require execution in the described order. For example, some operations or steps can be broken down, combined, or partially merged, so the actual execution order may change depending on the actual situation.

[0029] It should be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0030] It should be understood that, in order to clearly describe the technical solutions of the embodiments of the present invention, the terms "first" and "second" are used in the embodiments of the present invention to distinguish identical or similar items with essentially the same function and effect. For example, "first control information" and "second control information" are only used to distinguish different control information and do not limit their order.

[0031] Those skilled in the art will understand that the words "first" and "second" do not limit the quantity or the order of execution, and that the words "first" and "second" do not necessarily imply that they are different.

[0032] It should also be understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0033] With the rapid development of UAV technology, UAVs have been widely used in many fields such as remote sensing and mapping, environmental monitoring, emergency rescue, urban inspection, and border patrol. Real-time transmission of air-to-ground images via air-to-ground links is a core component for UAVs to perform various tasks. UAV air-to-ground wireless links have inherent characteristics such as strong channel time-varying nature, large dynamic fluctuation range of signal-to-noise ratio, and limited and unstable bandwidth resources. At the same time, the computing power, storage resources, and power consumption of UAV onboard devices are subject to strict constraints. Therefore, extremely high requirements are placed on the lightweight nature of image transmission schemes, channel adaptability, anti-interference robustness, bandwidth adaptation flexibility, and image reconstruction fidelity.

[0034] Currently, technologies related to UAV air-to-ground image transmission can be mainly divided into three categories: traditional two-stage image transmission schemes, end-to-end joint source-channel coding schemes based on deep learning, and end-to-end communication schemes based on long dependency modeling structures. All of these schemes have technical shortcomings that make them difficult to adapt to the core requirements of UAV air-to-ground transmission scenarios, as detailed below: The first type is the traditional two-stage transmission scheme of source compression and channel coding. This type of scheme uses standardized image compression algorithms such as JPEG to achieve source compression, and combines classic channel coding methods such as LDPC and Turbo to achieve anti-interference transmission. After long-term engineering iteration, it has a high degree of technical maturity and is currently the most widely used scheme in engineering. However, this type of scheme has the following inherent limitations: First, it lacks end-to-end global optimization. The image compression and channel coding stages are independent and designed separately, making it impossible to achieve a globally optimal trade-off between overall image transmission quality, transmission delay, and bandwidth utilization. It is difficult to adapt to the dynamic channel characteristics of UAV air-to-ground links. Second, it has weak anti-interference capability in low signal-to-noise ratio environments. Under complex channel conditions with low signal-to-noise ratio, the features after JPEG compression and channel noise will have a serious cumulative effect, resulting in a significant loss of key details such as texture and edges in the reconstructed image, and a sharp decline in image reconstruction quality. Third, it has poor bandwidth adaptability. The compression ratio and coding parameters are mostly fixed settings and cannot be flexibly adjusted according to the real-time bandwidth changes of the air-to-ground link. In bandwidth-limited scenarios, information redundancy or transmission interruption is likely to occur, resulting in poor universality.

[0035] The second category is end-to-end joint source-channel coding schemes based on deep learning, with DeepJSCC, Variational Autoencoder (VAE), and JSCCformer as typical examples. With the development of deep learning and computer vision technologies, this type of end-to-end scheme has gradually become a research hotspot. Through deep learning structures such as convolutional neural networks and attention mechanisms, it directly maps the input image into a transmittable, compact feature representation, achieving integrated modeling of image compression and robust transmission, breaking through the theoretical upper limit of traditional two-stage architectures. However, this type of solution still has significant shortcomings in UAV air-to-ground transmission scenarios: First, it lacks channel adaptability. Existing solutions mostly train models under fixed signal-to-noise ratio (SNR) conditions, resulting in significant fluctuations in transmission performance under different SNR environments. Especially in low SNR scenarios, the image reconstruction quality drops drastically, making it difficult to adapt to the real-time dynamic changes in SNR of UAV air-to-ground links. Second, it has limited bandwidth adaptability. Most solutions only optimize for fixed compression rates and transmission rates, lacking an effective mechanism for dynamically adjusting the amount of feature transmission based on channel bandwidth. This makes it impossible to achieve a dynamic balance between transmission efficiency and reconstruction quality in bandwidth-constrained and bandwidth-rich scenarios, resulting in a severe lack of flexibility in actual UAV operation scenarios. Third, it suffers from uneven feature utilization efficiency. Existing solutions use a homogenized transmission processing method for different channel features after encoding, failing to distinguish feature channels with different contributions to image reconstruction. This easily leads to insufficient anti-interference capability for key feature transmission and redundant features occupying valuable bandwidth resources, making it difficult to achieve optimal transmission results under limited bandwidth.

[0036] The third category comprises end-to-end communication solutions based on the Transformer architecture or improved state-space models. In recent years, to enhance the ability to capture long-range spatial dependencies in images, this type of solution has begun to be applied in the field of image transmission. It leverages self-attention mechanisms and selective state-space modeling to capture global dependencies, thereby strengthening the modeling effect on global image features and long-range spatial dependencies, and improving the efficiency of image feature representation to some extent. However, this type of solution still has insurmountable shortcomings in UAV air-to-ground transmission scenarios: on the one hand, the computational complexity and the number of model parameters are enormous. The self-attention mechanism of the Transformer and the inference computation of complex state-space models have extremely high overhead, while the computing resources and power consumption of UAV edge devices are strictly constrained, making it difficult to achieve real-time deployment and online inference on the UAV's onboard end; on the other hand, channel adaptability still has significant shortcomings. Existing solutions still lack explicit modeling and adaptive adjustment mechanisms for real-time channel signal-to-noise ratio and bandwidth conditions, and cannot dynamically adjust the coding strategy according to the real-time channel state of the UAV air-to-ground link. The transmission robustness and adaptability in complex time-varying channel environments still cannot meet the core requirements of actual UAV operations.

[0037] In summary, existing UAV air-to-ground image transmission technologies either suffer from poor image reconstruction quality under complex channel conditions with low signal-to-noise ratio and limited bandwidth, or are unable to adaptively adjust to time-varying channel conditions, or have excessively complex models that are difficult to deploy on UAV edge devices. They cannot simultaneously meet the multiple requirements of lightweight deployment, adaptive channel adjustment, dynamic bandwidth adaptation, and high-fidelity image transmission, resulting in core technical problems such as poor transmission robustness, low bandwidth utilization, and insufficient scene adaptability in UAV air-to-ground image transmission tasks. Therefore, there is an urgent need to develop a novel channel-adaptive end-to-end image transmission scheme to simultaneously address the aforementioned shortcomings of existing technologies and achieve efficient, robust, and high-fidelity transmission of UAV air-to-ground images in complex channel environments.

[0038] To address the aforementioned technical problems, this embodiment proposes a channel-adaptive encoding / decoding method for UAV image transmission. For example... Figure 2 As shown, the channel-adaptive end-to-end image transmission network used in this embodiment is divided into three core parts: the encoding end, the channel transmission link, and the decoding end. The encoding end is deployed on the UAV's airborne end, and the decoding end is deployed on the ground receiving end. At the UAV end, the acquired air-to-ground images first enter the lightweight feature extraction module to extract local details and global long-dependency features, generating spatial-spectral compact features with high expressive efficiency. Subsequently, the spatial-spectral compact features enter the channel enhancement module and the transmission rate adaptation module in sequence. The channel enhancement module performs dynamic weighted enhancement and suppression of feature channels according to the real-time channel signal-to-noise ratio of the UAV air-to-ground link, and the transmission rate adaptation module performs dynamic filtering and compression of features according to the real-time channel bandwidth ratio, finally generating transmission features adapted to the current channel state. After power normalization processing, the transmission features are transmitted through the air-to-ground wireless channel to complete the air-to-ground link transmission. At the ground receiving end, the decoding end receives the damaged features after transmission through the channel and, following the reverse processing order from the encoding end, sequentially completes the reverse recovery and reconstruction operations corresponding to transmission rate adaptation, channel enhancement, and lightweight feature extraction. Finally, it outputs the reconstructed air-to-ground image, realizing end-to-end adaptive transmission of UAV air-to-ground images in complex channel environments.

[0039] Specifically, this embodiment provides a channel adaptive encoding and decoding method for UAV image transmission, such as... Figure 1 As shown, it includes the following steps: Step S100: Obtain aerial images of the ground collected by the drone.

[0040] In this embodiment, the aerial image is an optical image of the ground scene acquired by the UAV through an onboard remote sensing imaging device, using a vertical overhead view or a small-angle tilt shooting method. Specifically, it can be a visible light remote sensing image containing typical ground features such as mountains, bridges, water bodies, buildings, industrial areas, and urban roads, or it can be a multispectral or hyperspectral remote sensing image. The spatial resolution and spectral dimensions of the image can be adapted and adjusted according to different UAV missions. The onboard imaging device for acquiring aerial and ground images can be a CCD area array camera, a complementary metal-oxide-semiconductor camera, or a multispectral imaging payload integrated into the UAV flight control system. The imaging device and the UAV's onboard processing unit complete real-time transmission of image data through a high-speed data interface, providing the raw image data source for subsequent encoding and decoding processing. In this embodiment, the aerial and ground images can uniformly adopt a three-channel RGB format, and the image size is adapted to the input requirements of the subsequent lightweight feature extraction module, ensuring the continuity and stability of subsequent feature extraction and encoding processing.

[0041] Step S200: Perform lightweight feature extraction processing on the open ground image to extract local detail features of the open ground image, and extract long dependency features based on the local detail features to obtain the spatial-spectral compact features of the open ground image.

[0042] In this embodiment, lightweight feature extraction processing is an image feature extraction method adapted to the limited computing, storage, and power constraints of UAV-borne edge devices. Its core is to extract and compress high-dimensional image features using low-complexity network structures and operators while ensuring the expressive power of image features. Specifically, it can be a feature extraction process based on lightweight convolutional neural networks and lightweight state-space models, which differs from traditional high-complexity Transformer and deep convolutional network feature extraction methods, enabling real-time inference computation on the UAV's onboard device. Local detail features are feature vectors that characterize the texture, edges, contours, color gradients, and other local spatial information of ground objects in aerial images. They are fundamental features for achieving high-fidelity image reconstruction. Specifically, they can be multi-scale local spatial features obtained through convolution operations and residual structure extraction, accurately capturing the fine structural information of different ground objects in aerial images. Long-dependency features are feature vectors that characterize global spatial associations and long-distance pixel dependencies in air-to-ground images. They enable the modeling of global contextual information in air-to-ground images. Specifically, they can be global features extracted through state-space models and self-attention mechanisms. This compensates for the difficulty of capturing global dependencies in local convolution operations, improving the overall expressive efficiency of features. Spatial-spectral compact features are high-compact feature representations that integrate local image detail information and global long-dependency information while achieving dimensionality compression. They significantly reduce the dimensionality and data volume of features while ensuring the integrity of all information required for image reconstruction, adapting to the transmission bandwidth constraints of air-to-ground wireless channels. In this embodiment, lightweight feature extraction processing is implemented through a two-level structure of local feature extraction and global long-dependency modeling. While ensuring lightweight feature extraction, it also takes into account the complete capture of local details and global information. It can generate highly expressive and compact features with limited computing resources, providing a foundation for subsequent channel adaptive coding and efficient transmission.

[0043] In one implementation, the lightweight feature extraction process on the open ground image to extract local detail features of the open ground image, and the extraction of long-dependency features based on the local detail features to obtain the spatial-spectral compact features of the open ground image, specifically includes the following steps: Step S210: Perform convolution and activation processing on the empty ground image to obtain the embedding features of the empty ground image; Step S220: Input the embedded features into the first residual block and the second residual block in sequence to obtain the local detail features of the empty ground image. The first residual block and the second residual block have the same structure. The first residual block and the second residual block include convolution, activation and skip connection operations. Step S230: Generate a horizontal forward token sequence from the local detail features using the first token generation block, and perform long dependency feature modeling on the horizontal forward token sequence using the first long dependency feature module to obtain the first long dependency feature; Step S240: Perform global context and long-range spatial dependency feature interaction processing on the first long-dependency feature through a self-attention layer to obtain self-attention processed features; Step S250: Generate a horizontal backward token sequence from the self-attention processing features through the second token generation block, and perform long dependency feature modeling on the horizontal backward token sequence through the second long dependency feature module to obtain the second long dependency feature; Step S260: Use the second long-dependency feature as the spatial-spectral compact feature of the empty ground image.

[0044] In this embodiment, as Figure 3 As shown, the lightweight feature extraction module consists of a convolutional layer, a ReLU nonlinear activation layer, two residual blocks with the same structure, two token generation blocks, two long dependency feature modules with the same structure, and a self-attention layer cascaded in sequence. It can complete the extraction of local details and global long dependency features with low computational complexity.

[0045] In the process of obtaining embedded features through convolution and activation processing of the empty ground image, an initial convolutional layer with a kernel size of 3×3 is used to complete the dimensionality transformation and shallow feature extraction of the input image. Then, a ReLU nonlinear activation layer is used to complete the nonlinear mapping of features, finally generating the embedded features. The mathematical expression of this process is as follows:

[0046] in, This represents the input empty ground image. , This represents the number of pixels in the vertical direction of the original open space image. This indicates the number of pixels in the horizontal direction of the original open ground image. This represents an initial convolutional layer with a kernel size of 3×3, where ReLU represents the rectified linear activation function. This represents the embedded features of the final output. A 3×3 convolutional kernel achieves an optimal balance between receptive field and computational cost, while the ReLU activation function effectively alleviates the gradient vanishing problem, making it suitable for the lightweight inference needs of drone edge devices.

[0047] The embedded feature input consists of two structurally identical residual blocks to extract local detail features. Each residual block comprises three 1×1 convolutional layers, three ReLU nonlinear activation layers, and skip connection structures. The specific three convolutional and activation processes of the residual block are as follows:

[0048]

[0049]

[0050] in , , These represent three convolutional layers with a kernel size of 1×1. , , These represent the output features after three convolutional activations.

[0051] Furthermore, the fusion of embedded features and residual convolutional features is achieved through skip connection operations. The output of the first residual block is specifically as follows:

[0052] in, This represents the output characteristics of the first residual block. The output of the second residual block can be obtained by operating exactly the same way as the first residual block. This feature represents the local detail features of the empty ground image. The two-level residual block cascade method can achieve complete capture of multi-scale local detail information of the empty ground image, while the skip connection structure can effectively solve the gradient degradation problem in the training process of deep networks and completely preserve the original detail information in the input features.

[0053] like Figure 4 As shown, the token generation blocks are derived from local detail features. This method generates token sequences in both the horizontal forward and horizontal backward directions. This generation method is adapted to the vertical overhead shooting characteristics of UAV air-to-ground images. Regardless of the scanning direction, the spatial feature distribution of the image exhibits high similarity. Complete modeling of global spatial information can be achieved using only the two horizontal token sequences, eliminating the need for a vertical token sequence and significantly reducing computational complexity. The horizontal forward token sequence is denoted as... The horizontal backward token sequence is denoted as Two sequences are input into two long dependency feature modules with the same structure to complete the modeling of long-distance spatial dependency in the corresponding direction, and generate the first long dependency feature and the second long dependency feature respectively.

[0054] Between two long-dependent feature modules, a self-attention layer is set up to handle feature interaction between global context and long-range spatial dependencies. The processing of the self-attention layer is as follows:

[0055] in , , These represent the query matrix, key matrix, and value matrix, respectively, obtained from the first longest dependency feature through a linear transformation. Represents the normalized exponential function, with superscript indicating the function. This represents the matrix transpose operation. The self-attention layer can establish spatial dependencies between any two pixels in the feature map, further enhancing the fusion effect of global contextual information and compensating for the information loss in unidirectional long dependency modeling.

[0056] The features processed by the self-attention layer are then processed by the second token generation block and the second long dependency feature module, ultimately outputting the second long dependency feature. This feature integrates local image detail information and global long dependency information, possessing high compactness and high expressive efficiency, and is directly used as the spatial-spectral compact feature output of the empty ground image. This spatial-spectral compact feature will subsequently be processed by the channel enhancement module and the transmission rate adaptation module, ultimately generating transmission features adapted to channel transmission. The entire encoding process can be expressed as a whole by a formal formula, specifically:

[0057] in, For the final generated transmission characteristics, To adapt the adaptive coding model for UAV air-to-ground transmission, This formula fully characterizes the entire coding transformation logic from the original ground image to the transmitted features, encompassing all network parameters included in the coding model.

[0058] In alternative implementations, the basic network structure of the lightweight feature extraction module can also adopt other lightweight convolutional network structures or other lightweight Transformer variants to further adapt to UAV onboard equipment with different computing power levels; the generation direction of the token sequence can also be adjusted according to the change of the UAV shooting angle, by adding a vertical token sequence to adapt to the feature extraction requirements of aerial and ground images in oblique photography scenarios.

[0059] In one implementation, both the first long dependency feature and the second long dependency feature are Mamba features, and the first long dependency feature module and the second long dependency feature module are structurally identical Mobile Mamba modules. The process of modeling the long dependency features of the token sequence using the Mobile Mamba module specifically includes the following steps: Step S231: Normalize the horizontal forward token sequence or the horizontal backward token sequence through a normalization layer to obtain normalized features; Step S232: Perform convolution operation on the normalized features through a one-dimensional convolutional layer to obtain the convolution result; Step S233: Activate the convolution result using a non-linear activation function to obtain intermediate features; Step S234: Perform a linear transformation on the intermediate features through a linear layer to obtain the linear transformation result; Step S235: Extract long-dependency features from the linear transformation results using a selective state-space model to obtain the extraction results; Step S236: Normalize the extracted results through a normalization layer to obtain the first long dependency feature or the second long dependency feature.

[0060] In this embodiment, both the first long dependency feature and the second long dependency feature are Mamba features. The first long dependency feature module and the second long dependency feature module are Mobile Mamba modules with the same structure. The Mobile Mamba module is a lightweight selective state space modeling module. Compared with the traditional self-attention mechanism, it has linear temporal complexity, faster inference speed, and lower computational overhead, and can adapt to the deployment constraints of UAV airborne edge devices.

[0061] Horizontal forward token sequence Taking the processing of tokens as an example, the Mobile Mamba module first generates intermediate features when modeling the token sequence, specifically:

[0062] in, For the generated horizontal forward intermediate features, The SiLU activation function is used. For normalization layer, As a one-dimensional convolutional layer, this process completes the initial feature transformation of the token sequence through normalization, convolution, and nonlinear activation, preparing for subsequent long dependency modeling.

[0063] After completing the intermediate feature extraction, the token sequence is further modeled using the Mobile Mamba module to perform complete long-dependency feature modeling. The overall process consists of six steps: normalization, one-dimensional convolution, non-linear activation, linear transformation, selective state space modeling, and quadratic normalization. The specific processing flow is as follows: First, the input horizontal forward token sequence is processed through a normalization layer. Or horizontal backward token sequence Layer normalization is performed to obtain normalized features. Layer normalization can standardize and adjust the feature distribution of the token sequence, alleviate the internal covariate shift problem, and improve the stability and convergence speed of model training.

[0064] The normalized features are then convolved using a one-dimensional convolutional layer to obtain the convolution result. One-dimensional convolution operations can fuse local contextual information in the temporal dimension of the token sequence, further strengthening the local dependencies of the sequence and laying the foundation for subsequent global long-term dependency modeling.

[0065] The convolution result is then activated using the SiLU activation function to obtain intermediate features. The SiLU activation function, also known as the Swish activation function, is mathematically expressed as follows: It possesses the characteristics of being unbounded at the upper limit but bounded at the lower limit, smooth, and non-monotonic. In lightweight deep models, its fitting ability and generalization effect are superior to the traditional ReLU activation function, which can effectively improve the richness of feature representation.

[0066] Then, a linear transformation is performed on the intermediate features through a linear layer to adjust the feature dimensions, resulting in a linear transformation result that adapts to the input dimensions of the selective state-space model. This step is similar to the aforementioned horizontal forward intermediate features. The subsequent linear transformation logic is consistent.

[0067] Subsequently, long dependency features are extracted from the linear transformation result using a selective state-space model to obtain the extraction result. The selective state-space model is a state-space model with sequence selective modeling capabilities. It can adaptively adjust the state update and memory methods according to the content of the input sequence, possessing extremely strong modeling capabilities for long-distance dependencies in long sequences. Furthermore, its computational complexity increases only linearly with the sequence length, far lower than the quadratic complexity of self-attention mechanisms. In this embodiment, the selective state-space model is implemented using the S6 model, which is the core computational module of the Mamba architecture. The result after extracting long dependency features using this model needs to undergo normalization processing to finally generate horizontal forward Mamba features, specifically:

[0068] in, Forward-looking Mamba features, For linear layers, It is a selective state-space model. As a normalization layer, this formula fully represents the core logic of the Mobile Mamba module in generating Mamba features. This applies to the horizontal backward token sequence. Using the exact same processing flow, we can obtain the horizontal backward Mamba feature, which is the second longest dependency feature.

[0069] Finally, the extraction results are normalized a second time through a normalization layer to obtain the final first or second long dependency feature, thus completing the entire long dependency feature modeling process.

[0070] The output of the lightweight feature extraction module is the output of the second Mobile Mamba module, which is... .

[0071] The Mobile Mamba module used in this embodiment combines the long dependency capture capability of state space modeling with a lightweight network structure. While ensuring the global feature modeling effect, it significantly reduces the amount of computation and parameters, solving the problem of high complexity and difficulty in deploying traditional Transformer-type models on edge devices of drones.

[0072] Step S300: Based on the real-time channel signal-to-noise ratio of the UAV air-to-ground link, channel enhancement is performed on the spatial-spectral compact features, and different channel features are adaptively enhanced or suppressed to obtain channel-enhanced features.

[0073] In this embodiment, the channel signal-to-noise ratio (SNR) is the ratio of the average power of the transmitted signal to the average power of the channel noise in the UAV air-to-ground wireless link. It is a core parameter characterizing the current channel transmission quality, measured in decibels (dB). A higher SNR indicates better channel transmission quality, while a lower SNR indicates stronger channel interference and poorer transmission quality. The real-time SNR can be measured in real-time by the channel estimation algorithm of the UAV's onboard communication module. The measurement results can be synchronized in real-time to the channel enhancement module at the encoding end via a low-bandwidth feedback link. Channel enhancement is a feature processing method that dynamically weights and adjusts feature channels based on their different importance levels. Its core is to enhance feature channels with high contribution to image reconstruction and weak anti-interference ability, while suppressing feature channels with low contribution to image reconstruction and high redundancy, thereby improving the overall anti-interference ability and information utilization efficiency of the features. Adaptive enhancement or suppression is a processing logic that automatically adjusts the weighting coefficients of different feature channels based on the dynamic changes in the real-time channel signal-to-noise ratio (SNR). In low SNR environments, it can focus on enhancing the weights of core feature channels and suppressing the transmission of invalid information from redundant channels, ensuring reliable transmission of core information. In high SNR environments, it can balance the weights of each channel, preserving richer image detail information. Channel-enhanced features are feature representations that have undergone dynamic weighting optimization after channel enhancement processing. They possess stronger channel anti-interference capabilities and can ensure stable transmission of core feature information in complex, time-varying channel environments. In this embodiment, channel enhancement processing establishes a direct mapping relationship between channel state and feature channel weights. It can dynamically adjust the channel distribution of features according to real-time channel quality, effectively improving the adaptability and transmission robustness of the coding scheme under different SNR conditions, and solving the problem of significant degradation in image reconstruction quality in low SNR environments in existing technologies.

[0074] In one implementation, the real-time channel signal-to-noise ratio based on the UAV air-to-ground link is used to enhance the spatial-spectral compact features, and different channel features are adaptively enhanced or suppressed to obtain channel-enhanced features. Specifically, this includes the following steps: Step S310: The real-time channel signal-to-noise ratio based on the UAV air-to-ground link is used to generate a channel weight vector through a signal-to-noise ratio adjustment block; Step S320: Perform three involution operations on the spatial-spectral compact features in sequence, and after each involution operation, multiply the output feature of the involution operation with the channel weight vector broadcast channel by channel to obtain the involution channel features; Step S330: Perform linear transformation and softmax operation on the involute channel features in sequence to generate global channel weight parameters. Multiply the global channel weight parameters with the involute channel features channel by channel broadcast to obtain the channel enhancement features. The involution operations are all the same channel-dimensional adaptive feature aggregation operations, and the signal-to-noise ratio adjustment block includes multi-level fully connected layers and non-linear activation functions to map the real-time signal-to-noise ratio into a weight vector that matches the number of spatial-spectral compact feature channels.

[0075] In this embodiment, as Figure 5 As shown, the channel enhancement module consists of three signal-to-noise ratio adjustment blocks, three involution operations, a linear layer, and a softmax operation module. It can perform dynamic adaptive enhancement and suppression of the feature channel according to the real-time channel signal-to-noise ratio, thereby improving the transmission robustness of the feature under different channel conditions.

[0076] First, a channel weight vector is generated based on the real-time channel signal-to-noise ratio (SNR) of the UAV air-to-ground link using a signal-to-noise ratio (SNR) adjustment block. The SNR adjustment block consists of a cascaded four-level fully connected layer and a nonlinear activation function, which maps the single-dimensional real-time channel SNR value to a value matching the number of spatially-spectrally compact feature channels. Dimensional weight vector. The specific processing procedure for the signal-to-noise ratio adjustment block is as follows:

[0077]

[0078]

[0079]

[0080] in, This represents the real-time channel signal-to-noise ratio. and They represent the first Weights and biases of a fully connected layer , and These represent the nonlinear activation function and the linear rectified activation function, respectively. Indicates the final output A four-dimensional channel weight vector. A fully connected layer can establish a non-linear mapping relationship between signal-to-noise ratio and channel weights, enabling adaptive generation of channel weights under different signal-to-noise ratio conditions.

[0081] In this embodiment, the involution operation is an adaptive feature aggregation operation in the channel dimension. Unlike the traditional convolution operation, which shares weights in the spatial domain and is independent in the channel domain, the involution operation shares weights in the spatial domain and is independent in the channel domain. Its convolution kernel is generated by the input features themselves and can adaptively assign different weights to different spatial locations, summarizing contextual information in a larger spatial range while significantly reducing computational complexity.

[0082] like Figure 5As shown, the kernel of the involution operation From input features Through kernel generation function get:

[0083] in, This represents the spatial dimensions of the involution kernel. Representing coordinates Involution kernel of position Generated by kernel function In pixels Generated under certain conditions. Representing coordinates Location-based input features, Represents the kernel generating function. and This represents the weights of the two linear transformation layers. This represents batch normalization and the ReLU nonlinear activation function. Number of channels. Involution kernel. and Having the same spatial size, it is able to summarize contextual information over a larger spatial range, and since the convolution kernel is generated from the input features themselves, it can adaptively assign different weights to different spatial locations. Indicates in Multiply the broadcasts on each channel, Indicates in Aggregation and summation operations within a spatial region.

[0084] The spatial-spectral compact features are subjected to three involution operations in sequence. After each involution operation, the output feature of that involution operation is multiplied by the channel weight vector broadcast channel by channel. The process of the three operations is as follows:

[0085]

[0086]

[0087] in, Indicates in Element-by-element broadcast multiplication operation on each channel. , , This represents three involution operations with identical structure and operation methods. , , This represents the output feature after three rounds of involution weighting, and the final result is... This is the involution channel feature. Three cascaded involution and weighting operations can gradually enhance the core feature channel while suppressing invalid information from redundant channels.

[0088] Features of involution channels obtained from triple involution and weighted involution. The global channel weight parameters are generated by sequentially performing linear transformation and softmax operation, as follows:

[0089] in, This represents the generated global channel weight parameters. This represents the normalized exponential function.

[0090] Finally, the global channel weight parameters are combined with the involution channel features. The channel-by-channel broadcasts are multiplied to obtain the final channel enhancement features, specifically:

[0091] in, This represents the channel enhancement features of the final output.

[0092] In this embodiment, the channel enhancement module directly integrates channel state information into the feature encoding process, achieving adaptive feature enhancement for channel awareness. In low signal-to-noise ratio (SNR) environments, it prioritizes the protection of core feature transmission, while in high SNR environments, it retains richer detailed information, significantly improving the system's robustness in complex time-varying channels. In alternative implementations, the channel enhancement module can also introduce other channel attention mechanisms such as squeeze excitation modules and convolutional block attention modules, or introduce normalization methods such as instance normalization and layer normalization to further enhance adaptability to complex channels.

[0093] Step S400: Based on the real-time channel bandwidth ratio of the UAV air-to-ground link, the transmission rate of the channel enhancement features is adaptively adjusted, and the features are dynamically filtered and compressed to obtain the transmission features.

[0094] In this embodiment, the channel bandwidth ratio is the ratio of the currently available transmission bandwidth to the nominal maximum bandwidth for image transmission in the UAV air-to-ground wireless link. It is a core parameter characterizing the current channel transmission rate carrying capacity; a higher value indicates a higher transmission rate that the channel can carry, while a lower value indicates more limited channel bandwidth resources. The real-time channel bandwidth ratio can be obtained in real time through the bandwidth detection algorithm of the UAV air-to-ground communication link, and the detection result is synchronized to the transmission rate adaptive module at the encoding end in real time. Transmission rate adaptive adjustment is the processing method that automatically adjusts the amount of data and compression rate of the features to be transmitted based on the dynamic changes in the available channel bandwidth. Its core is to retain more feature information to improve image reconstruction quality when the channel bandwidth is abundant, and to compress and filter feature information to adapt to bandwidth constraints when the channel bandwidth is limited, avoiding transmission congestion and information loss. Dynamic filtering and compression are achieved by generating channel masks to dynamically select and filter feature channels of different importance, retaining core feature channels that contribute highly to image reconstruction and filtering out non-core feature channels with high redundancy, thereby realizing dynamic adjustment of feature data volume and completing lossless compression and efficient filtering of features. The transmission feature is the final representation of the feature to be transmitted after adaptive adjustment of the transmission rate to adapt to the current channel bandwidth capacity. Its data volume is perfectly matched with the current channel bandwidth, enabling the most efficient transmission under the current channel conditions. In this embodiment, the adaptive adjustment of the transmission rate establishes a dynamic matching relationship between the channel bandwidth and the feature transmission volume, which can flexibly adapt to the bandwidth fluctuations of the UAV air-to-ground link, effectively improve bandwidth utilization, and solve the problems of low transmission efficiency and poor adaptability of existing technologies in bandwidth-constrained scenarios.

[0095] In one implementation, the transmission rate of the channel enhancement features is adaptively adjusted based on the real-time channel bandwidth ratio of the UAV air-to-ground link, and the features are dynamically filtered and compressed to obtain transmission features. Specifically, this includes the following steps: Step S410: The real-time channel bandwidth ratio based on the UAV air-to-ground link is used to generate a rate vector through a rate adjustment block; Step S420: Perform three convolution operations on the channel enhancement features in sequence, and after each convolution operation, multiply the output feature of the convolution operation with the rate vector broadcast channel by channel to obtain the convolution weighted intermediate features; Step S430: Perform a convolution operation on the convolution-weighted intermediate features to obtain rate-adjusted intermediate features; Step S440: Generate a mask vector from the rate adjustment intermediate feature using a rate mask block, and multiply the mask vector with the channel enhancement feature by channel-by-channel broadcast to obtain the transmission feature; The rate adjustment block includes a multi-level fully connected layer and a nonlinear activation function, which is used to map the real-time channel bandwidth ratio into a rate vector that matches the number of channels of the channel enhancement feature. The rate mask block generates a mask vector through average pooling, linear transformation and nonlinear activation function, and the operation mode of each convolution operation is the same.

[0096] In this embodiment, as Figure 6 As shown, the transmission rate adaptive module consists of three rate adjustment blocks, four 1×1 convolutional layers, and a rate mask block. It can adaptively adjust the amount of data of the features to be transmitted according to the real-time channel bandwidth ratio, so as to achieve dynamic matching between the transmission rate and the channel bandwidth.

[0097] First, a rate vector is generated based on the real-time channel bandwidth ratio of the UAV air-to-ground link using a rate adjustment block. The rate adjustment block consists of a cascaded three-level fully connected layer and a nonlinear activation function, which maps the single-dimensional real-time channel bandwidth ratio to a value matching the number of channels in the channel enhancement feature. A 3D rate vector. The specific processing procedure for the rate adjustment block is as follows:

[0098]

[0099]

[0100] in, Indicates the real-time channel bandwidth ratio. and They represent the first Weights and biases of a fully connected layer , and These represent the linear rectification activation function and the nonlinear activation function, respectively. Indicates the final output A three-dimensional rate vector. A fully connected layer can establish a nonlinear mapping relationship between the channel bandwidth ratio and the rate vector, enabling adaptive generation of rate weights under different bandwidth conditions.

[0101] The channel enhancement features are sequentially subjected to three 1×1 convolution operations. After each convolution operation, the output feature of that operation is multiplied by the rate vector channel by channel broadcast to obtain the convolutionally weighted intermediate features. The 1×1 convolution operation can complete the fusion and transformation of feature channels without changing the feature space size. The channel-by-channel broadcast multiplication dynamically weights the features of different channels according to the rate vector, realizing the initial screening of feature channels with different levels of importance.

[0102] The weighted intermediate features are then subjected to another convolution operation with a kernel size of 1×1 to obtain rate-adjusted intermediate features, thus completing further feature fusion and dimensionality adaptation. In this embodiment, the kernel size and operation method are the same for all four convolution operations, ensuring the consistency and stability of feature processing.

[0103] The intermediate features of the rate adjustment are input into the rate mask block to generate a mask vector. The specific processing procedure for the rate mask block is as follows:

[0104]

[0105] in, This indicates intermediate characteristics of rate regulation. This represents the feature vector after global average pooling. Indicates the final output 3D mask vector, This indicates a global average pooling operation. Represents a linear transformation layer. This represents a non-linear activation function. Each element in the generated mask vector takes a value between 0 and 1, corresponding to the transmission weight of each feature channel.

[0106] Finally, the mask vector is multiplied by the channel enhancement feature broadcast channel by channel to obtain the final transmission feature, specifically:

[0107] in, This indicates the transmission characteristics of the final output. This indicates the channel enhancement features output by the channel enhancement module. This represents a channel-by-channel broadcast multiplication operation. Channels with values ​​close to 1 in the mask vector are core feature channels, which will be fully preserved and transmitted; channels with values ​​close to 0 are redundant feature channels, which will be suppressed and filtered out, thus achieving dynamic adjustment of the feature data volume and adaptive adjustment of the transmission rate. When the channel bandwidth ratio is high, more channels in the mask vector have values ​​close to 1, preserving more feature information to improve reconstruction quality; when the channel bandwidth ratio is low, the mask vector only retains the high weight of core channels, compressing the feature data volume to adapt to bandwidth constraints and avoiding the cliff effect of traditional methods under low bandwidth.

[0108] In this embodiment, the transmission rate adaptive module achieves dynamic matching between the amount of transmitted information and the channel bandwidth, significantly improving bandwidth utilization and scenario adaptability. In alternative implementations, the transmission rate adaptive module can also be combined with methods such as variable bit rate coding and arithmetic entropy coding to further improve bandwidth utilization efficiency; the mask generation method can also be replaced with dynamic gating mechanisms such as gated recurrent units and gated convolution to achieve more refined feature selection.

[0109] Step S500: Normalize the power of the transmission characteristics and transmit them through an air-to-ground wireless channel.

[0110] In this embodiment, power normalization is a process of adjusting the amplitude of the feature data to be transmitted, ensuring that the average transmit power of the feature data meets the transmit power constraints of the wireless communication system. Its core purpose is to guarantee that the average transmit power of the feature data conforms to the communication protocol during wireless channel transmission, while avoiding signal distortion and reception errors caused by power fluctuations. The air-to-ground wireless channel is the wireless communication transmission link between the UAV's onboard unit and the ground receiver. Specifically, it can be a narrowband communication link using orthogonal frequency division multiplexing modulation or a broadband ad hoc network communication link. Its channel characteristics conform to typical air-to-ground wireless channel propagation patterns, and it is subject to interference such as additive white Gaussian noise and multipath Rayleigh fading during transmission. It serves as the transmission carrier for image feature data in this embodiment. In this embodiment, power normalization ensures that the transmit power of the features to be transmitted conforms to system constraints, effectively reducing the probability of signal distortion during wireless transmission, improving the feature demodulation accuracy at the receiver, and providing stable feature input for subsequent image reconstruction.

[0111] In one implementation, the step of power normalizing the transmission characteristics and transmitting them via an air-to-ground wireless channel specifically includes the following steps: Step S510: Obtain the global element expectation of the preset transmit power and the second-order Euclidean norm of the transmission characteristics; Step S520: Calculate the power normalization coefficient based on the preset transmit power and the expected global element; Step S530: Multiply the power normalization coefficient element by element with the transmission characteristics to obtain the normalized transmission characteristics; Step S540: Transmit the normalized transmission features via the UAV air-to-ground link.

[0112] In this embodiment, the core purpose of power normalization of transmission features is to ensure that the average transmit power of the feature data to be transmitted strictly conforms to the preset transmit power constraints of the wireless communication system, thereby avoiding signal distortion, reception demodulation errors, and compliance issues caused by fluctuations in feature power, as well as exceeding the power limits of the communication protocol.

[0113] During the power normalization process, the expected value of the global elements of the preset transmit power and the second-order Euclidean norm of the transmission characteristics is first obtained. The preset transmit power is the maximum average transmit power preset in the UAV air-to-ground communication system, denoted as . Its value is set according to the communication protocol, the hardware parameters of the UAV communication module, and the air-to-ground communication distance. The second-order Euclidean norm, also known as the L2 norm, is the global element expectation of the second-order Euclidean norm of the transmission feature. It is the mathematical expectation of the elements of all dimensions of the transmission feature, representing the overall power level of the transmission feature. In this embodiment, it is denoted as... ,in This represents the expected value of all elements of the feature. Let denote the second-order Euclidean norm.

[0114] Then, the power normalization coefficient is calculated based on the preset transmit power and the expected global element power. The calculation logic of the power normalization coefficient is: the square root of the ratio of the preset transmit power to the expected global element power. This coefficient can linearly scale the overall power of the transmission characteristics to the constraint range of the preset transmit power.

[0115] Finally, the power normalization coefficient is multiplied element-wise by the transmission characteristics to obtain the normalized transmission characteristics, thus completing the entire power normalization process, as shown below:

[0116] in, Indicates the transmission characteristics of the input. This represents the normalized transmission characteristics of the final output.

[0117] After power normalization, the normalized transmission characteristics are... Transmitted via UAV air-to-ground link, the normalized feature data can be directly mapped to the baseband modulation symbols of wireless communication, and transmitted to the air-to-ground wireless channel through the radio frequency module, thus completing the feature transmission of the air-to-ground link.

[0118] Step S600: Receive the damaged features transmitted through the air-to-ground wireless channel, and sequentially perform reverse recovery and reconstruction processing on the damaged features corresponding to transmission rate adaptation, channel enhancement, and lightweight feature extraction to obtain the reconstructed air-to-ground image.

[0119] In this embodiment, the damaged feature refers to the receiver's characteristic data that, after transmission via an air-to-ground wireless channel, suffers data distortion and information loss due to interference such as channel noise and multipath fading. In this embodiment, this damaged feature is denoted as... This is the input data source for the reverse recovery and reconstruction processing at the decoding end. The overall processing at the decoding end can be expressed by the following formal formula:

[0120] in, This is the reconstructed image of the empty ground. For adaptive decoding model, These are all the parameters of the decoding model. The inverse recovery and reconstruction process is a completely symmetrical inverse operation to the encoding process. Following the reverse order of the encoding process, it sequentially performs the inverse operations corresponding to transmission rate adaptation, channel enhancement, and lightweight feature extraction, gradually mapping the damaged features back to the image pixel domain, ultimately generating the reconstructed ground image corresponding to the formula. The processing modules at the decoding end and their corresponding modules at the encoding end employ a symmetrical network structure and parameter configuration to ensure end-to-end matching between the encoding and decoding processes, enabling the decoding model to accurately reproduce the feature transformation logic of the encoding model. In this embodiment, the symmetrical reverse recovery and reconstruction processing, together with the adaptive encoding processing at the encoding end, forms an end-to-end global optimization, maximizing the restoration of image information in the damaged features and achieving high-fidelity reconstruction of air-to-ground images under complex channel interference conditions.

[0121] In one implementation, the end-to-end training process of the method specifically includes the following steps: Step S710: Construct a channel adaptive coding / decoding network that matches the method described above; Step S720: Obtain the training image and input it into the encoding model of the encoding / decoding network. After lightweight feature extraction, channel enhancement, and adaptive transmission rate adjustment, the training transmission features are obtained. Step S730: Perform power normalization processing on the training transmission features to obtain training normalized transmission features; Step S740: Simulate a real UAV air-to-ground link channel environment, introduce additive white Gaussian noise interference and Rayleigh fading interference into the training normalized transmission features to obtain training impaired features; Step S750: Input the training damaged features into the decoding model of the encoder-decoder network, and obtain the training reconstructed image after inverse recovery and reconstruction processing; Step S760: Calculate the reconstruction loss in the pixel domain based on the training image and the training reconstructed image; Step S770: Calculate the perceptual loss of the feature domain based on the training normalized transmission features and the training impaired features; Step S780: Construct a composite loss function based on the reconstruction loss and perception loss, and iteratively optimize the parameters of the encoding and decoding network with the goal of minimizing the composite loss function, to obtain the trained channel adaptive encoding and decoding network.

[0122] In this embodiment, the channel adaptive coding and decoding method completes network training in an end-to-end manner. During the training process, the coding model and the decoding model are jointly optimized, eliminating the need for separate design of source coding and channel coding, and achieving the optimal balance of global performance.

[0123] During end-to-end training, a channel adaptive encoding and decoding network matching the method is first constructed. The encoding and decoding network consists of three parts: an encoding model, a channel simulation layer, and a decoding model. The encoding and decoding models adopt a symmetrical network structure. The encoding model includes a lightweight feature extraction module, a channel enhancement module, and a transmission rate adaptive module. The decoding model includes a corresponding inverse recovery and reconstruction module. The channel simulation layer is used to simulate the real UAV air-to-ground link channel environment.

[0124] A training dataset containing UAV air-to-ground images is obtained. These training images are then input into the encoding model of the encoder-decoder network. After lightweight feature extraction, channel enhancement, and adaptive transmission rate adjustment, training transmission features are obtained. In this embodiment, the training dataset uses the typical remote sensing image dataset AID. This dataset contains air-to-ground images of various typical UAV operation scenarios, such as mountains, bridges, waterways, industrial areas, and playgrounds, covering samples with different resolutions, lighting conditions, and terrain types, ensuring that the trained model possesses excellent generalization ability. In alternative implementations, other remote sensing image datasets can be used as the training dataset, or a private dataset can be constructed using air-to-ground images collected during actual UAV operations to further improve the model's performance in specific operation scenarios.

[0125] The training transmission features are subjected to power normalization to obtain training normalized transmission features. The power normalization process is consistent with the actual transmission process, ensuring the matching between the training process and the actual inference process.

[0126] By simulating the real UAV air-to-ground link channel environment through a channel simulation layer, additive white Gaussian noise interference and Rayleigh fading interference are introduced into the training normalized transmission features to obtain training impaired features. The damaged feature is completely consistent with the damaged feature definition of the input at the decoding end.

[0127] The specific process of introducing additive white Gaussian noise interference is as follows:

[0128] in, This indicates that additive white Gaussian noise follows a pattern with a mean of 0 and a variance of . Gaussian distribution, Indicates symbolic power, A linear scale representing the signal-to-noise ratio. The signal-to-noise ratio of the simulated channel set for the channel simulation layer.

[0129] The specific process of introducing Rayleigh fading interference is as follows:

[0130] in, This represents the channel coefficient vector, composed of complex Gaussian random variables, with dimensions consistent with the trained normalized transmission features. The specific expression is , and Let be mutually independent random variables, and let them all follow a Gaussian distribution with a mean of zero and a variance of 1. The imaginary unit, This represents complex Gaussian noise that matches the dimension of the channel coefficient vector.

[0131] During training, channel conditions with different signal-to-noise ratios and bandwidth ratios are randomly generated to simulate the time-varying characteristics of UAV air-to-ground links. This allows the model to complete training in diverse channel environments, improving its channel adaptability and generalization. In alternative implementations, the channel simulation layer can also introduce channel models that more closely resemble real UAV flight scenarios, such as Ricean fading, multipath channels, and Doppler shift, further enhancing the model's robustness in real-world scenarios.

[0132] The damaged training features are input into the decoding model of the encoder-decoder network, and the training reconstructed image is obtained after inverse recovery and reconstruction processing. The reconstructed image is completely consistent with the definition in the formal formula at the decoding end.

[0133] Based on the training image and the training reconstructed image, the pixel-domain reconstruction loss is calculated. The reconstruction loss adopts the mean square error loss function, specifically:

[0134] in, Represents the original training images. Indicates the training and reconstruction of images. This represents the reconstruction loss, which penalizes pixel-level numerical differences to ensure that the decoded and reconstructed image is as spatially consistent as possible with the original image.

[0135] Based on the normalized transmission features and the impaired transmission features, the perceptual loss of the feature domain is calculated. This perceptual loss is used to constrain the consistency of information in the feature domain before and after transmission, ensuring that the two images remain consistent not only in pixel values ​​but also at the perceptual level. The formula is:

[0136] in, This represents the normalized features after noisy transmission, i.e., the feature domain data corresponding to the training impaired features. Indicates the original normalized transmission characteristics. , , These represent the number of channels, height, and width of the feature, respectively. This indicates perceived loss.

[0137] A composite loss function is constructed based on the reconstruction loss and the perception loss. The composite loss function is a weighted combination of the reconstruction loss and the perception loss, specifically as follows:

[0138] in, This is a hyperparameter greater than 0, used to control the weight of perceptual loss in the overall loss function, and can be adjusted according to the training effect.

[0139] With the goal of minimizing the composite loss function, the parameters of the encoding / decoding network are iteratively optimized using the backpropagation algorithm until the composite loss function converges, resulting in a trained channel adaptive encoding / decoding network. The training process employs an adaptive moment estimation algorithm as the optimizer, setting corresponding hyperparameters such as learning rate, batch size, and training epochs, and performs parallel training on a graphics processing unit (GPU) for accelerated processing. In alternative implementations, generative adversarial training and knowledge distillation mechanisms can be introduced into the training process to further improve the perceptual quality of the reconstructed images and the lightweight performance of the model; federated learning can also be used to complete collaborative training of the model among multiple UAV nodes, while protecting the privacy and security of the training data.

[0140] Through multiple comparative experiments, the superiority of the technical solution in this embodiment is demonstrated.

[0141] First, for overall performance comparison, systematic experiments were conducted on the typical remote sensing image dataset AID, and the results were compared with several existing state-of-the-art methods. Specifically, under different channel conditions, our method significantly outperforms traditional coding schemes (JPEG+LDPC), deep joint source-channel coding (Deep JSCC, VAE), and the latest JSCCformer-f method. JPEG+LDPC uses a JPEG codec for compression and combines it with LDPC for channel coding; Deep JSCC is a method combining deep learning technology and joint source-channel coding; VAE is a variational autoencoder model that uses SSIM as a perception-based metric; and JSCCformer-f uses a single encoder to facilitate transmission across multiple blocks, thereby improving receiver estimation at each block.

[0142] Table 1 shows a comparison of the classification performance of different methods on the AID dataset.

[0143] Table 1

[0144] As shown in Table 1, on the AID dataset, this method achieves a maximum improvement of 6.05 dB compared to the traditional JPEG+LDPC method.

[0145] Secondly, to further verify the transmission effect, the image restoration results were visualized and compared. Transmission restoration images for typical scenes (such as mountains, bridges, ponds, industrial areas, and playgrounds) were displayed on the AID dataset, as shown below. Figure 7 As shown. Figure 7 The first three rows are visual comparison samples under an additive white Gaussian noise channel with a signal-to-noise ratio of 1dB. The last two rows are visual comparison samples under a Rayleigh fading channel with a signal-to-noise ratio of 1dB. The first two columns show the original image and the original patch. The last five columns show the reconstruction patches using different methods. Figure 7 The restoration results of the original image and each method were compared: Deep JSCC and VAE methods showed obvious blurring and loss of detail; JSCCformer-f still showed edge blurring at low signal-to-noise ratio; this method can clearly restore details and structure, and the overall visual effect is closest to the original image.

[0146] Subsequently, robustness and adaptability were verified. Under different signal-to-noise ratios (SNR) and transmission rates, the relationship between peak SNR performance and SNR was plotted in additive white Gaussian noise (AWG) and Rayleigh fading (FF) channels. Figure 8 As shown, when the signal-to-noise ratio decreases, this method can still maintain a high peak signal-to-noise ratio and multi-scale structural similarity, while the performance of the comparison method deteriorates significantly. When the transmission rate decreases, this method can maintain stable image quality through TRAM (Transmission Rate Adaptive Module), while traditional methods are prone to cliff effect.

[0147] Finally, ablation experiments were conducted to verify the contribution of key modules. Table 2 shows the results after removing CFEM (Channel Feature Enhancement Module), TRAM, or both.

[0148] Table 2

[0149] The results show that removing either module significantly reduces performance, indicating that both modules are indispensable for the effectiveness of the overall method.

[0150] In summary, the method proposed in this embodiment is innovative compared to existing technologies in that it combines residual blocks with the MobileMamba module, enabling rapid extraction of local detail features and capture of long-range spatial dependencies, thus improving feature representation capabilities while maintaining computational efficiency. Furthermore, the introduction of involution operations and adaptive signal-to-noise ratio adjustment blocks dynamically enhances useful channel features based on different channel conditions, thereby improving the system's anti-interference capability under complex channels. In addition, dynamic adjustment of feature dimensions is achieved through rate adjustment blocks and rate mask blocks, allowing the system to adaptively compress and select features according to different transmission rates, improving bandwidth utilization and adaptability. At the end-to-end joint optimization level, a composite loss function of reconstruction loss and perceptual loss is used for training to ensure that the transmitted image maintains high fidelity at both the pixel and perceptual levels.

[0151] like Figure 10 As shown in the figure, this embodiment of the invention provides a channel adaptive encoding and decoding system for UAV image transmission. The system includes: a data acquisition module 10, a lightweight feature extraction module 20, a channel enhancement module 30, a transmission rate adaptive module 40, a transmission module 50, and a receiving and reconstruction module 60.

[0152] Specifically, the data acquisition module 10 is used to acquire air-to-ground images collected by the UAV; the lightweight feature extraction module 20 is used to perform lightweight feature extraction processing on the air-to-ground images, extract local detail features of the air-to-ground images, and extract long-dependency features based on the local detail features to obtain the spatial-spectral compact features of the air-to-ground images; the channel enhancement module 30 is used to perform channel enhancement on the spatial-spectral compact features based on the real-time channel signal-to-noise ratio of the UAV air-to-ground link, and adaptively enhance or suppress different channel features to obtain channel-enhanced features; the transmission rate adaptive module 40 is used to adaptively adjust the transmission rate of the channel-enhanced features based on the real-time channel bandwidth ratio of the UAV air-to-ground link, and dynamically filter and compress the features to obtain transmission features; the transmission module 50 is used to normalize the power of the transmission features and transmit them through the air-to-ground wireless channel; the receiving and reconstruction module 60 is used to receive the damaged features after transmission through the air-to-ground wireless channel, and sequentially perform inverse recovery and reconstruction processing on the damaged features corresponding to the transmission rate adaptation, channel enhancement, and lightweight feature extraction to obtain the reconstructed air-to-ground image.

[0153] Based on the above embodiments, the present invention also provides a terminal device, the principle block diagram of which can be as follows: Figure 11As shown, the terminal device includes a processor, memory, network interface, display screen, and temperature sensor connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When executed by the processor, the computer program implements a channel-adaptive encoding / decoding method for UAV image transmission. The display screen can be an LCD screen or an e-ink screen. The temperature sensor is pre-installed inside the terminal device to detect the operating temperature of the internal components.

[0154] Those skilled in the art will understand that Figure 11 The block diagram shown is merely a partial structural diagram related to the present invention and does not constitute a limitation on the terminal device to which the present invention is applied. A specific terminal device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0155] In one embodiment, a terminal device is provided, including a memory and one or more programs, wherein one or more programs are stored in the memory and configured to be executed by one or more processors, the one or more programs including instructions for performing operations as described in the embodiments of the methods above.

[0156] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0157] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0158] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A channel-adaptive encoding and decoding method for UAV image transmission, applied to image transmission in UAV air-to-ground links, characterized in that, The method includes: Acquire aerial images of the ground captured by the drone; Lightweight feature extraction processing is performed on the open ground image to extract local detail features of the open ground image. Based on the local detail features, long dependency features are extracted to obtain the spatial-spectral compact features of the open ground image. Based on the real-time channel signal-to-noise ratio of the UAV air-to-ground link, channel enhancement is performed on the spatial-spectral compact features, and different channel features are adaptively enhanced or suppressed to obtain channel-enhanced features; Based on the real-time channel bandwidth ratio of the UAV air-to-ground link, the transmission rate of the channel enhancement features is adaptively adjusted, and the features are dynamically filtered and compressed to obtain the transmission features. The transmission characteristics are power normalized and transmitted via an air-to-ground wireless channel; The damaged features transmitted through the air-to-ground wireless channel are received, and the damaged features are sequentially subjected to inverse recovery and reconstruction processing corresponding to transmission rate adaptation, channel enhancement, and lightweight feature extraction to obtain the reconstructed air-to-ground image.

2. The channel adaptive encoding and decoding method for UAV image transmission according to claim 1, characterized in that, The lightweight feature extraction process on the open ground image, extracting local detail features of the open ground image, and extracting long-dependency features based on the local detail features to obtain the spatial-spectral compact features of the open ground image, includes: The empty ground image is subjected to convolution and activation processing to obtain the embedding features of the empty ground image; The embedded features are sequentially input into the first residual block and the second residual block to obtain the local detail features of the open ground image. The first residual block and the second residual block have the same structure, and the first residual block and the second residual block include convolution, activation and skip connection operations. A horizontal forward token sequence is generated from the local detail features through the first token generation block, and the horizontal forward token sequence is modeled with long dependency features through the first long dependency feature module to obtain the first long dependency feature. The first long-dependency feature is processed by a self-attention layer to perform feature interaction processing on global context and long-range spatial dependency, resulting in self-attention processed features. A horizontal backward token sequence is generated from the self-attention processing features through the second token generation block, and the horizontal backward token sequence is modeled with long dependency features through the second long dependency feature module to obtain the second long dependency feature. The second long-dependency feature is used as the spatial-spectral compact feature of the empty ground image.

3. The channel adaptive encoding and decoding method for UAV image transmission according to claim 2, characterized in that, Both the first long dependency feature and the second long dependency feature are Mamba features. The first long dependency feature module and the second long dependency feature module are Mobile Mamba modules with the same structure. The process of modeling the long dependency features of the token sequence using the Mobile Mamba module includes: The horizontal forward token sequence or the horizontal backward token sequence is normalized by a normalization layer to obtain normalized features. The normalized features are convolved by a one-dimensional convolutional layer to obtain the convolution result; Intermediate features are obtained by activating the convolution results using a non-linear activation function. The intermediate features are linearly transformed using a linear layer to obtain the linear transformation result; The extraction results are obtained by extracting long-dependency features from the linear transformation results using a selective state-space model. The extraction results are normalized by a normalization layer to obtain the first or second longest dependency feature.

4. The channel adaptive encoding and decoding method for UAV image transmission according to claim 1, characterized in that, The real-time channel signal-to-noise ratio based on the UAV air-to-ground link is used to enhance the spatial-spectral compact features, and to adaptively enhance or suppress different channel features to obtain channel-enhanced features, including: The real-time channel signal-to-noise ratio based on the UAV air-to-ground link is used to generate a channel weight vector through a signal-to-noise ratio adjustment block; The spatial-spectral compact features are subjected to three involution operations in sequence, and after each involution operation, the output features of the involution operation are multiplied by the channel weight vector broadcast channel by channel to obtain the involution channel features; The involute channel features are sequentially subjected to linear transformation and softmax operation to generate global channel weight parameters. The global channel weight parameters are then multiplied by the involute channel features channel by channel broadcast to obtain the channel enhancement features. The involution operations are all the same channel-dimensional adaptive feature aggregation operations, and the signal-to-noise ratio adjustment block includes multi-level fully connected layers and non-linear activation functions to map the real-time signal-to-noise ratio into a weight vector that matches the number of spatial-spectral compact feature channels.

5. The channel adaptive encoding and decoding method for UAV image transmission according to claim 1, characterized in that, The real-time channel bandwidth ratio based on the UAV air-to-ground link is used to adaptively adjust the transmission rate of the channel enhancement features, and the features are dynamically filtered and compressed to obtain transmission features, including: The real-time channel bandwidth ratio based on the UAV air-to-ground link generates a rate vector through a rate adjustment block. The channel enhancement features are subjected to three convolution operations in sequence, and after each convolution operation, the output feature of the convolution operation is multiplied by the rate vector broadcast channel by channel to obtain the convolution weighted intermediate features; Perform a convolution operation on the convolution-weighted intermediate features to obtain rate-adjusted intermediate features; The rate adjustment intermediate feature is used to generate a mask vector through a rate mask block. The mask vector is then multiplied by the channel enhancement feature broadcast channel by channel to obtain the transmission feature. The rate adjustment block includes multi-level fully connected layers and nonlinear activation functions, which are used to map the real-time channel bandwidth ratio into a rate vector that matches the number of channels of the channel enhancement feature. The rate mask block generates a mask vector through average pooling, linear transformation and nonlinear activation, and the operation mode of each convolution operation is the same.

6. The channel adaptive encoding and decoding method for UAV image transmission according to claim 1, characterized in that, The step of power normalizing the transmission characteristics and transmitting them via an air-to-ground wireless channel includes: Obtain the global element expectation of the second-order Euclidean norm of the preset transmit power and the transmission characteristics; The power normalization coefficient is calculated based on the preset transmission power and the expected global element power. The power normalization coefficient is multiplied element-wise by the transmission characteristic to obtain the normalized transmission characteristic; The normalized transmission characteristics are transmitted via the UAV air-to-ground link.

7. The channel adaptive encoding and decoding method for UAV image transmission according to claim 1, characterized in that, The end-to-end training process of the method includes: Construct a channel-adaptive coding / decoding network that matches the method described above; The training images are acquired and input into the encoding model of the encoding and decoding network. After lightweight feature extraction, channel enhancement, and adaptive transmission rate adjustment, the training transmission features are obtained. The training transmission features are subjected to power normalization to obtain training normalized transmission features; By simulating a real UAV air-to-ground link channel environment, additive white Gaussian noise interference and Rayleigh fading interference are introduced into the training normalized transmission characteristics to obtain training impaired characteristics. The training-damaged features are input into the decoding model of the encoder-decoder network, and the training-reconstructed image is obtained after inverse recovery and reconstruction processing. Based on the training image and the training reconstructed image, calculate the reconstruction loss in the pixel domain; Based on the training normalized transmission features and the training impaired features, the perceptual loss of the feature domain is calculated. Based on the reconstruction loss and perception loss, a composite loss function is constructed, and the parameters of the encoding and decoding network are iteratively optimized with the goal of minimizing the composite loss function, to obtain the trained channel adaptive encoding and decoding network.

8. A channel-adaptive encoding and decoding system for UAV image transmission, characterized in that, The system includes: The data acquisition module is used to acquire aerial and ground images collected by the drone; The lightweight feature extraction module is used to perform lightweight feature extraction processing on the ground image, extract local detail features of the ground image, extract long dependency features based on the local detail features, and obtain the spatial-spectral compact features of the ground image. The channel enhancement module is used to enhance the spatial-spectral compact features based on the real-time channel signal-to-noise ratio of the UAV air-to-ground link, and to adaptively enhance or suppress different channel features to obtain channel enhancement features; The transmission rate adaptive module is used to adaptively adjust the transmission rate of the channel enhancement features based on the real-time channel bandwidth ratio of the UAV air-to-ground link, and to dynamically filter and compress the features to obtain the transmission features. A transmission module is used to normalize the power of the transmission characteristics and transmit them through an air-to-ground wireless channel; The receiving and reconstruction module is used to receive the damaged features transmitted through the air-to-ground wireless channel, and sequentially perform reverse recovery and reconstruction processing on the damaged features corresponding to transmission rate adaptation, channel enhancement, and lightweight feature extraction to obtain the reconstructed air-to-ground image.

9. A terminal device, characterized in that, The terminal device includes a memory, a processor, and a channel adaptive encoding / decoding program for UAV image transmission stored in the memory and executable on the processor. When the processor executes the channel adaptive encoding / decoding program for UAV image transmission, it implements the steps of the channel adaptive encoding / decoding method for UAV image transmission as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a channel adaptive encoding and decoding program for UAV image transmission. When the channel adaptive encoding and decoding program for UAV image transmission is executed by a processor, it implements the steps of the channel adaptive encoding and decoding method for UAV image transmission as described in any one of claims 1-7.

Images