Unmanned aerial vehicle inspection image transmission method and system based on asymmetric codec image compression technology

Abstract

The application discloses a kind of unmanned aerial vehicle inspection image transmission method and system based on asymmetric codec.First, the sensitive area of equipment and the non-sensitive area of environment in image are identified by semantic segmentation network;Then the sensitive area of equipment is encrypted and compressed using compression sensing technology based on public key mechanism, and the non-sensitive area of environment is efficiently compressed by lightweight encoder;The processed code stream is transmitted to the ground station through wireless channel;The ground station differentiates reconstruction using high-performance decoder according to user permission level;In the training phase, the codec is jointly optimized using a hybrid objective function;Finally, the reconstruction image is used to complete the inspection state analysis and decision.The application realizes the safe and efficient transmission of inspection image, improves the transmission efficiency while ensuring the security of sensitive data, and provides a complete solution for unmanned aerial vehicle intelligent inspection.

Description

Technical Field

[0001] This invention relates to the field of image processing and transmission technology, and in particular to a method and system for transmitting UAV inspection images based on asymmetric encoding and decoding image compression technology. Background Technology

[0002] With the widespread application of drone technology in fields such as power line inspection and pipeline monitoring, how to efficiently and securely transmit and process inspection image data has become a key focus of the industry. Images collected by drones during inspections typically contain sensitive equipment information (such as details of power equipment) and non-sensitive environmental information (such as the sky and vegetation). Traditional image transmission methods often employ uniform compression and transmission strategies, which cannot balance security and transmission efficiency.

[0003] The current mainstream image transmission methods have the following shortcomings: First, they use the same compression standard for the entire image and fail to differentiate based on the importance of different areas, resulting in insufficient protection of sensitive information or low transmission efficiency. Second, they use a single encryption method, which is difficult to resist complex network attacks. Third, they lack a differentiated reconstruction mechanism based on user permissions, which cannot meet the needs of multi-level security management. Fourth, the adaptability and reconstruction quality of existing codecs in complex inspection environments need to be improved.

[0004] Furthermore, traditional methods are relatively weak in image analysis and decision support, lacking a complete solution that effectively combines image transmission with intelligent analysis. Therefore, there is an urgent need for a UAV inspection image transmission method that can achieve secure and efficient transmission, support hierarchical reconstruction, and possess intelligent analysis capabilities. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a drone inspection image transmission solution that can balance security and transmission efficiency.

[0006] In a first aspect, embodiments of this application provide a method for transmitting UAV inspection images based on asymmetric encoding and decoding image compression technology, the method comprising: S1: The drone acquires raw images of the inspection area, identifies and marks sensitive equipment areas and non-sensitive environmental areas in the images; S2: Asymmetric compression sampling and chaotic encryption processing are performed on the sensitive areas of the device using two-dimensional compressed sensing technology based on the public key mechanism. Meanwhile, the non-sensitive areas of the environment are efficiently compressed using a lightweight encoder that includes dual-channel feature extraction and detail enhancement modules. S3: Transmit the differentiated compressed bitstream to the ground station via a wireless channel; S4: The ground station uses a high-performance decoder to reconstruct the image from the received bitstream. Based on the user's permission level, it fully reconstructs all areas of content for authorized users, and only reconstructs non-sensitive areas of content for some authorized users. S5: During the training phase, a hybrid objective function combining multi-scale image fidelity measurement and pixel-level error measurement is used to jointly optimize the encoder and decoder parameters. S6: Ground stations use reconstructed images to perform inspection status analysis and decision-making.

[0007] Secondly, embodiments of this application provide a UAV inspection image transmission system based on asymmetric encoding and decoding image compression technology, applied to the UAV inspection image transmission method based on asymmetric encoding and decoding image compression technology as described in the first aspect, the system comprising: The image acquisition and region recognition module is used by the UAV to acquire raw images of the inspection area, identify and label sensitive equipment areas and non-sensitive environmental areas in the images; The asymmetric encoding processing module is used to perform asymmetric compression sampling and chaotic encryption processing on the sensitive areas of the device using two-dimensional compressed sensing technology based on the public key mechanism, while the non-sensitive areas of the environment are efficiently compressed using a lightweight encoder containing dual-channel feature extraction and detail enhancement modules. The secure transmission module is used to transmit the differentiated compressed bitstream to the ground station via a wireless channel; The high-performance decoding and reconstruction module is used by the ground station to reconstruct images from the received bitstream using a high-performance decoder. Based on the user's permission level, it can fully reconstruct all areas of content for authorized users, and only reconstruct non-sensitive areas of content for some authorized users. The model training optimization module is used to jointly optimize the encoder and decoder parameters during the training phase by employing a hybrid objective function that combines multi-scale image fidelity metrics and pixel-level error metrics. The inspection analysis and decision-making module is used by ground stations to perform inspection status analysis and decision-making using reconstructed images.

[0008] Thirdly, embodiments of this application provide an electronic device, including: processor; Memory used to store processor-executable instructions; The processor is configured to implement the UAV inspection image transmission method based on asymmetric encoding and decoding image compression technology as described in the first aspect when executing the instructions.

[0009] Fourthly, embodiments of this application provide a computer-readable storage medium storing a program that instructs a device to execute the UAV inspection image transmission method based on asymmetric encoding and decoding image compression technology as described in the first aspect.

[0010] Beneficial effects:

[0011] 1. Achieving a balance between security and transmission efficiency: By distinguishing between sensitive areas of the device and non-sensitive areas of the environment, and employing differentiated compression and encryption strategies, the security of critical device data is ensured while significantly improving overall transmission efficiency. Sensitive areas are protected by a dual approach of public-key-based compression sensing and chaotic encryption to ensure data security during transmission.

[0012] 2. A comprehensive hierarchical authorization mechanism has been established: Differentiated image reconstruction strategies are implemented based on user permission levels. Fully authorized users can obtain complete image content, while partially authorized users can only view non-sensitive areas, effectively realizing hierarchical management and access control of data.

[0013] 3. Improved image transmission reliability: Adopting an adaptive transmission mechanism, the modulation method and error correction strategy are dynamically adjusted according to the channel quality, and a transmission quality monitoring and retransmission mechanism is established to ensure the stability and reliability of transmission in complex wireless environments.

[0014] 4. Improved image reconstruction quality: By jointly optimizing the encoder and decoder through a hybrid objective function and a progressive training strategy, and combining multi-scale feature fusion reconstruction technology, the visual quality and detail preservation ability of the reconstructed images are significantly improved.

[0015] 5. A complete intelligent inspection closed loop has been constructed: image transmission and intelligent analysis are deeply integrated, and equipment defect detection, environmental risk assessment and status trend prediction are realized through multi-task analysis models. A hierarchical decision support system has been established to provide comprehensive data support and decision-making basis for inspection work.

[0016] 6. Enhanced system practicality: The data augmentation strategy and lightweight encoder designed for the characteristics of UAV inspection scenarios effectively improve the system's adaptability and generalization ability in real-world application environments, while reducing computing resource requirements. Attached Figure Description

[0017] Figure 1 This is a schematic flowchart of a UAV inspection image transmission method based on asymmetric encoding and decoding image compression technology, provided in an embodiment of this application.

[0018] Figure 2 This application provides an architecture diagram of a drone inspection image transmission system based on asymmetric encoding and decoding image compression technology.

[0019] Figure 3 A schematic diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0021] It should be noted that in the embodiments of this application, "at least one" refers to one or more, and "more than one" refers to two or more. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the specification of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application.

[0022] Based on the embodiments described in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] Example 1

[0024] Figure 1 This is a schematic flowchart of a UAV inspection image transmission method based on asymmetric encoding and decoding image compression technology, provided as an embodiment of this application. Figure 1 As shown, a method for transmitting UAV inspection images based on asymmetric encoding and decoding image compression technology includes: S1: The drone acquires raw images of the inspection area, identifies and labels sensitive equipment areas and non-sensitive environmental areas within the images. Image acquisition and area recognition complete the conversion from the physical world to digital images, and perform intelligent analysis and understanding based on the image content. The drone collects raw inspection images, and then uses semantic segmentation technology to perform pixel-level analysis of the images, distinguishing between sensitive equipment areas requiring key protection (such as core components of power equipment) and non-sensitive environmental areas that can be handled conventionally, laying the foundation for subsequent differentiated processing.

[0025] Specifically, in this embodiment, the step S1 of identifying and labeling device-sensitive areas and environment-insensitive areas in the image includes: A deep learning-based semantic segmentation network is used to perform pixel-level region division on the original inspection image, ensuring the accuracy and automation of region recognition. This avoids the low accuracy and inefficiency of traditional threshold segmentation or manual annotation methods, providing a high-quality input foundation for subsequent differential processing. Specifically, the sensitive equipment regions include insulators, transformer bushings, line hardware, and their connections; these critical components fall within the scope of the protected sensitive information. The non-sensitive environmental regions include the sky background, vegetation cover areas, and building outlines. These non-sensitive regions do not require high-intensity security processing, providing a basis for implementing efficient and low-complexity compression coding in the S2 stage.

[0026] The semantic segmentation network employs an encoder-decoder architecture. The encoder uses a ResNet-50 backbone to extract multi-scale features, leveraging its powerful deep feature extraction capabilities to ensure accurate capture of device features at different scales. The decoder fuses feature maps from different scales through a feature pyramid network and employs an attention mechanism to enhance the feature response of key regions. The feature pyramid network (FPN) effectively fuses features from different levels of the encoder, taking into account both semantic information and spatial details, improving the segmentation accuracy for small targets (such as hardware) and complex contours. The attention mechanism enables the network to focus on key regions (i.e., device regions) in the image, suppressing background interference, thereby further improving the accuracy and robustness of segmentation.

[0027] During the region labeling phase, tiered privacy labels are automatically added to identified sensitive areas of the device. Based on the device's criticality, these areas are divided into Level 1 and Level 2 sensitive areas. Level 1 sensitive areas correspond to core components and employ enhanced encryption strategies, while Level 2 sensitive areas correspond to auxiliary components and employ standard encryption strategies. This tiered labeling mechanism introduces a refined security management hierarchy based on the identification process. The automatic addition of tiered privacy labels automatically links security management strategies to the identification results, automating the process. The division into Level 1 and Level 2 sensitive areas allows for finer-grained differentiation based on the criticality of device components (e.g., core components vs. auxiliary components). This tiering is the direct prerequisite and basis for subsequent steps (such as dynamically configuring different sampling rates and encryption strategies for different levels of areas in S2), and is the core foundation for achieving differentiated security processing.

[0028] This step defines a precise, automated, and hierarchical intelligent visual perception front-end. It goes beyond simply classifying images into sensitive and non-sensitive categories; it ensures segmentation accuracy through advanced deep learning models and provides fine-grained control for subsequent processes by introducing a hierarchical mechanism. This is the primary and crucial step in achieving a balance between safety and efficiency in the entire method.

[0029] S2: For sensitive areas of the device, asymmetric compression sampling and chaotic encryption are performed using two-dimensional compressed sensing technology based on a public-key mechanism. Simultaneously, for non-sensitive environmental areas, efficient compression is achieved using a lightweight encoder incorporating dual-channel feature extraction and detail enhancement modules. Asymmetric encoding processing implements differentiated data compression and encryption based on the importance of each area, balancing security and efficiency. The use of compressed sensing + chaotic encryption asymmetric technology for sensitive areas integrates high-strength encryption during compression to ensure data security. For non-sensitive environmental areas, a high-efficiency lightweight encoder is employed, focusing on improving compression efficiency and reducing data volume. This step is crucial for achieving a balance between security and efficiency.

[0030] Specifically, in this embodiment, step S2 employs a two-dimensional compressed sensing technology based on a public-key mechanism for asymmetric compression sampling and chaotic encryption processing of the sensitive areas of the device, thereby compressing the data volume in the sensitive areas and embedding security mechanisms. These include: A two-dimensional compressed sensing sampling matrix is ​​constructed using a public-key mechanism based on elliptic curve cryptography. The sampling matrix Φ satisfies the finite isometric property, and the sampling rate is dynamically configured according to the sensitivity level of the device: a sampling rate of 0.4 is used for Level 1 sensitive areas, and a sampling rate of 0.5 is used for Level 2 sensitive areas. This part introduces the concept of asymmetric encryption into compressed sensing based on the public-key mechanism of elliptic curve cryptography. The construction of the sampling matrix is ​​bound to the public key, allowing anyone to use the public key (sampling matrix) for compressed sampling, but only authorized parties with the corresponding private key can perform correct reconstruction. This provides access control for data from the source (sampling stage). The sampling rate is dynamically configured, with 0.4 / 0.5 sampling rates configured according to the Level 1 / Level 2 sensitive areas. This reflects a refined resource allocation strategy: a lower sampling rate (higher compression ratio) is used for the more critical Level 1 areas, retaining less data during the compression stage, further increasing the difficulty of reconstruction for unauthorized parties; for less important Level 2 areas, the sampling rate is appropriately relaxed to balance security and the final reconstruction quality.

[0031] Suppose the original insulator image has 1000 pixels (N=1000). The system generates a special sampling matrix based on the public key. Because this is a level-one sensitive area, the sampling rate is set to 0.4, so only 400 measurements are collected (M=400). The image data size is reduced from 1000 to 400, achieving compression. Simultaneously, since the sampling matrix is ​​generated using the public key, this compression process itself is lock-enabled.

[0032] After compressed sampling, a three-dimensional hyperchaotic system is used to generate a random scrambling matrix to perform diffusion encryption and chaotic encryption on the sampled data. This high-strength encryption ensures the confidentiality of the compressed measurement values ​​(data) during transmission. The state equation of the three-dimensional hyperchaotic system is expressed as: , , , in, : indicates the first The three-dimensional state variables of the system at the next iteration : indicates the first The three-dimensional state variables of the system at the next iteration; , , The system parameters are derived from the hash value of the public key using a key expansion algorithm. Leveraging the extreme sensitivity of chaotic systems to initial conditions and their long-term unpredictability, a scrambled matrix with quasi-randomness is generated. Through the coupling of sine and Chebyshev mappings, the system's complex structure is difficult to predict and reconstruct, thus ensuring encryption strength. The security of the encryption system is ultimately anchored to an asymmetric cryptosystem. This strongly links the encryption process to a specific public / private key pair; only the legitimate key holder can derive the correct chaotic parameters and perform decryption.

[0033] Chaotic encryption scrambles the compressed data, turning it into a jumbled mess. Although the 400 measurements obtained in the previous step are compressed, the data patterns can still be analyzed. The system uses parameters derived from the public key (…) , , This process drives a three-dimensional hyperchaotic system, generating an extremely chaotic scrambling matrix. This matrix is ​​then used to diffusely encrypt 400 measurements, much like shuffling a deck of cards, completely disrupting the values ​​and order of the data. Even if a hacker intercepts these 400 data points, they will only see a jumble of random characters.

[0034] The encrypted sampled data is numerically normalized using modulo-256 arithmetic. The initial parameters and regional location information of the chaotic system are then embedded into the compressed sensing measurements using an orthogonal matrix, enabling covert transmission of the key data. This covert key transmission securely delivers the initial parameters of the chaotic system required for decryption to the authorized receiver. Embedding measurements using an orthogonal matrix is ​​an information hiding technique. It cleverly hides key information (initial parameters, location information) within the compressed sensing measurements without significantly altering the data structure or adding extra transmission overhead. This covert key transmission allows the receiver, after obtaining the private key, to perform reverse operations to extract these hidden parameters. For attackers, they cannot distinguish between image data and key data from the transmitted bitstream, thus enhancing the overall security and covertness of the system.

[0035] Authorized ground stations need to know the initial parameters of the chaotic system to generate the same scrambling matrix for reverse decryption. The system embeds these initial parameters and the insulator's position information in the image into the 400 encrypted scrambled data using mathematical methods (orthogonal matrices). Authorized parties with the private key can easily extract these hidden parameters from the data. Hackers, however, cannot distinguish between image data and the decryption key from the scrambled data. The public key (lock) determines who can begin decompression (reconstruction), while the hidden chaotic parameters determine who can ultimately decrypt. Both are indispensable, jointly ensuring a security effect where data is transmitted, but not completely; received, but only understandable to specific individuals.

[0036] Specifically, in this embodiment, step S2 efficiently compresses environmentally insensitive areas using a lightweight encoder that includes dual-path feature extraction and detail enhancement modules. This achieves extremely efficient compression of environmentally insensitive areas while maintaining image quality. It utilizes a carefully designed lightweight encoder that focuses on extracting and preserving the most important visual information, significantly reducing the amount of data. This includes: For environmentally insensitive areas, efficient compression is performed using a lightweight encoder, which includes: a) Dual-path feature extraction module: The first path captures fine-grained spatial features through local window division and internal pixel correlation calculation, while the second path captures semantic structural features through channel importance evaluation. The two paths of features are then aggregated to comprehensively understand the image content from both spatial detail and semantic structure dimensions.

[0037] b) Detail Enhancement Module: Employing a feedforward network with a reverse structure and a feature selection switch, this module enhances and filters key contour and texture information from the aggregated features. The fused information is refined, intelligently selecting the most valuable data for transmission and enhancing key details. The reverse feedforward network enhances the features by first expanding the feature dimensions for depth analysis, then compressing them back to the original dimensions. Nonlinear transformations are performed using activation functions such as GELU. This process enhances the expressive power of the features, particularly strengthening the clarity of cloud edges and the contrast of leaf textures, preventing these crucial details from becoming blurred during compression.

[0038] Faced with a large background of sky and vegetation, this encoder doesn't blindly transmit all pixels. Instead, it acts like an AI editor: Dual-path analysis: It sends surveyors and artists to jointly analyze the scene. Intelligence fusion: It merges detail maps and color compositions. Refinement: It uses software to sharpen key details (such as cloud edges and foliage). Intelligent filtering: The editor decisively removes redundant information describing a smooth sky, retaining only the most essential and difficult-to-reconstruct outlines and textures. Ultimately, the amount of data to be transmitted is drastically reduced, but because only the essential elements are transmitted, the ground station can still reconstruct a visually high-quality, clearly detailed background image during reconstruction.

[0039] Specifically, the efficient compression processing using a lightweight encoder includes: In the dual-path feature extraction module, the first path uses non-overlapping local window partitioning. Within each window, the covariance matrix eigenvalues ​​between pixels are calculated to capture fine-grained spatial features. The second path employs channel importance evaluation based on a channel attention mechanism, calculating the importance weights of each channel through global average pooling and fully connected layers to extract semantic structure features. The two feature paths are concatenated along the channel dimension and then convolved to achieve feature aggregation. The dual-path feature extraction module uses operable algorithms to extract spatial details and semantic structure respectively. The first path (capturing fine-grained spatial features): uses non-overlapping 8×8 local windows to partition the image. Within each window, the covariance matrix eigenvalues ​​between pixels are calculated. The eigenvalues ​​of the covariance matrix reflect the degree of pixel variation within this small block. A large eigenvalue indicates a strong change in that direction (such as a sharp edge between white clouds and blue sky); a small eigenvalue indicates a smooth area (such as a clear blue sky).

[0040] The second approach (capturing semantic structural features) employs a channel attention mechanism based on SENet. First, global average pooling is used to compress each channel into a value representing its overall importance. Then, two fully connected layers analyze these values, generating a weight between 0 and 1 for each channel. The network learns autonomously that in sky and vegetation scenes, channels representing blue and green have the highest weights because they occupy the majority of the area; while some unimportant noise channels have very low weights.

[0041] After concatenating the two feature streams, a 1×1 convolution is performed. Concatenation is like reporting work, putting detailed information and importance scores together. The 1×1 convolution is like a team leader, intelligently integrating these two types of information and deciding which to keep and which to discard.

[0042] The detail enhancement module adopts a backfeedforward network structure, in which the first layer expands the feature dimension, the middle layer uses the GELU activation function for nonlinear transformation, and the last layer compresses the feature dimension back to the original dimension; a feature selection switch based on a gating mechanism is set in the middle layer of the network. This switch generates selection weights through the Sigmoid function based on the global average pooling result of the feature map, and only features with weights higher than a preset threshold are retained for subsequent transmission.

[0043] The detail enhancement module purifies and condenses the aggregated features, transmitting only the most essential parts. A feedforward network performs a non-linear transformation using a three-layer structure: expansion (4x) → transformation (GELU) → compression (back to original dimension). Features are first deeply analyzed and reorganized in a high-dimensional space (making it easier to capture complex patterns), then compressed back into an efficient and compact form. This better represents complex textures such as cloud transitions.

[0044] At the encoder output, the enhanced features are compressed using entropy coding based on arithmetic coding. The probabilistic model adaptively updates based on the statistical characteristics of environmentally insensitive regions from historical transmissions. Entropy coding losslessly packages the selected essential features to achieve ultimate compression. Arithmetic coding, on the other hand, uses a probabilistic model that adaptively updates based on historical transmission data. Arithmetic coding is a highly efficient lossless compression technique. It assigns short codes to frequently occurring feature patterns (such as edges at various angles) and long codes to infrequent patterns. For example, since sky and vegetation backgrounds are consistently transmitted, the probabilistic model learns that smooth blue regions (although most are filtered out, their patterns still exist in the remaining data) and green textures are very common. Therefore, when encoding the final retained features, they can be represented with a very small number of bits.

[0045] The above steps thoroughly engineer an intelligent compression process: it first quantitatively analyzes the local details and global structure of the image (dual-path feature extraction), then performs intelligent refinement (detail enhancement and filtering), and finally performs efficient encoding (entropy encoding). This ensures that the drone can generate a very small but visually high-quality compressed bitstream for the background area with limited computing and bandwidth resources.

[0046] S3: Transmits the differentiated compressed bitstream to the ground station via a wireless channel. Secure transmission reliably delivers the processed data to the ground station through an unstable wireless channel. The two bitstreams are encapsulated, and adaptive modulation and error correction technologies are used to combat noise, attenuation, and interference in the wireless channel, ensuring data packet integrity and reliable transmission.

[0047] Specifically, in this embodiment, step S3 transmits the differentiated compressed bitstream to the ground station via a wireless channel. A reliable and adaptive logistics packaging and transportation system is designed for the two types of compressed and encrypted data (sensitive area bitstream and non-sensitive area bitstream) to ensure stable and efficient data transmission to the ground station in complex wireless environments. This includes: The encrypted compressed bitstream in sensitive areas of the equipment and the high-efficiency compressed bitstream in non-sensitive areas of the environment are subjected to channel coding and encapsulation processing respectively. Differentiated channel coding and encapsulation are used to provide different levels of error protection based on the importance of the data, and the data is packaged into a unified format. Among these: Type I error correction coding is used for the bitstream of sensitive areas of the equipment; Type I error correction coding (such as RS code) is used for the bitstream of sensitive areas of the equipment (such as insulator data). RS code is a powerful error correction code, and even if a certain degree of error or loss occurs during transmission, the receiving end can completely recover the original data. This provides the highest level of protection for the most critical data.

[0048] Type II error correction coding is used for bitstreams in environmentally insensitive areas; Type II error correction coding (such as LDPC codes) is used for bitstreams in environmentally insensitive areas (such as sky background data). LDPC codes are highly efficient, providing good error correction capabilities while requiring fewer redundant check bits. This is suitable for data with high real-time requirements and where a small number of errors are permissible.

[0049] Two types of bitstreams are encapsulated into transmission frames according to a preset protocol. The transmission frame structure includes a frame header, frame type identifier, data length field, payload data area, and checksum. The frame type identifier distinguishes between bitstreams from sensitive areas of the device and bitstreams from non-sensitive areas of the environment. Encapsulation into transmission frames involves packaging the two types of bitstreams according to the preset protocol. The frame type identifier is crucial; it tells the ground station: this package contains valuable porcelain (sensitive bitstream), please handle with care, or this package contains ordinary clothing (non-sensitive bitstream), which can be handled normally.

[0050] During transmission, the modulation method is adaptively adjusted according to the wireless channel quality, and the most suitable mode of transportation (modulation method) is selected in real time according to road conditions (channel quality).

[0051] When the channel quality is above the first threshold, a higher-order modulation scheme is used; when the channel quality is above the first threshold (e.g., high signal-to-noise ratio, smooth and unobstructed road), a higher-order modulation scheme (e.g., 64QAM) is used. 64QAM is like a large freight truck that can carry a lot of cargo (data) at once. When road conditions are good, it is used to transport data at high speed with the highest efficiency.

[0052] When the channel quality is between the first and second thresholds, a mid-order modulation scheme is used; when the channel quality is between the first and second thresholds (like normal road conditions), a mid-order modulation scheme (such as 16QAM) is used. 16QAM is like a medium-sized van. It has a smaller carrying capacity than a truck but a larger capacity than a car, and it can ensure a balance between stability and speed when road conditions are not ideal.

[0053] When the channel quality is below the second threshold, a low-order modulation method is used; when the channel quality is below the second threshold (e.g., low signal-to-noise ratio, rough road), a low-order modulation method (e.g., QPSK) is used. QPSK is like a rugged off-road jeep. It can only carry a small amount of cargo (data) at a time, but it is very stable and reliable, ensuring data delivery even in the worst road conditions without overturning (data corruption).

[0054] Establish a transmission quality monitoring mechanism to statistically analyze the bit error rate and packet loss rate in real time. When the bit error rate of multiple consecutive transmission frames exceeds a set threshold, an automatic transmission rate degradation and retransmission mechanism is triggered. Real-time monitoring of logistics status is also implemented; if a high risk of cargo damage is detected, an emergency plan is immediately activated. Real-time statistics of the bit error rate (cargo damage rate) and packet loss rate (cargo loss rate) are maintained. When the bit error rate of multiple consecutive transmission frames exceeds a set threshold (e.g., ... When the system determines that the current channel quality is continuously deteriorating, the following trigger mechanisms are implemented: 1. Transmission rate degradation: Automatically switch from 64QAM (large truck) to QPSK (small jeep), prioritizing data delivery even at a slower rate. 2. Retransmission mechanism: For packets (data frames) that have been corrupted or lost, request the sender to retransmit them.

[0055] This step establishes an intelligent and robust transmission assurance system. Through three main methods—differentiated packaging (channel coding), intelligent vehicle selection (adaptive modulation), and end-to-end monitoring and emergency handling (quality monitoring and retransmission)—it ensures that both important sensitive data and large amounts of non-sensitive data can safely, reliably, and efficiently reach their destination in complex and ever-changing wireless channels.

[0056] S4: The ground station uses a high-performance decoder to reconstruct images from the received bitstream. Based on user access levels, authorized users fully reconstruct all areas, while partially authorized users only reconstruct non-sensitive environmental areas. This differentiated reconstruction, based on user permissions, enables hierarchical decryption and reconstruction of data content at the receiving end. This is the ultimate manifestation of the access control policy. Fully authorized users can fully reconstruct all areas and obtain all information; partially authorized users can only reconstruct non-sensitive environmental areas, while sensitive areas remain encrypted or obscured, thus achieving secure hierarchical information sharing.

[0057] The ground station acts as an intelligent receiving and restoration center. Based on different authorization credentials (user permissions), it parses, reconstructs, and seamlessly stitches received data packets, ultimately restoring a clear image that meets the permission requirements. Specifically, in this embodiment, in step S4, the ground station performs image reconstruction using a high-performance decoder, including: The high-performance decoder adopts a multi-scale feature fusion architecture based on an attention mechanism, including a feature parsing module, a region reconstruction module, and an image fusion module; wherein: The feature parsing module performs channel decoding and protocol parsing on the received bitstream, separating the device-sensitive area bitstream and the environment-non-sensitive area bitstream based on the frame type identifier. The feature parsing module acts as a warehouse sorter, unpacking and classifying the delivered goods. It performs channel decoding (unpacking the external shipping packaging) and protocol parsing (checking the waybill) on the received bitstream. Based on the frame type identifier, the module accurately separates the bitstream: this box contains insulator encrypted data (sensitive bitstream), requiring special handling; that box contains sky background data (non-sensitive bitstream), which follows the standard procedure.

[0058] The domain reconstruction module acts as a product restorer, implementing permission-based differentiated restoration strategies. The domain reconstruction module implements differentiated reconstruction strategies for users with different permissions, including: For fully authorized users (such as senior engineers), a joint reconstruction approach is used. Sensitive areas of the device's bitstream are decrypted and reconstructed using a private key-based 2D compressed sensing reconstruction algorithm, while non-sensitive areas of the environment are reconstructed using a convolutional neural network-based super-resolution reconstruction. Specifically, for sensitive bitstreams: the private key-based 2D compressed sensing reconstruction algorithm is used. This algorithm uses the private key for decryption and, combined with hidden chaotic initial parameters, reverses the chaotic system, much like using a unique key to open a safe and perfectly reconstructing shredded paper into a document according to hidden instructions, ultimately obtaining a clear image of the insulator. For non-sensitive bitstreams: super-resolution reconstruction based on a convolutional neural network is used. This is like using powerful image restoration software to enhance details and increase resolution in a compressed background image, resulting in a clearer sky and vegetation. Senior engineers can then see a complete and high-definition inspection image.

[0059] For authorized users (such as public viewers), only the non-sensitive areas of the environment are reconstructed, while sensitive areas of the device remain encrypted and are displayed in a masked form in the reconstructed image. Alternatively, only the non-sensitive areas are reconstructed, while sensitive areas remain encrypted and are displayed in a masked form (such as a mosaic) in the reconstructed image. Public viewers can only see a clear background, while critical device areas are masked, thus sharing environmental information while protecting core confidential information.

[0060] The image fusion module employs an adaptive weighted fusion algorithm to seamlessly stitch together the reconstructed sensitive and non-sensitive areas. A weighted averaging algorithm is used at the area boundaries to eliminate stitching artifacts. The image fusion module acts as a seamless stitcher, perfectly combining the two separately reconstructed images into a single image. The adaptive weighted fusion algorithm uses a weighted averaging algorithm at area boundaries. If the reconstruction is fully authorized by the user, a clear insulator image and a clear background image are obtained. Direct stitching might result in a harsh, knife-cut look at the insulator edges. The fusion module performs a weighted averaging of the pixels from both images at the boundary between the insulator and the sky. The closer to the inside of the insulator, the higher the weight of the insulator image; the closer to the outer sky, the higher the weight of the sky image. This achieves seamless stitching, eliminating stitching artifacts and making the entire image appear as a seamless whole.

[0061] During the reconstruction process, the decoder verifies data integrity based on the checksum in the transmission frame and requests retransmission of data frames that fail verification. Data integrity verification acts as a quality inspector, ensuring that the raw materials (data) used for reconstruction are intact. The decoder verifies based on the checksum (such as CRC) in the transmission frame. Before sorting and reconstruction, the checksum of the received data is calculated and compared with the checksum embedded in the transmission frame. If the comparison fails, it indicates that the data packet was corrupted during transmission (equivalent to a damaged package), and the decoder immediately requests the drone to retransmit the corrupted data packet. This ensures that the data used to reconstruct the image is 100% correct, guaranteeing the reliability of the final reconstructed image from the source.

[0062] The defined ground station decoder is a high-performance processing system that integrates intelligent sorting, access control restoration, seamless stitching, and quality inspection. It ensures that users with different access levels receive high-quality image results matching their identities, while strictly guaranteeing data security and integrity.

[0063] S5: During the training phase, a hybrid objective function combining multi-scale image fidelity metrics and pixel-level error metrics is employed to jointly optimize the encoder and decoder parameters. Model training optimization, through offline training, continuously improves the performance of the encoder and decoder. This is the core driver ensuring the overall system performance (such as image reconstruction quality and compression efficiency). By designing a hybrid objective function and a progressive training strategy, the parameters of the encoder and decoder are jointly optimized, enabling the system to learn how to achieve the optimal balance between compression ratio, security, and reconstruction quality.

[0064] Specifically, in this embodiment, step S5 employs a hybrid objective function to jointly optimize the encoder and decoder parameters. The hybrid objective function establishes comprehensive evaluation criteria, using a holistic scoring standard to guide the model's learning direction, rather than focusing on a single indicator. This includes: The hybrid objective function consists of a multi-scale structural similarity loss. Pixel-level mean square error loss Feature perception loss and region adaptive loss The four-item weighted sum is expressed as follows: , Among them, multi-scale structural similarity loss The similarity between the reconstructed image and the original image in terms of brightness, contrast, and structure is calculated at five different scales, with weighting coefficients... An adaptive adjustment strategy was adopted, with an initial value set to 0.4. The study evaluated whether the brightness, contrast, and structure of the restored image remained consistent with the original image at different magnification levels. Pixel-level mean square error loss was also assessed. Calculate the mean square error of image pixel values ​​and the weighting coefficients. Set to 0.3 to rigorously assess whether the color value of each pixel is accurately reproduced. Feature-aware loss. Features are extracted and differences are calculated using a pre-trained VGG-19 network at the ReLU3_3 layer, along with weight coefficients. Set to 0.2, and have a senior expert (using a pre-trained VGG network) evaluate the similarity between the restored image and the original image in high-level semantic features (e.g., whether both have the fluffy feel of clouds or the texture of leaves). Region adaptive loss. Calculate the local reconstruction quality difference between sensitive and non-sensitive areas separately, and assign weights to the relevant factors. Set the value to 0.1 for differentiated assessment, and check the restoration quality of sensitive and non-sensitive areas separately to ensure that the system is not biased.

[0065] A progressive training strategy is adopted during the optimization process: The first stage uses an optimizer with weight decay to focus on optimizing the compression and reconstruction quality of non-sensitive environmental regions with a first learning rate. In the first stage, a solid foundation is laid (focusing on optimizing non-sensitive regions). An optimizer with weight decay (such as AdamW) is used with a high first learning rate to allow the model to learn how to efficiently compress and clearly reconstruct the background (such as sky and vegetation).

[0066] The second stage introduces the sensitive region reconstruction task, reducing the learning rate to the second learning rate and adding an edge protection loss term to improve the ability to preserve edge details in sensitive areas of the equipment. The second stage enhances expertise (introducing the sensitive region task). The learning rate is reduced to the second learning rate, the sensitive region reconstruction task is introduced, and an edge protection loss term is added. Building upon the existing foundation, the model begins to learn to process complex images containing critical equipment. The reduced learning rate is for fine-tuning. The edge protection loss forces the model to pay special attention to key details such as insulator edges and hardware outlines, preventing them from becoming blurred.

[0067] The third stage involves end-to-end joint fine-tuning. The learning rate is decayed using a cosine annealing strategy, while a gradient clipping strategy is introduced to limit the gradient norm. This third stage is a full-process collaborative optimization (end-to-end joint fine-tuning), connecting editors and restorers for end-to-end joint fine-tuning. The learning rate uses a cosine annealing strategy (smoothly decaying from a high value to near 0), and gradient clipping is introduced. Joint fine-tuning: allows editors and restorers to work together to find the optimal collaboration method. Cosine annealing: like simulated annealing, helps the model escape local optima and find a better global solution. Gradient clipping: prevents excessively large learning steps (gradients) during training, which could lead to model collapse (training instability), ensuring a smooth training process.

[0068] Training data augmentation includes targeted enhancement strategies based on the characteristics of UAV inspection scenarios: random rotation, brightness adjustment, Gaussian noise addition, atmospheric turbulence simulation, and sensor noise simulation. Training data augmentation creates complex simulated combat environments by artificially generating various complex image scenes to improve the model's adaptability and robustness in the real world. This includes random rotation, brightness adjustment, Gaussian noise addition, atmospheric turbulence simulation, and sensor noise simulation. The codec, trained in this rigorous manner, can still operate stably and reliably when encountering various harsh inspection environments in actual operations.

[0069] This solution describes a scientific, systematic, and efficient model training system. By setting comprehensive evaluation objectives, adopting progressive training phases, and using a highly realistic data environment, it ensures that the final deployed codec can intelligently complete image compression and reconstruction tasks, performing excellently in real-world drone inspection applications.

[0070] S6: Ground stations utilize reconstructed images to perform inspection status analysis and decision-making. Intelligent analysis and decision-making transform reconstructed images into valuable maintenance decisions. Moving beyond traditional transmission-display models, deep learning models are used to automatically analyze reconstructed images, enabling defect detection, risk assessment, and trend prediction. This directly generates maintenance work orders, early warning information, and other decision support, forming a complete closed loop from data acquisition to intelligent action.

[0071] Specifically, in this embodiment, the implementation of the ground station using reconstructed images to complete inspection status analysis and decision-making in step S6 includes: A deep learning-based multi-task analysis model is established, which performs the following analysis tasks in parallel: Equipment defect detection: The YOLOv5 architecture is used to detect defects such as insulator breakage, conductor strand breakage, and hardware corrosion in sensitive areas of equipment, with a detection confidence threshold set at 0.75; Environmental risk assessment: The U-Net segmentation network is used to identify risks such as vegetation encroachment and building intrusion in non-sensitive areas of the environment, and the risk level is calculated.

[0072] Status Trend Prediction: Based on reconstructed images from continuous time series, an LSTM network is used to predict equipment status change trends. A hierarchical decision support system is constructed to automatically generate handling suggestions based on the analysis results: For general defects detected, the system generates maintenance work orders and automatically arranges inspection plans; for identified major hidden dangers, real-time alarms are immediately triggered and pushed to mobile terminals; based on the status prediction results, preventive maintenance time windows are intelligently recommended; a visualization feedback mechanism for the analysis results is established, overlaid on the reconstructed images: different colors are used to mark the equipment defect level (red for urgent, yellow for important, and green for normal); a heat map is used to display the distribution of environmental risks; trend curves are used to display the predicted equipment status changes; the analysis and decision results are correlated with historical databases to establish a full life cycle health record for the equipment, providing data support for subsequent inspection strategy optimization.

[0073] These six steps constitute a complete technology chain from data acquisition → intelligent understanding → differentiated processing → reliable transmission → on-demand reconstruction → value extraction, systematically solving the comprehensive challenges of security, efficiency, and intelligent application in UAV inspection image transmission.

[0074] Example 2

[0075] like Figure 2 As shown in the figure, this application provides an architecture diagram of a UAV inspection image transmission system based on asymmetric encoding and decoding image compression technology, which is applied to the UAV inspection image transmission system based on asymmetric encoding and decoding image compression technology as described in Embodiment 1. It includes an image acquisition and region recognition module 11, an asymmetric encoding processing module 12, a secure transmission module 13, a high-performance decoding and reconstruction module 14, a model training and optimization module 15, and an inspection analysis and decision module 16.

[0076] The image acquisition and region recognition module 11 is used by the UAV to acquire the original images of the inspection area, identify and label the equipment sensitive areas and environmental non-sensitive areas in the images.

[0077] The asymmetric encoding processing module 12 is used to perform asymmetric compression sampling and chaotic encryption processing on the sensitive areas of the device using two-dimensional compressed sensing technology based on a public key mechanism, while efficiently compressing the non-sensitive areas of the environment using a lightweight encoder containing dual-channel feature extraction and detail enhancement modules.

[0078] The secure transmission module 13 is used to transmit the differentiated compressed bitstream to the ground station via a wireless channel.

[0079] The high-performance decoding and reconstruction module 14 is used by the ground station to reconstruct the image from the received bitstream using a high-performance decoder. Based on the user's permission level, it fully reconstructs all areas of content for authorized users, and only reconstructs non-sensitive areas of content for some authorized users.

[0080] The model training optimization module 15 is used to jointly optimize the encoder and decoder parameters during the training phase by employing a hybrid objective function that combines multi-scale image fidelity metrics and pixel-level error metrics.

[0081] The inspection analysis and decision-making module 16 is used by ground stations to perform inspection status analysis and decision-making using reconstructed images.

[0082] Figure 3 This is an electronic device provided in one embodiment of this application. For example... Figure 3 As shown, the electronic device includes at least the following components: processor 101 and memory 100, communication interface 103, and bus 102.

[0083] In this embodiment of the application, memory 100 is used to store executable instructions of processor 101, which, when configured to execute instructions, implements the method as described in the first aspect.

[0084] In embodiments of this application, a computer-readable storage medium includes instructions that instruct a device to perform the method as described in the first aspect. For example, the instructions instruct the device to perform... Figure 1 The method is shown in the process steps.

[0085] In one embodiment of this application, the program operating in the electronic device may be a program that controls a central processing unit (CPU) or similar device to achieve the functions of the above-described embodiments of the present invention (a program that enables the computer to function). Information processed by these systems is then temporarily stored in random access memory (RAM) during processing, and subsequently stored in various ROMs such as read-only memory (FlashROM) and hard disk drives (HDDs), and read, corrected, and written by the CPU as needed.

[0086] It should be noted that a portion of the electronic device described above can also be implemented using a computer. In this case, the program for implementing the control function can be recorded on a computer-readable recording medium, and the program recorded on the recording medium can be read into the computer and executed.

[0087] It should be noted that the computer mentioned here refers to a computer built into an electronic device, employing hardware including an operating system and peripheral devices. Furthermore, computer-readable recording media refers to removable media such as floppy disks, magneto-optical disks, ROMs, and CD-ROMs, as well as storage systems such as hard drives built into the computer.

[0088] Furthermore, computer-readable recording media can include: media that dynamically stores programs for short periods of time, such as communication lines used when transmitting programs via networks like the Internet or communication lines like telephone lines; and media that store programs for fixed periods of time, such as volatile memory inside a computer that serves as a server or client in this case. In addition, the aforementioned program can be a program used to implement the above-mentioned functions, or it can be a program that can implement the above-mentioned functions by combining them with programs already recorded in the computer.

[0089] Furthermore, the electronic device in the above embodiments can also be implemented as an assembly (system group) composed of multiple systems. Each system constituting the system group can possess some or all of the functions or functional blocks of the electronic device in the above embodiments. As a system group, it is sufficient to have all the functions or functional blocks of the electronic device.

[0090] Those skilled in the art should recognize that the above embodiments are only used to illustrate this application and are not intended to limit this application. Any appropriate changes and variations made to the above embodiments within the essential spirit and scope of this application fall within the scope of protection claimed in this application.

Claims

1. A method for transmitting UAV inspection images based on asymmetric encoding and decoding image compression technology, characterized in that, Includes the following steps: S1: The drone acquires raw images of the inspection area, identifies and marks sensitive equipment areas and non-sensitive environmental areas in the images; S2: Asymmetric compression sampling and chaotic encryption processing are performed on the sensitive areas of the device using two-dimensional compressed sensing technology based on the public key mechanism. Meanwhile, the non-sensitive areas of the environment are efficiently compressed using a lightweight encoder that includes dual-channel feature extraction and detail enhancement modules. In step S2, asymmetric compression sampling and chaotic encryption processing are performed on the sensitive areas of the device using a two-dimensional compressed sensing technology based on a public key mechanism, including: A two-dimensional compressed sensing sampling matrix is ​​constructed using a public-key mechanism based on elliptic curve cryptography. The sampling matrix Φ satisfies the finite isometry property, and the sampling rate is dynamically configured according to the level of the sensitive area of ​​the device: a sampling rate of 0.4 is used for the first-level sensitive area, and a sampling rate of 0.5 is used for the second-level sensitive area. After compression sampling is completed, a three-dimensional hyperchaotic system is used to generate a random scrambling matrix to diffuse and encrypt the sampled data. The state equation of the three-dimensional hyperchaotic system is expressed as: ， ， ， in, : indicates the first The three-dimensional state variables of the system at the next iteration : indicates the first The three-dimensional state variables of the system at the next iteration; , , The system parameters are derived from the hash value of the public key using a key expansion algorithm; The encrypted sampled data is numerically normalized using modulo 256 operations, and the initial parameters and regional location information of the chaotic system are embedded into the compressed sensing measurement values ​​through an orthogonal matrix to achieve the covert transmission of key data. In S2, environmentally insensitive regions are efficiently compressed using a lightweight encoder that includes dual-path feature extraction and detail enhancement modules, including: For environmentally insensitive areas, efficient compression is performed using a lightweight encoder, which includes: a) Dual-path feature extraction module: The first path captures fine-grained spatial features through local window partitioning and internal pixel association calculation, while the second path captures semantic structural features through channel importance evaluation. The two paths of features are then aggregated. b) Detail enhancement module: Employs a feedforward network with an inverse structure and a feature selection switch to enhance and filter key contour and texture information in aggregated features; S3: Transmit the differentiated compressed bitstream to the ground station via a wireless channel; S4: The ground station uses a high-performance decoder to reconstruct the image from the received bitstream. Based on the user's permission level, it fully reconstructs all areas of content for authorized users, and only reconstructs non-sensitive areas of content for some authorized users. S5: During the training phase, a hybrid objective function combining multi-scale image fidelity measurement and pixel-level error measurement is used to jointly optimize the encoder and decoder parameters; the hybrid objective function is composed of four weighted terms: multi-scale structural similarity loss, pixel-level mean square error loss, feature perception loss and region adaptation loss. S6: Ground stations use reconstructed images to perform inspection status analysis and decision-making.

2. The UAV inspection image transmission method based on asymmetric encoding and decoding image compression technology according to claim 1, characterized in that, The process of identifying and labeling device-sensitive areas and environment-insensitive areas in the image in S1 includes: A deep learning-based semantic segmentation network is used to perform pixel-level region division on the original inspection image, where: The sensitive areas of the equipment include power equipment insulators, transformer bushings, line fittings and their connection parts; The environmentally non-sensitive areas include the sky background, vegetation cover areas, and building outlines; The semantic segmentation network adopts an encoder-decoder architecture, in which the encoder uses a ResNet-50 backbone network to extract multi-scale features, and the decoder fuses feature maps of different scales through a feature pyramid network and uses an attention mechanism to enhance the feature response of key regions. During the region labeling phase, hierarchical privacy labels are automatically added to the identified sensitive areas of the device. Based on the criticality of the device, the areas are divided into Level 1 sensitive areas and Level 2 sensitive areas. Level 1 sensitive areas correspond to the core components of the device and are subject to enhanced encryption strategies, while Level 2 sensitive areas correspond to the auxiliary components of the device and are subject to standard encryption strategies.

3. The UAV inspection image transmission method based on asymmetric encoding and decoding image compression technology according to claim 1, characterized in that, The efficient compression processing via a lightweight encoder includes: In the dual-path feature extraction module, the first path uses non-overlapping local window partitioning, and captures fine-grained spatial features by calculating the covariance matrix eigenvalues ​​between pixels within each window; the second path uses channel importance evaluation based on channel attention mechanism, and calculates the importance weights of each channel through global average pooling and fully connected layers to extract semantic structure features; the two paths of features are concatenated by channel dimensions and then convolved to achieve feature aggregation; The detail enhancement module adopts a backfeedforward network structure, in which the first layer expands the feature dimension, the middle layer uses the GELU activation function for nonlinear transformation, and the last layer compresses the feature dimension back to the original dimension; a feature selection switch based on a gating mechanism is set in the middle layer of the network. This switch generates selection weights based on the global average pooling result of the feature map via the Sigmoid function, and only features with weights higher than a preset threshold are retained for subsequent transmission. At the encoder output, the enhanced features are compressed using an entropy coding technique based on arithmetic coding, where the probability model is adaptively updated based on the statistical characteristics of environmental non-sensitive regions from historical transmission.

4. The UAV inspection image transmission method based on asymmetric encoding and decoding image compression technology according to claim 1, characterized in that, In step S3, the compressed bitstream, after differential processing, is transmitted to the ground station via a wireless channel, including: Channel coding and encapsulation processes are performed on the encrypted compressed bitstream in the sensitive area of ​​the device and the high-efficiency compressed bitstream in the non-sensitive area of ​​the environment, respectively. Type I error correction coding is used for the bitstream in the sensitive areas of the device; Type II error correction coding is used for the bitstream in non-sensitive environmental areas; Two types of bitstreams are encapsulated into transmission frames according to a preset protocol. The transmission frame structure includes a frame header, a frame type identifier, a data length field, a payload data area, and a checksum. The frame type identifier is used to distinguish between bitstreams in sensitive areas of the device and bitstreams in non-sensitive areas of the environment. During transmission, the modulation scheme is adaptively adjusted according to the quality of the wireless channel; Establish a transmission quality monitoring mechanism to statistically analyze the bit error rate and packet loss rate in real time. When the bit error rate of multiple consecutive transmission frames exceeds the set threshold, the transmission rate degradation and retransmission mechanism will be automatically triggered.

5. The UAV inspection image transmission method based on asymmetric encoding and decoding image compression technology according to claim 1, characterized in that, In S4, the ground station performs image reconstruction using a high-performance decoder, including: The high-performance decoder adopts a multi-scale feature fusion architecture based on an attention mechanism, including a feature parsing module, a region reconstruction module, and an image fusion module; wherein: The feature parsing module performs channel decoding and protocol parsing on the received bitstream, and separates the device sensitive area bitstream and the environment non-sensitive area bitstream according to the frame type identifier; The region reconstruction module implements differentiated reconstruction strategies for users with different permissions; The image fusion module uses an adaptive weighted fusion algorithm to seamlessly stitch together the reconstructed sensitive and non-sensitive regions, wherein a weighted average algorithm is used at the region boundaries to eliminate stitching marks. During the reconstruction process, the decoder verifies data integrity based on the checksum in the transmitted frame and requests retransmission for data frames that fail the verification.

6. The UAV inspection image transmission method based on asymmetric encoding and decoding image compression technology according to claim 1, characterized in that, The S5 method employs a hybrid objective function to jointly optimize the encoder and decoder parameters, including: A progressive training strategy is adopted during the optimization process: The first stage uses an optimizer with weight decay to focus on optimizing the compression and reconstruction quality of environmentally insensitive regions with the first learning rate. The second stage introduces a reconstruction task for sensitive areas, reduces the learning rate to the second learning rate, and adds an edge protection loss term to improve the device's ability to preserve edge details in sensitive areas. In the third stage, end-to-end joint fine-tuning is performed, with the learning rate decaying using a cosine annealing strategy, while a gradient pruning strategy is introduced to limit the gradient norm. Training data augmentation includes targeted enhancement strategies based on the characteristics of UAV inspection scenarios: random rotation, brightness adjustment, Gaussian noise addition, atmospheric turbulence simulation, and sensor noise simulation.

7. The UAV inspection image transmission method based on asymmetric encoding and decoding image compression technology according to claim 1, characterized in that, In S6, the ground station uses reconstructed images to perform inspection status analysis and decision-making, including: Establish a multi-task analysis model to execute the following analysis tasks in parallel: Equipment defect detection: identify component defects in sensitive areas of equipment; Environmental risk assessment: identify potential risks in non-sensitive areas of the environment and assess the risk level; Status trend prediction: predict the status change trend of equipment based on time series images. Construct a hierarchical decision support system to generate handling suggestions based on analysis results: generate maintenance work orders and arrange inspection plans for general defects; trigger real-time alarms and push information for major hidden dangers; and recommend preventive maintenance time windows based on prediction results. Establish a visualization feedback mechanism to overlay analysis results onto reconstructed images: use color coding to label equipment defect levels; employ heatmaps to display environmental risk distribution; and use trend curves to display state change predictions. By linking the analysis results with historical data, a health record for the entire life cycle of the equipment can be established.

8. A UAV inspection image transmission system based on asymmetric encoding and decoding image compression technology, applied to the UAV inspection image transmission method based on asymmetric encoding and decoding image compression technology as described in any one of claims 1 to 7, characterized in that, The system includes: The image acquisition and region recognition module is used by the UAV to acquire raw images of the inspection area, identify and label sensitive equipment areas and non-sensitive environmental areas in the images; The asymmetric encoding processing module is used to perform asymmetric compression sampling and chaotic encryption processing on the sensitive areas of the device using two-dimensional compressed sensing technology based on the public key mechanism, while the non-sensitive areas of the environment are efficiently compressed using a lightweight encoder containing dual-channel feature extraction and detail enhancement modules. The process of using a two-dimensional compressed sensing technology based on a public-key mechanism to perform asymmetric compressed sampling and chaotic encryption processing on sensitive areas of the device includes: A two-dimensional compressed sensing sampling matrix is ​​constructed using a public-key mechanism based on elliptic curve cryptography. The sampling matrix Φ satisfies the finite isometry property, and the sampling rate is dynamically configured according to the level of the sensitive area of ​​the device: a sampling rate of 0.4 is used for the first-level sensitive area, and a sampling rate of 0.5 is used for the second-level sensitive area. After compression sampling is completed, a three-dimensional hyperchaotic system is used to generate a random scrambling matrix to diffuse and encrypt the sampled data. The state equation of the three-dimensional hyperchaotic system is expressed as: ， ， ， in, : indicates the first The three-dimensional state variables of the system at the next iteration : indicates the first The three-dimensional state variables of the system at the next iteration; , , The system parameters are derived from the hash value of the public key using a key expansion algorithm; The encrypted sampled data is numerically normalized using modulo 256 operations, and the initial parameters and regional location information of the chaotic system are embedded into the compressed sensing measurement values ​​through an orthogonal matrix to achieve the covert transmission of key data. The environmentally insensitive regions are efficiently compressed using a lightweight encoder that includes dual-path feature extraction and detail enhancement modules, including: For environmentally insensitive areas, efficient compression is performed using a lightweight encoder, which includes: a) Dual-path feature extraction module: The first path captures fine-grained spatial features through local window partitioning and internal pixel association calculation, while the second path captures semantic structural features through channel importance evaluation. The two paths of features are then aggregated. b) Detail enhancement module: Employs a feedforward network with an inverse structure and a feature selection switch to enhance and filter key contour and texture information in aggregated features; The secure transmission module is used to transmit the differentiated compressed bitstream to the ground station via a wireless channel; The high-performance decoding and reconstruction module is used by the ground station to reconstruct images from the received bitstream using a high-performance decoder. Based on the user's permission level, it can fully reconstruct all areas of content for authorized users, and only reconstruct non-sensitive areas of content for some authorized users. The model training optimization module is used to jointly optimize the encoder and decoder parameters during the training phase by employing a hybrid objective function that combines multi-scale image fidelity measurement and pixel-level error measurement. The hybrid objective function consists of four weighted components: multi-scale structural similarity loss, pixel-level mean square error loss, feature perception loss, and region adaptation loss. The inspection analysis and decision-making module is used by ground stations to perform inspection status analysis and decision-making using reconstructed images.

Images