Image compression and recovery method for unmanned aerial vehicle power inspection
By performing lossless compression and terminal decoding and recovery of inspection images on the drone's end, the problems of slow data transmission and large storage requirements during power inspection of drone are solved, fast transmission and real-time identification are achieved, and the safety and efficiency of power grid operation are improved.
Patent Information
- Application Number
- CN202510678787.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-26
- Publication Date
- 2025-07-18
AI Technical Summary
During the drone power inspection, a large amount of inspection image data collected leads to high data communication costs and storage needs, and the transmission speed is slow, affecting the intelligent identification cycle and endangering the safety of the power grid operation.
The image compression algorithm model is used to compress the original patrol image losslessly, and combine pruning, weight sharing and quantization to reduce the amount of model parameters; compress it on the drone side, decoding and recovery of the terminal equipment, and use neural network technology and a variety of filtering methods to ensure image quality.
While ensuring image quality and intelligent recognition effect, greatly compress the data size, improve transmission speed and recognition cycle, reduce storage needs, and improve grid operation safety.
Smart Images
Figure QLYQS_1 
Figure QLYQS_2 
Figure QLYQS_3
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of image processing, and particularly to an image compression and restoration method for UAV power inspection. Background Technique
[0002] UAV power inspection is to use a UAV to inspect the power grid, collect inspection images of power facilities through the UAV, and the UAV uploads the inspection images to the server through a mobile network for terminal devices to access the server to view the inspection images. UAV power inspection can greatly improve the efficiency of power inspection.
[0003] However, a large number of inspection images need to be collected during the UAV power inspection process, which will result in ultra-large amounts of image data, high data communication costs, and a large investment in storage devices. Moreover, during the upload process, the large dataset will cause a great time delay to the intelligent recognition means deployed in the background due to the slow transmission speed, which has a great impact on the recognition cycle in various fields and is not conducive to the safe and stable operation of the large power grid. Summary of the Invention
[0004] To solve the defects of the prior art, the present invention provides an image compression and restoration method for UAV power inspection, including the following steps: the UAV inspects the power grid and collects the original inspection images of power facilities; the image compression algorithm model built in the UAV performs lossless compression on the original inspection images to obtain compressed inspection images; the UAV uploads the compressed inspection images to the server through a mobile network; the terminal device downloads the compressed inspection images from the server, and the image decoding and restoration module built in the terminal device decodes and restores the compressed inspection images to obtain decoded inspection images.
[0005] Preferably, the image compression algorithm model is lightweight processed, combined with pruning, weight sharing, quantization, etc. to reduce the model parameter quantity and size, so that the model is applicable to the UAV.
[0006] Preferably, the image compression algorithm model built in the UAV performs lossless compression on the original inspection images, including the following steps: adopting a transformation step to convert the data from the spatial domain to the frequency domain to discover redundant information; performing a quantization operation on the data to reduce the data representation precision and thus compress the data; using methods such as Huffman coding to achieve efficient storage of the data.
[0007] Preferably, the image decoding and recovery module is obtained through the following steps: 1) model initialization: setting the neural network model and confirming the number of residual blocks; 2) data preparation: reading in the compressed data and normalizing it; 3) model training: using random block cutting for training, using the stochastic gradient descent method to solve the parameters w and b of the neural network model, with the learning rate set to 0.001, and calculating the gradient and updating the model parameters during each iteration; after 100 iterations, using larger blocks to fine tune the training results.
[0008] Preferably, the image decoding and recovery module decodes and recovers the compressed inspection image, including the following steps: adopting a block reasoning method, merging different block data during output, using a block segmentation method with overlapping areas, and in the block segmentation process, an overlapping block reasoning scheme with a certain number of pixels on the edges of adjacent blocks, dividing the original image into a series of blocks with overlapping areas, and after inputting into the recovery network, splicing the output block images at a fixed position, using the output data of the block that is closer to the overlapping part, or using the distance-weighted data of all blocks with the overlapping part.
[0009] For more specific details of the image compression algorithm model and the image decoding and recovery module of the present invention, please refer to the specific implementation manner.
[0010] The advantages and beneficial effects of the present invention are: providing an image compression and recovery method for UAV power inspection, which can greatly compress the inspection image data size while ensuring the inspection image quality and the intelligent identification effect of hidden dangers, and then under limited bandwidth, make the transmission speed of uploaded inspection images faster and the defect identification cycle shorter, ultimately making the safer operation of the power grid more secure.
[0011] The present invention also has the following characteristics: 1) The present invention adopts an image compression algorithm model to perform lossless compression on the original inspection image, utilizes neural network technology and combines lossless compression algorithm to retain the complete image information of the original inspection image as much as possible, and seeks a high compression ratio, which can reach 10 times the compression ratio; 2) The present invention also performs lightweight processing on the image compression algorithm model, which can greatly reduce the model's hardware performance requirements on the edge, making the model suitable for drones; 3) The present invention compresses the inspection image before uploading it to the server. Under limited bandwidth, the transmission speed of the uploaded inspection image is faster, the defect recognition cycle is shorter, and the operation safety of the power grid is finally more guaranteed. The data set has a great influence on the background intelligent recognition cycle due to the transmission rate. Realizing fast transmission and real-time processing is of great significance to the operation safety of large power grids. 4) The server of the present invention only needs to store the compressed inspection images, which can greatly reduce the consumption of storage capacity and extend the traceability cycle; 5) The present invention uses an image decoding and restoration module to decode and restore the compressed inspection images. The image decoding and restoration combines various means such as inverse filtering method, Wiener filtering method, constrained least squares filtering method, and super-resolution algorithm based on deep learning to form an image restoration algorithm that meets the requirements of the power industry, ensuring that the restored images are suitable for intelligent recognition applications in power industry scenarios. 6) The compressed file of the present invention can be in a private format that cannot be interpreted by a third party, improving the security during storage and transmission. Specific embodiments
[0012] The following combines embodiments to further describe the specific embodiments of the present invention. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and cannot be used to limit the protection scope of the present invention.
[0013] The specific technical solutions implemented by the present invention are as follows: The present invention provides an image compression and restoration method for drone power inspection, including the following steps: The drone inspects the power grid and collects the original inspection images of power facilities; the image compression algorithm model built in the drone performs lossless compression on the original inspection images to obtain compressed inspection images; the drone uploads the compressed inspection images to the server through a mobile network; the terminal device downloads the compressed inspection images from the server, and the image decoding and restoration module built in the terminal device decodes and restores the compressed inspection images to obtain decoded inspection images. Specifically: 1. Regarding the image compression algorithm model Lightweight processing is performed on the image compression algorithm model by combining pruning, weight sharing, and quantization to reduce the number of model parameters and size, so that the model is suitable for drones. The image compression algorithm model performs lossless compression on the original inspection images, including the following steps: Using a transformation step to convert the data from the spatial domain to the frequency domain to discover redundant information; performing quantization operations on the data to reduce the data representation precision and thus compress the data; using methods such as Huffman coding to achieve efficient storage of the data. 2. Regarding the image decoding and restoration module The image decoding and restoration module is obtained through the following steps: 1) Model initialization: Set the neural network model and confirm the number of residual blocks. 2) Data preparation: Read in the compressed data and perform normalization. Select the image pair (x, t) as the training set to obtain the objective function: , Among them, x is the compressed image, t is the original high-resolution image with extremely low compression or compression ratio, f(w, b, x) is the prediction result of the convolutional neural network model, w and b are the network model parameters, and n is the number of pixels of the input image; 3) Model training: Use the random chunking method for training, use the stochastic gradient descent method to solve the parameters w and b of the neural network model, set the learning rate to 0.001, and in each iteration, calculate the gradient and update the model parameters; after 100 iterations, use larger-sized chunks to fine-tune the training results. The image decoding and restoration module decodes and restores the compressed inspection images, including the following steps: Adopt the block inference method, merge different block data during output, use the block segmentation method with overlapping regions, and during the block segmentation process, there is a certain pixel overlapping block inference scheme at the edges of adjacent blocks. Divide the original image into a series of blocks with overlapping regions. After inputting into the restoration network, splice the output block images at fixed positions. For the overlapping part, use the output data of the block closer in distance, or use the distance-weighted data of all blocks with this overlapping part; If using the block closer in distance, then: ; If using weighting, then: ; Among them, p is the pixel point in the overlapping region, 0 to n are the n image blocks with this overlapping region, center k is the center of the k-th image block, and dis(x, y) is the distance from pixel point x to pixel point y.
[0014] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the technical principles of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.
Claims
1. An image compression and restoration method for UAV power inspection, characterized in that, The method includes the following steps: a drone inspects the power grid and acquires the original inspection images of power facilities; An image compression algorithm model built in the drone performs lossless compression on the original inspection images to obtain compressed inspection images; the drone uploads the compressed inspection images to a server through a mobile network; a terminal device downloads the compressed inspection images from the server, and an image decoding and restoration module built in the terminal device decodes and restores the compressed inspection images to obtain decoded inspection images.
2. The method for image compression and restoration of drone power inspection according to claim 1, characterized in that The image compression algorithm model is light-weighted by combining pruning, weight sharing, and quantization methods to reduce the number of model parameters and the size.
3. The method for image compression and restoration of drone power inspection according to claim 1, wherein, The image compression algorithm model built in the drone performs lossless compression on the original inspection images, including the following steps: adopting a transformation step to convert data from the spatial domain to the frequency domain to discover redundant information; performing a quantization operation on the data to reduce the data representation precision and thus compress the data; using the Huffman coding method to achieve efficient storage of the data.
4. The method for image compression and restoration of drone power inspection according to claim 1, wherein The image decoding and restoration module is obtained through the following steps: 1) Model initialization: setting a neural network model and confirming the number of residual blocks; 2) Data preparation: reading in the compressed data and performing normalization; 3) Model training: using a random chunking method for training, using the stochastic gradient descent method to solve the parameters w and b of the neural network model, setting the learning rate to 0.001, and in each iteration, calculating the gradient and updating the model parameters.
5. The method for image compression and restoration of drone power inspection according to claim 4, characterized in that, The data preparation step includes: Selecting image pairs (x, t) as the training set to obtain the objective function: ， where x is the compressed image, t is the original high-resolution image with extremely low compression or compression ratio, f(w, b, x) is the prediction result of the convolutional neural network model, w and b are the network model parameters, and n is the number of pixels of the input image.
6. The method for image compression and restoration of drone power inspection according to claim 4, wherein In the model training step, after 100 iterations, the training result is fine-tuned using larger-sized chunks.
7. The method for image compression and restoration in the power inspection of an unmanned aerial vehicle according to claim 1, characterized in that, The image decoding and restoration module decodes and restores the compressed inspection images, including the following steps: adopting a block-based inference method, merging different block data during output, using a block segmentation method with overlapping regions, and in the process of segmenting the blocks, having a certain pixel overlapping block inference scheme at the edges of adjacent blocks, dividing the original image into a series of blocks with overlapping regions, after inputting into the restoration network, splicing the output block images at fixed positions, using the output data of the block closer in distance for the overlapping part, or using the distance-weighted data of all blocks with this overlapping part.
8. The method for image compression and restoration of drone power inspection according to claim 7, wherein: If the block closer in distance is used, then: ； If weighting is used, then: ； Among them, p is the pixel of the overlapping area, 0 to n are n image patches with this overlapping area, and center k is the center of the k-th image patch, and dis(x, y) is the distance from pixel point x to pixel point y.