Unmanned aerial vehicle intelligent inspection system based on AI vision and detection switch cabinet

Abstract

The invention discloses an unmanned aerial vehicle intelligent inspection system based on AI vision and a detection switch cabinet, and relates to the technical field of unmanned aerial vehicle intelligent inspection, the unmanned aerial vehicle intelligent inspection system comprises an unmanned aerial vehicle inspection platform, and the unmanned aerial vehicle inspection platform is in communication connection with the following modules: an unmanned aerial vehicle end, which is used for collecting and preprocessing video stream data of an inspection area; extracting a key frame from the preprocessed video stream data; and the AI visual analysis module is used for analyzing the extracted key frame by using an AI visual algorithm and identifying key information and abnormal fragments in the key frame. According to the invention, through the AI vision algorithm based on the convolutional neural network model, abnormal features can be automatically learned and identified, the abnormal types and specific conditions can be rapidly determined through deep analysis of the key frames, accurate positioning of abnormal segments and comparison with the preset abnormal feature database, compared with manual detection, the accuracy is greatly improved, and the detection efficiency is improved. Tiny abnormal changes can be found in time, and potential faults can be warned in advance.

Description

Technical Field

[0001] This invention relates to the field of drone intelligent inspection technology, specifically to a drone intelligent inspection system and inspection switch cabinet based on AI vision. Background Technology

[0002] With the rapid development of the power industry, the number of power facilities such as transmission lines and substations is constantly increasing, and the demand for inspection is also growing. Traditional manual inspection methods require a large amount of manpower, are slow, and are greatly limited by natural conditions such as terrain and weather, making it difficult to achieve high-frequency and large-scale inspections. In some dangerous or hard-to-reach areas, such as high-voltage transmission lines, bridges, and tunnels, manual inspection poses significant safety risks.

[0003] In the field of image communication, especially in UAV intelligent inspection systems, video transcoding is a crucial step in ensuring efficient transmission and storage of video data. However, existing technologies suffer from the following problems: 1. The limited computing resources of drones make it difficult to handle complex video decoding and re-encoding tasks, resulting in low transcoding efficiency; 2. Unoptimized video data transmission consumes a large amount of bandwidth, increasing communication costs; 3. Existing transcoding technology lacks intelligent analysis capabilities, making it difficult to dynamically adjust video quality or extract key information according to inspection needs; Therefore, the problem to be solved by this invention is how to partially decode the input video stream on the drone, extract key frames, and use AI vision algorithms to analyze the extracted key frames, identify key information and abnormal segments, and then adjust the video quality in the cloud and perform in-depth analysis of abnormal segments. To this end, an AI vision-based drone intelligent inspection system and inspection switch cabinet are proposed. Summary of the Invention

[0004] The purpose of this invention is to provide an AI vision-based drone intelligent inspection system and switch cabinet for testing, in order to solve the problems mentioned in the background art.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: Firstly, an AI vision-based intelligent drone inspection system includes a drone inspection platform, which is communicatively connected to the following modules: On the drone side, video stream data of the inspection area is collected and preprocessed, and key frames are extracted from the preprocessed video stream data. The AI ​​visual analysis module is used to analyze the extracted keyframes using AI visual algorithms, identify key information and abnormal segments in the keyframes, mark keyframes containing abnormal information, and then extract the marked keyframes. The cloud processing module is used to adjust the quality of the original video stream based on the results of AI visual analysis and to perform in-depth analysis of abnormal segments. The results integration and display engine is used to integrate the results of cloud analysis and display them to inspection personnel, presenting abnormal information, key information, and video quality adjustment information, so that inspection personnel can quickly understand the inspection results.

[0006] A further improvement to the technical solution of the present invention is that the UAV terminal includes a video acquisition and processing unit and a key frame extraction unit; The video acquisition and processing unit uses a high-definition camera mounted on the drone to collect raw video stream data of the inspection area in real time, and performs preprocessing operations including format conversion and frame rate adjustment to ensure that the input video data has a uniform format and a stable frame rate. The keyframe extraction unit employs a video decoding algorithm to partially decode the preprocessed video stream. By analyzing the differences between video frames, it extracts representative keyframes, summarizes the main content of the video, and reduces the amount of data that needs to be processed.

[0007] A further improvement to the technical solution of the present invention is that the video acquisition and processing unit specifically includes: The drone is equipped with a high-definition camera. After flying to the target area according to the preset inspection route, it collects raw video stream data of the inspection area in real time. Based on the system's preset standard format, the corresponding format conversion algorithm is called to decode and re-encode the original video stream data, converting various video formats into a specific format that can be processed by the system, ensuring that the converted video stream format is consistent. The frame rate of the converted video stream data is adjusted to a stable and appropriate value by using frame interpolation or frame extraction techniques, thus obtaining stable video stream data.

[0008] A further improvement to the technical solution of the present invention is that the keyframe extraction unit specifically includes: It receives preprocessed video stream data and uses a lightweight video decoding algorithm, which, unlike full decoding, only partially decodes the video stream. Inter-frame difference analysis is performed on a portion of the decoded video frame sequence. Using image analysis technology, the differences between adjacent video frames are compared frame by frame. By calculating multi-dimensional indicators such as pixel changes, color differences, and object movement, the differences between each pair of frames are quantified. Based on a preset difference threshold, adjacent video frames with significant changes are identified, effectively filtering redundant frames and reducing the amount of data. Based on the results of inter-frame difference analysis, adjacent video frames with differences exceeding the difference threshold are selected as keyframes. Keyframes can represent important scenes or changes in the video at different stages. Finally, the extracted keyframes are output, which greatly reduces the amount of data that needs to be processed later and improves the overall processing efficiency of the system.

[0009] A further improvement to the technical solution of the present invention is that the AI ​​visual analysis module specifically includes: The system receives keyframes output from the drone and performs preprocessing operations, including grayscale conversion and normalization. Grayscale conversion converts color images to grayscale, reducing data volume and computational complexity. Normalization ensures that image pixel values ​​are within a specific range, eliminating scale differences between different images. Simultaneously, filtering algorithms are used to remove noise interference and enhance image quality. The system also resizes and crops keyframes to meet the input requirements of AI vision algorithms, ensuring keyframe image quality, standardizing input data, and improving analysis accuracy and efficiency. The pre-trained AI vision algorithm model is loaded to perform in-depth analysis on the pre-processed keyframes. The AI ​​vision algorithm model is based on a convolutional neural network model as its basic architecture. It automatically extracts feature information from the keyframes and matches it with preset key information patterns and abnormal segment features. During the recognition process, it distinguishes between normal scenes and abnormal situations, identifies key information in the keyframes, and accurately locates abnormal segments. Then, it marks the keyframes containing abnormal information. Based on the results of AI visual analysis and combined with the labeling information, keyframes containing abnormal information are selected from all keyframes and organized into the required data format. Then, the extracted and organized labeled keyframes are output.

[0010] A further improvement to the technical solution of the present invention is that the cloud processing module includes a video quality adjustment unit and an abnormal segment analysis unit; The video quality adjustment unit adjusts the quality of the original video stream based on the keyframes and abnormal segment marker information extracted from the UAV terminal. The abnormal segment analysis unit is used to perform in-depth analysis on the abnormal segments marked by the AI ​​visual analysis module to determine the specific circumstances of the abnormality.

[0011] A further improvement to the technical solution of the present invention is that the video quality adjustment unit specifically includes: The system receives keyframes and corresponding abnormal segment markers extracted from the drone. It integrates and analyzes the received information to clarify the time position and correspondence of the keyframes in the original video stream. Based on the timestamps carried by the keyframes, it locates their time position in the original video stream and sorts out the correspondence between the keyframes and the original video frames. Through logical association and timestamp matching, it distinguishes between abnormal and non-abnormal segments in the original video stream. Based on the parsed tagging information, differentiated quality adjustments are made to abnormal and non-abnormal segments in the original video stream; After completing the quality adjustment operation, the adjusted video stream is verified. By checking the smoothness of video playback, the clarity of the picture, and the presentation of key information and abnormal segments, it is ensured that the quality adjustment achieves the expected goal. If the adjustment effect is not ideal, fine-tuning and optimization are performed. After verification, the adjusted video stream is output.

[0012] A further improvement to the technical solution of the present invention is that the abnormal fragment analysis unit specifically includes: After receiving the abnormal segments marked by the AI ​​visual analysis module, the abnormal segments are extracted from the original video stream and the relevant metadata, including timestamps, shooting locations, etc., are organized. Then, the abnormal segments are analyzed to observe their overall image features and determine the approximate type of the anomaly. Image processing is used to extract deep features from abnormal fragments, analyzing the texture and color changes of abnormal regions at the pixel level to extract key feature vectors, including shape, texture, and color. These key feature vectors are then compared with a pre-established abnormal feature database, which contains detailed feature information on various known abnormal situations. By comparing the abnormal fragments with known abnormal types, the abnormal fragments are quickly located. Weighted cosine similarity is used to calculate the abnormal similarity, and then the abnormal similarity is analyzed against a preset similarity threshold. When the abnormal similarity exceeds the similarity threshold, the specific situation of the abnormality is determined.

[0013] A further improvement to the technical solution of this invention lies in that: the result integration and display engine specifically includes: The system collects and analyzes the data from the cloud processing module, including detailed information on abnormal segments, key information, and video quality adjustments. It then integrates the data and organizes it according to a preset format and structure to ensure data integrity and consistency. The integrated results data are classified and processed, and abnormal information, key information and video quality adjustment are classified separately. For different types of inspection results, corresponding visualization display methods are designed, with abnormal information marked in red and key information highlighted in blue. The processed visualization results are presented to inspection personnel through a user interface. The display includes a detailed description of the anomaly information, a highlighting of key information, and specific details of video quality adjustments. Interactive functions are also provided, such as clicking on the anomaly information to view detailed video clips and adjusting the display ratio, allowing inspection personnel to thoroughly view and analyze the inspection results.

[0014] Secondly, the AI ​​vision-based drone intelligent inspection and detection switch cabinet includes at least one control processor and a memory for communicating with the at least one control processor; the memory stores instructions that can be executed by the at least one control processor, and the instructions are executed by the at least one control processor to enable the at least one control processor to execute the AI ​​vision-based drone intelligent inspection system.

[0015] Due to the adoption of the above technical solution, the technical progress achieved by this invention compared to the prior art is as follows: 1. This invention provides an AI vision-based drone intelligent inspection system and switch cabinet inspection system. Through AI vision algorithms based on convolutional neural network models, it can automatically learn and identify abnormal features. By performing in-depth analysis of key frames, it can accurately locate abnormal segments and compare them with a preset abnormal feature database to quickly determine the type and specific situation of the abnormality. Compared with manual inspection, its accuracy is greatly improved, and it can detect subtle abnormal changes in a timely manner and provide early warning of potential faults.

[0016] 2. This invention provides an AI vision-based drone intelligent inspection system and switch cabinet. By preprocessing the video stream and extracting key frames at the drone end, the cloud performs differentiated quality adjustments on the original video stream based on the AI ​​analysis results. Non-abnormal segments have their quality reduced to save storage space and transmission bandwidth, while abnormal segments maintain or improve their quality to ensure analysis quality. This optimizes the storage and transmission efficiency of video data, reduces data redundancy, lowers storage and communication costs, and ensures the clarity and detail of abnormal segments. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0018] Figure 1 This is a schematic diagram of the workflow of the present invention; Figure 2 This is a schematic diagram illustrating the workflow of the AI ​​visual analysis module of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1, as Figure 1 , Figure 2 As shown, this invention provides an AI vision-based intelligent drone inspection system, including a drone inspection platform. The drone inspection platform has the following communication modules, wherein: The drone terminal is used to collect and preprocess video stream data of the inspection area and extract key frames from the preprocessed video stream data. The drone terminal includes a video acquisition and processing unit and a key frame extraction unit. The video acquisition and processing unit utilizes a high-definition camera mounted on a drone to collect raw video stream data of the inspection area in real time. It performs preprocessing operations, including format conversion and frame rate adjustment, to ensure that the input video data has a uniform format and stable frame rate. The drone, equipped with a high-definition camera, flies to the target area along a pre-set inspection route and collects raw video stream data of the inspection area in real time. The camera continuously records every detail of the inspection scene with a specific scanning frequency and angle, converting light signals into electrical signals to form a continuous raw video stream. Based on the system's pre-set standard format, the corresponding format conversion algorithm is called to decode and re-encode the raw video stream data, uniformly converting various video formats into a specific format that can be processed by the system, ensuring that the converted video stream format is consistent. The frame rate of the converted video stream data is adjusted to a stable and appropriate value using frame interpolation or frame extraction techniques, obtaining stable video stream data. The video stream data after frame rate adjustment is not only smoother to play but also reduces data redundancy, lowering the pressure on subsequent transmission and storage. The keyframe extraction unit employs a video decoding algorithm that partially decodes the preprocessed video stream. By analyzing the differences between video frames, it extracts representative keyframes to summarize the main content of the video, reducing the amount of data that needs to be processed. It receives the preprocessed video stream data and uses a lightweight video decoding algorithm, unlike full decoding, to partially decode the video stream and extract only key information. This avoids the complex and time-consuming full decoding operation of the entire video stream, enabling it to quickly obtain the basic structure and content information of video frames with limited UAV computing resources. It performs inter-frame difference analysis on the partially decoded video frame sequence, using image analysis technology to compare the differences between adjacent video frames frame by frame. By calculating multi-dimensional indicators such as pixel changes, color differences, and object movement, it quantifies the differences between each pair of frames and, based on a preset difference threshold, identifies adjacent video frames with significant changes, effectively filtering redundant frames and reducing data volume. Based on the results of the inter-frame difference analysis, it selects adjacent video frames with differences exceeding the difference threshold as keyframes. Keyframes represent important scenes or changes in the video at different stages. Finally, the extracted keyframes are output, significantly reducing the amount of data that needs to be processed subsequently and improving the overall processing efficiency of the system. The AI ​​vision analysis module utilizes AI vision algorithms to analyze extracted keyframes, identify key information and abnormal segments within them, and mark keyframes containing abnormal information. It then extracts these marked keyframes. The module receives keyframes output from the drone and performs preprocessing operations, including grayscale conversion and normalization. Grayscale conversion converts color images to grayscale, reducing data volume and computational complexity. Normalization ensures pixel values ​​are within a specific range, eliminating scale differences between images. Simultaneously, filtering algorithms remove noise interference, enhancing image quality. Keyframes are resized and cropped to meet the input requirements of the AI ​​vision algorithm, ensuring image quality, standardizing input data, and improving analysis accuracy and efficiency. Finally, a pre-trained AI vision algorithm model is loaded to perform deep analysis on the preprocessed keyframes. This AI vision algorithm model is based on a convolutional neural network model and automatically extracts feature information from the keyframes. The system collects and matches information with preset key information patterns and abnormal segment features. During the recognition process, it distinguishes between normal scenes and abnormal situations, identifies key information in key frames, and accurately locates abnormal segments. Then, it marks key frames containing abnormal information. The training process of the AI ​​vision algorithm model is as follows: collect a large amount of image data related to the inspection task, covering various normal and abnormal scenes to ensure the diversity and representativeness of the data, including different lighting conditions, weather conditions, shooting angles, etc. Use the LabelImg annotation tool to annotate the collected image data to clarify whether there are abnormalities in the image, as well as the specific location and type of the abnormality. Then, perform preprocessing operations on the image data, including grayscale, normalization, and size adjustment, to make it meet the requirements of the model input. Select a suitable convolutional neural network (CNN) architecture as the basic model, define the input layer, convolutional layer, pooling layer and fully connected layer of the network, construct a complete network structure, and add regularization technology to the network to prevent the model from overfitting.The preprocessed data is divided into training and validation sets. The model is trained using the training set and evaluated and tuned using the validation set. A cross-entropy loss function and the Adam optimizer are selected, and reasonable learning rates and training epochs are set. During training, the model's progress and performance metrics, including accuracy, recall, and F1 score, are monitored in real time. Training parameters are adjusted as needed. The convolutional neural network model is evaluated using a test set to test its performance on unseen data. A comprehensive performance assessment is then conducted, and the model's confusion matrix is ​​analyzed to understand its performance across different categories and identify any shortcomings. Based on the evaluation results, the model is optimized. Optimization methods include adjusting the network structure, modifying hyperparameters, and increasing the amount of data. Model pruning and quantization are performed to reduce the computational and storage requirements of the model and improve its running efficiency. Transfer learning technology is used to apply the pre-trained model to similar tasks to further improve the model's performance. The optimized model is deployed to the drone to ensure that the model can run efficiently in actual inspection tasks. Based on the results of AI visual analysis and combined with the labeling information, key frames containing abnormal information are selected from all key frames and organized into the required data format. Then, the extracted and organized labeled key frames are output. The cloud processing module is used to adjust the quality of the original video stream based on the results of AI visual analysis and to perform in-depth analysis of abnormal segments. The results integration and display engine is used to integrate the results of cloud analysis and display them to inspection personnel, presenting abnormal information, key information, and video quality adjustment information, so that inspection personnel can quickly understand the inspection results.

[0021] Example 2, as Figure 1 , Figure 2 As shown, based on Embodiment 1, the present invention provides a technical solution: preferably, the cloud processing module includes a video quality adjustment unit and an abnormal segment analysis unit; The video quality adjustment unit adjusts the quality of the original video stream based on keyframes and abnormal segment marker information extracted from the UAV. For non-abnormal segments, it reduces video quality to save storage space and transmission bandwidth; for abnormal segments, it maintains or improves video quality, achieving dynamic adjustment of video quality, optimizing the storage and transmission efficiency of video data, and ensuring the analysis quality of abnormal segments. It receives keyframes extracted from the UAV and corresponding abnormal segment marker information, integrates and analyzes the received information, clarifies the temporal position and correspondence of keyframes in the original video stream, locates their temporal position in the original video stream based on the timestamps carried by the keyframes, and establishes the correspondence between keyframes and original video frames. Through logical association and timestamp matching, it divides the original video stream into segments... Based on the parsed tagging information, differentiated quality adjustments are made to abnormal and non-abnormal segments in the original video stream. For non-abnormal segments, video quality is reduced by lowering resolution, frame rate, and adjusting encoding parameters, effectively reducing data volume and saving storage space and transmission bandwidth. For abnormal segments, their original quality is maintained or their resolution and color saturation are further increased to ensure that the details of the abnormal segments are clearly visible. After the quality adjustment is completed, the adjusted video stream is verified by checking the playback smoothness, image clarity, and the presentation of key information and abnormal segments to ensure that the quality adjustment achieves the expected goals. If the adjustment effect is not ideal, fine-tuning and optimization are performed. After verification, the adjusted video stream is output. The abnormal segment analysis unit is used to perform in-depth analysis on abnormal segments marked by the AI ​​visual analysis module to determine the specific situation of the anomaly. After receiving the abnormal segments marked by the AI ​​visual analysis module, it extracts the abnormal segments from the original video stream and organizes relevant metadata, including timestamps, shooting locations, etc. Then, it analyzes the abnormal segments, observes their overall image features, judges the approximate type of anomaly, and uses image processing to perform in-depth feature extraction on the abnormal segments. It analyzes the texture and color changes of the abnormal area at the pixel level and extracts key feature vectors, including shape, texture, and color. The key feature vectors are compared with a pre-established abnormal feature database, which contains detailed feature information of various known anomalies. By comparing, the abnormal segments are quickly located with known anomaly types. Weighted cosine similarity is used to calculate the anomaly similarity, and then the anomaly similarity is analyzed with a preset similarity threshold. When the anomaly similarity exceeds the similarity threshold, the specific situation of the anomaly is determined. The formula for calculating anomaly similarity is as follows: ; ; ; In the formula, The anomaly similarity score represents the similarity between the anomaly fragment and known anomaly types. The abnormal similarity is given by the key feature vector extracted from the abnormal fragments, which is represented as: , (i=1,2,…,n) represents the anomalous segment at the _th ... Feature values ​​on each feature dimension, and a known anomaly type in a pre-built anomaly feature database. The key feature vector is represented as =, (i=1,2,…,n) represents the known anomaly type in the nth position. Feature values ​​in each feature dimension Feature vector of abnormal fragment The weighted average is used to center the feature vector. For known exception types Feature vector The weighted average is used to center the feature vector. For the first The weights of each feature dimension, when When, it indicates that the abnormal fragment is different from the known abnormal type. The feature vectors of the two are perfectly positively correlated, meaning their trends of change across all feature dimensions are completely consistent. In this case, the abnormal segment can be considered completely similar to the known abnormal type. The value gradually increases from 0 to 1, indicating that the abnormal fragment is related to a known abnormal type. The similarity is getting higher and higher, when When the similarity exceeds a preset threshold, the abnormal segment is considered to belong to the known abnormal type or to be highly similar to the known abnormal type. The results integration and display engine specifically includes: The system collects and analyzes data from the cloud processing module, including detailed information on abnormal segments, key information, and video quality adjustments. This data is then systematically integrated and organized according to a pre-defined format and structure to ensure data integrity and consistency. The integrated data is categorized, with abnormal information, key information, and video quality adjustments grouped separately. Corresponding visualization methods are designed for different types of inspection results, such as highlighting abnormal information in red and key information in blue. The processed visualizations are presented to inspection personnel through a user interface, displaying detailed descriptions of abnormal information, highlighted key information, and specific details of video quality adjustments. Interactive functions are also provided, such as clicking on abnormal information to view detailed video segments and adjusting the display scale, allowing inspection personnel to thoroughly review and analyze the inspection results.

[0022] Example 3, as Figure 1 , Figure 2 As shown, based on Embodiments 1-2, the present invention also provides an AI vision-based drone intelligent inspection and detection switch cabinet, including at least one control processor and a memory for communicating with the at least one control processor; the memory stores instructions that can be executed by the at least one control processor, and the instructions are executed by the at least one control processor to enable the at least one control processor to execute the AI ​​vision-based drone intelligent inspection system.

[0023] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An AI vision-based drone intelligent inspection system, including a drone inspection platform, characterized in that, The drone inspection platform has the following communication modules, among which: On the drone side, video stream data of the inspection area is collected and preprocessed, and key frames are extracted from the preprocessed video stream data. The AI ​​visual analysis module is used to analyze the extracted keyframes using AI visual algorithms, identify key information and abnormal segments in the keyframes, mark keyframes containing abnormal information, and then extract the marked keyframes. The cloud processing module is used to adjust the quality of the original video stream based on the results of AI visual analysis and to perform in-depth analysis of abnormal segments. The results integration and display engine is used to integrate the results of cloud analysis and display them to inspection personnel.

2. The AI ​​vision-based unmanned aerial vehicle (UAV) intelligent inspection system according to claim 1, characterized in that: The drone terminal includes a video acquisition and processing unit and a keyframe extraction unit; The video acquisition and processing unit uses a high-definition camera mounted on the drone to collect raw video stream data of the inspection area in real time and performs preprocessing operations including format conversion and frame rate adjustment. The keyframe extraction unit uses a video decoding algorithm to partially decode the preprocessed video stream, extracting representative keyframes to summarize the main content of the video.

3. The AI ​​vision-based unmanned aerial vehicle (UAV) intelligent inspection system according to claim 2, characterized in that: The video acquisition and processing unit specifically includes: The drone is equipped with a high-definition camera. After flying to the target area according to the preset inspection route, it collects raw video stream data of the inspection area in real time. Based on the system's preset standard format, the corresponding format conversion algorithm is called to decode and re-encode the original video stream data, converting various video formats into a specific format that can be processed by the system. To obtain stable video stream data, frame interpolation or frame extraction techniques are used to assess the frame rate of the converted video stream data.

4. The AI ​​vision-based unmanned aerial vehicle (UAV) intelligent inspection system according to claim 2, characterized in that: The keyframe extraction unit specifically includes: It receives preprocessed video stream data and uses a lightweight video decoding algorithm, which, unlike full decoding, only partially decodes the video stream. Inter-frame difference analysis is performed on a portion of the decoded video frame sequence. Using image analysis technology, the differences between adjacent video frames are compared frame by frame. By calculating multi-dimensional indicators such as pixel changes, color differences, and object movement, the differences between each pair of frames are quantified. Based on a preset difference threshold, adjacent video frames with significant changes are identified. Based on the results of inter-frame difference analysis, adjacent video frames with differences exceeding the difference threshold are selected as keyframes, and finally, the extracted keyframes are output.

5. The AI ​​vision-based unmanned aerial vehicle (UAV) intelligent inspection system according to claim 2, characterized in that: The AI ​​visual analysis module specifically includes: The system receives keyframes output from the drone and performs preprocessing operations, including grayscale conversion and normalization of the keyframes. At the same time, it uses filtering algorithms to remove noise interference, and adjusts and crops the keyframes to meet the input requirements of AI vision algorithms. The pre-trained AI vision algorithm model is loaded to perform in-depth analysis on the pre-processed keyframes. The AI ​​vision algorithm model is based on a convolutional neural network model as its basic architecture. It automatically extracts feature information from the keyframes and matches it with preset key information patterns and abnormal segment features. During the recognition process, it distinguishes between normal scenes and abnormal situations, identifies key information in the keyframes, and accurately locates abnormal segments. Then, it marks the keyframes containing abnormal information. Based on the results of AI visual analysis and combined with the labeling information, keyframes containing abnormal information are selected from all keyframes and organized into the required data format. Then, the extracted and organized labeled keyframes are output.

6. The AI ​​vision-based drone intelligent inspection system according to claim 5, characterized in that: The cloud processing module includes a video quality adjustment unit and an abnormal segment analysis unit; The video quality adjustment unit adjusts the quality of the original video stream based on the keyframes and abnormal segment marker information extracted from the UAV terminal. The abnormal segment analysis unit is used to perform in-depth analysis on the abnormal segments marked by the AI ​​visual analysis module to determine the specific circumstances of the abnormality.

7. The AI ​​vision-based unmanned aerial vehicle (UAV) intelligent inspection system according to claim 6, characterized in that: The video quality adjustment unit specifically includes: The system receives keyframes and corresponding abnormal segment markers extracted from the drone. It integrates and analyzes the received information to clarify the time position and correspondence of the keyframes in the original video stream. Based on the timestamps carried by the keyframes, it locates their time position in the original video stream and sorts out the correspondence between the keyframes and the original video frames. Through logical association and timestamp matching, it distinguishes between abnormal and non-abnormal segments in the original video stream. Based on the parsed tagging information, differentiated quality adjustments are made to abnormal and non-abnormal segments in the original video stream; After completing the quality adjustment operation, the adjusted video stream is verified by checking the smoothness of video playback, the clarity of the picture, and the presentation of key information and abnormal segments. If the adjustment effect is not ideal, fine-tuning and optimization are performed. After the verification is passed, the adjusted video stream is output.

8. The AI ​​vision-based unmanned aerial vehicle (UAV) intelligent inspection system according to claim 7, characterized in that: The abnormal fragment analysis unit specifically includes: After receiving the abnormal segments marked by the AI ​​visual analysis module, the abnormal segments are extracted from the original video stream, and the relevant metadata is organized. Then, the abnormal segments are analyzed to observe their overall image features and determine the approximate type of the abnormality. Image processing is used to extract deep features from abnormal fragments, analyzing the texture and color changes of abnormal regions at the pixel level to extract key feature vectors, including shape, texture, and color. These key feature vectors are then compared with a pre-established abnormal feature database, which contains detailed feature information on various known abnormal situations. By comparing the abnormal fragments with known abnormal types, the abnormal fragments are quickly located. Weighted cosine similarity is used to calculate the abnormal similarity, and then the abnormal similarity is analyzed against a preset similarity threshold. When the abnormal similarity exceeds the similarity threshold, the specific situation of the abnormality is determined.

9. The AI ​​vision-based unmanned aerial vehicle (UAV) intelligent inspection system according to claim 8, characterized in that: The results integration and display engine specifically includes: The system collects and analyzes the data from the cloud processing module, including detailed information on abnormal segments, key information, and video quality adjustments. It then integrates the data and organizes it according to a preset format and structure. The integrated results data are classified and processed, and abnormal information, key information and video quality adjustment are classified separately. Corresponding visualization display methods are designed for different types of inspection results. The processed visualization results are displayed to the inspection personnel through the user interface. The display includes a detailed description of the abnormal information, a highlighting of key information, and specific details of the video quality adjustments. At the same time, interactive functions are provided so that the inspection personnel can view and analyze the inspection results in depth.

10. A drone-based intelligent inspection and testing switch cabinet based on AI vision, characterized in that: It includes at least one control processor and a memory for communicatively connecting to the at least one control processor; The memory stores instructions that can be executed by the at least one control processor, which, when executed by the at least one control processor, enables the at least one control processor to execute the AI ​​vision-based drone intelligent inspection system as described in any one of claims 1 to 9.

Images