Unmanned aerial vehicle inspection image vector retrieval method based on space-time fragmentation

Abstract

The invention relates to an unmanned aerial vehicle inspection image vector retrieval method based on space-time fragmentation, and the method comprises the steps: obtaining an unmanned aerial vehicle inspection image and image metadata, and carrying out the space-time fragmentation division of the unmanned aerial vehicle inspection image according to the image metadata, and obtaining the space-time fragmentation of the unmanned aerial vehicle inspection image; performing feature vector extraction on the unmanned aerial vehicle inspection image to obtain an image feature vector; mapping the space-time fragment of the unmanned aerial vehicle inspection image to a physical fragment of a vector database, and storing the image feature vector and the image metadata to the vector database according to a mapping relationship; and according to a retrieval condition, obtaining the physical fragment of the vector database, and carrying out collaborative retrieval in the target physical fragment according to the mixed index of the content and the spatio-temporal information to obtain a target image set. According to the method, the retrieval efficiency and accuracy of large-scale unmanned aerial vehicle inspection images can be remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of UAV image retrieval technology, and in particular to a UAV inspection image vector retrieval method based on spatiotemporal segmentation, which can be applied to the efficient retrieval and management of various types of inspection images, such as power facilities, transportation infrastructure, agricultural and forestry resources, and geological environment. Background Technology

[0002] With the widespread application of drone technology in fields such as power facilities, transportation infrastructure, agricultural and forestry resources, and geological exploration, the number of drone inspection images is growing exponentially, forming massive, high-dimensional image datasets. Existing large-scale drone inspection image retrieval methods face the following two core challenges:

[0003] Inefficient data storage and retrieval: Existing solutions typically store image spatiotemporal metadata (such as GPS coordinates and timestamps) and image feature vectors (used for content similarity search) separately in different systems (e.g., metadata stored in a relational database, feature vectors stored in a vector database). This necessitates complex cross-system joint queries and result merging when performing a complete retrieval query. This separate retrieval model not only significantly increases system coupling but also results in high retrieval latency and high resource consumption, making it difficult to meet the real-time or near-real-time response requirements of inspection operations.

[0004] Traditional vector retrieval methods, in large-scale data scenarios, typically rely on performing a full vector similarity search across the entire database. This method fails to effectively utilize the inherent spatiotemporal constraints of images (i.e., an image can only be associated with historical images from the same time and location). Ignoring the spatiotemporal context results in a massive computational burden for the full search, wasting computing resources. More importantly, it reduces the accuracy of retrieval results and increases the burden on maintenance personnel to sift through irrelevant results for relevant information. Summary of the Invention

[0005] The purpose of this invention is to provide a vector retrieval method for UAV inspection images based on spatiotemporal slicing. This method can effectively utilize the spatiotemporal constraint information of images to meet the needs of spatiotemporal correlation query and image comparison in inspection operations, and significantly improve the retrieval efficiency and accuracy of large-scale UAV inspection images.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] A method for retrieving UAV inspection image vectors based on spatiotemporal segmentation includes:

[0008] Acquire drone inspection images and image metadata, and perform spatiotemporal segmentation of the drone inspection images based on the image metadata to obtain spatiotemporal segments of the drone inspection images;

[0009] Feature vectors are extracted from the UAV inspection images to obtain image feature vectors.

[0010] The spatiotemporal segments of the UAV inspection images are mapped to physical segments of the vector database, and the image feature vectors and image metadata are stored in the vector database according to the mapping relationship;

[0011] Based on the search criteria, physical segments of the vector database are obtained, and collaborative retrieval is performed within the physical segments using a hybrid index of content and spatiotemporal information to obtain a set of target images.

[0012] Optionally, the spatiotemporal segmentation of the UAV inspection images includes:

[0013] The spatial units of the UAV inspection images are defined with the inspected target object as the center.

[0014] The time intervals for the drone inspection images are defined by inspection shifts or rolling time windows.

[0015] Based on an image set that conforms to the spatiotemporal range and image type, the spatiotemporal slices of the UAV inspection images are obtained.

[0016] Optionally, feature vector extraction of the UAV inspection images includes:

[0017] The UAV inspection images are preprocessed, and a pre-trained deep convolutional neural network model is used to extract feature vectors from the preprocessed images to obtain high-dimensional feature vectors.

[0018] The high-dimensional feature vector is reduced in dimensionality to obtain the reduced feature vector.

[0019] Optionally, storing the feature vector and the spatiotemporal segmented image in a vector database includes:

[0020] Using a deterministic encoding function or a spatiotemporal hash function, the metadata of the spatiotemporal segment is used as input parameters to calculate and generate the physical segment ID;

[0021] After generating the physical fragment ID, the image feature vector and image metadata are stored in the vector database.

[0022] Optionally, collaborative retrieval of the physical segments based on a hybrid index of content and spatiotemporal information includes: filtering the associated metadata fields in the retrieval conditions based on a structured index established for the image metadata to obtain a set of candidate images.

[0023] For the candidate image set, a vector similarity search is performed based on the vector index established on the high-dimensional feature vector to obtain the target image set.

[0024] Optionally, after obtaining the target image set, the process includes: performing feature similarity comparison on the target image set to achieve change detection or target comparison analysis.

[0025] Optionally, performing feature similarity comparison on the target image set includes:

[0026] The feature vectors and contents of the images in the target image set are compared with the feature vectors and contents of historical benchmark images. If the calculated similarity is lower than a preset similarity threshold, potential anomalies or changes are automatically marked.

[0027] The beneficial effects of this invention are as follows: This invention enables unified and efficient management of image feature vectors and spatiotemporal metadata, and provides a method for direct addressing and collaborative retrieval based on spatiotemporal constraints, thereby significantly improving the retrieval efficiency and accuracy of large-scale UAV inspection images. This invention proposes a method for unified storage of spatiotemporal metadata and image vectors, and for achieving direct hit retrieval through spatiotemporal sharding and functional database sharding mapping. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 This is a flowchart of an image vector retrieval method for UAV inspection based on spatiotemporal slicing, according to an embodiment of the present invention. Detailed Implementation

[0030] 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, and 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.

[0031] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0032] To address the problem that existing technologies often rely solely on image feature similarity for retrieval, neglecting temporal and spatial information, resulting in inaccurate and inefficient retrieval results, such as... Figure 1 As shown, this embodiment proposes a UAV inspection image vector retrieval method based on spatiotemporal segmentation, including:

[0033] Acquire drone inspection images and image metadata, and perform spatiotemporal segmentation of drone inspection images based on relevant image metadata to obtain spatiotemporal segments of drone inspection images;

[0034] Feature vectors are extracted from images inspected by drones to obtain image feature vectors;

[0035] The spatiotemporal slices of the UAV inspection images are mapped to the physical slices of the vector database, and the image feature vectors and image metadata are stored in the vector database according to the mapping relationship (the original images are stored in the object storage service, and the corresponding image file URLs are stored as metadata fields in the vector database).

[0036] Based on the search criteria, the physical segments of the vector database are hit, and a collaborative search is performed within the physical segments based on a hybrid index of content and spatiotemporal information to obtain the target image set.

[0037] In this embodiment, the metadata of the UAV inspection image is obtained, including but not limited to: high-precision GPS coordinates (longitude, latitude, altitude), shooting timestamp, target object identifier (such as device ID, tower number, road segment ID) and image type (such as visible light, infrared, multispectral).

[0038] Furthermore, the spatiotemporal segmentation of UAV inspection images includes:

[0039] Spatial units are defined in the image of the drone inspection, centered on the target object being inspected.

[0040] The time interval for drone inspection images is defined by inspection shift or rolling time window;

[0041] Based on an image set that conforms to the spatiotemporal range and image type, obtain spatiotemporal slices of UAV inspection images.

[0042] Specifically, based on the inspection tasks and object characteristics, refined spatiotemporal division rules are formulated. Spatial division: Taking the target object (such as tower, bridge, sign) as the center, spatial units such as circles, polygons or buffer zones are defined to ensure that images within the same unit have spatial proximity. Temporal division: According to the inspection frequency and data volume, time windows (such as shifts, rolling hour windows) are defined according to the inspection task requirements to ensure that images within the same time window have temporal continuity.

[0043] For an image set that conforms to a specific spatiotemporal range and image type, a unique spatiotemporal shard identifier (shard_id) is generated for logical grouping of the image set, and the precise spatial range and time interval represented by the shard are recorded.

[0044] Furthermore, feature vector extraction for drone inspection includes:

[0045] The images inspected by the UAV are preprocessed, and a pre-trained deep convolutional neural network model is used to extract feature vectors from the preprocessed images to obtain high-dimensional feature vectors.

[0046] The high-dimensional feature vector is reduced in dimensionality to obtain the reduced feature vector.

[0047] Specifically, image preprocessing involves standardizing the acquired images through denoising, Gaussian filtering, and size normalization. Simultaneously, specific enhancement processing is applied based on image type. For example, visible light images undergo texture enhancement (such as histogram equalization) to highlight structural features, while infrared images are subjected to adaptive threshold segmentation to highlight anomalous temperature regions. Specific processing can also be applied based on image type to emphasize target features.

[0048] Pre-trained deep convolutional neural network models (such as ResNet, VGG, etc.) are used to infer the pre-processed images and extract their fixed-length, high-dimensional feature vectors to describe the texture, shape, and structural information of the images.

[0049] By using dimensionality reduction techniques such as principal component analysis (PCA), high-dimensional feature vectors can be optimized and reduced in dimensionality (e.g., from 2048 dimensions to 512 dimensions), while maintaining feature discriminative power, storage overhead and retrieval computation complexity can be reduced.

[0050] Furthermore, storing the image feature vector and image metadata in a vector database includes:

[0051] Using a deterministic coding function or a spatiotemporal hash function, the image metadata corresponding to the spatiotemporal segment is taken as input parameters to calculate and generate the ID of the physical segment. After generating the physical segment ID, the image feature vector and image metadata are stored in a vector database. The metadata of the spatiotemporal segment includes the acquisition time, spatial location, target object identifier, and image type.

[0052] Specifically, the steps for building a vector database are as follows:

[0053] (1) Following the principle that the spatiotemporal partitioning is consistent with the physical structure of the database, each spatiotemporal partition is corresponding to an independent Collection or Partition (collectively referred to as physical partitioning) in a vector database (such as Milvus).

[0054] (2) Within each physical shard of the database, feature vectors and structured metadata are stored in a unified manner. The shard stores the feature vectors of the image (such as the FLOAT_VECTOR field) and the associated metadata.

[0055] (3) The system adopts a mechanism based on deterministic coding rules or spatiotemporal hash functions to define the mapping relationship between spatiotemporal shards and physical shards. This mechanism takes the core metadata of the spatiotemporal shards (such as target object ID and inspection shift code) as input parameters, and directly calculates and generates the corresponding physical shard ID through a fixed and repeatable function. Through this functional mapping, the system can convert spatiotemporal conditions into target database shard addresses without querying additional mapping tables, realizing direct hits on physical shards and fast addressing with O(1) complexity.

[0056] The mapping mechanism is as follows:

[0057] (3.1) Deterministic Encoding Function: This method encodes based on discrete, identifiable metadata.

[0058] ;

[0059] The input parameter (M) represents the core metadata of the spatiotemporal partitioning, specifically... This is represented as the target object identifier (such as tower ID, road segment ID). This refers to the inspection shift code or time window code. Image type (e.g., visible light, infrared); Output result (ID): ID of the corresponding physical slice in the vector database; Function ( ): Represents a deterministic encoding rule or function. After calculating the physical address, the feature vector and metadata are stored together.

[0060] (3.2) Spatio-temporal Hash Function: This method performs hashing based on continuous coordinate and time data.

[0061] ;

[0062] Among them, the output parameter (M) is: GPS coordinates ( ), collection timestamp ( ).

[0063] (4) To achieve efficient collaborative retrieval, a hybrid index is constructed for the data within each shard. A vector index (such as IVF_SQ8) is built for the high-dimensional feature vectors to quickly perform similarity searches. A structured index is built for metadata fields (such as time and target object ID) to quickly filter the image set based on search criteria. After calculating the physical address, the feature vectors and metadata are stored together.

[0064] Furthermore, collaborative retrieval of physical slices based on a hybrid index of content and spatiotemporal information includes: filtering by associating metadata fields with a structured index established on the image metadata to obtain a candidate image set;

[0065] For the candidate image set, a vector similarity search is performed based on the vector index established by the high-dimensional feature vector to obtain the target image set.

[0066] Specifically, the steps for vector retrieval based on spatiotemporal sharding and hybrid indexing are as follows:

[0067] (1) Determine the target fragment and achieve physical hit: Based on the target object ID, time interval, and image type information in the inspection task, the system calculates and determines the corresponding spatiotemporal fragment ID through the functional mapping mechanism constructed in step three. With the help of this mapping relationship, the system directly hits the corresponding Collection / Partition in the database, achieving physical fragment hit, thereby effectively avoiding vector scanning of the entire database.

[0068] (2) Collaborative retrieval and rapid filtering within a segment: Within the matched segments, the associated metadata fields are first rapidly filtered through the metadata index to quickly return a set of candidate images that meet the spatiotemporal constraints. Then, for this set of candidate images, a vector similarity search is performed to find the image that is most similar to the feature vector of the target image or a preset feature vector, and the final retrieval result is determined.

[0069] (3) Output results: The returned image set includes the slice ID, shooting time, target object information and its feature vector, etc.

[0070] Furthermore, after obtaining the target image set, the process includes: comparing the feature similarity of the target image set to achieve change detection or target comparison analysis.

[0071] Furthermore, feature similarity comparison of the target image set includes:

[0072] The feature vectors and contents of the images in the target image set are compared with the feature vectors and contents of historical benchmark images. If the calculated similarity is lower than the preset similarity threshold, potential anomalies or changes are automatically marked.

[0073] Specifically, based on the returned set of target images and their feature vectors, automatic or manual comparison and analysis are performed to achieve the application goals of UAV inspection images. This mainly includes the following two tasks:

[0074] Change detection includes feature similarity comparison: comparing the feature vector of the current inspection image with the feature vector of images within the same spatiotemporal segment (same target object, similar location) in history; anomaly labeling: if the calculated feature vector similarity is lower than a preset similarity threshold, the image or target area is automatically labeled as having a potential anomaly or change; application: mainly used to detect temporal changes such as missing power facility components, new cracks in the road surface, and changes in agricultural and forestry resource coverage.

[0075] Target comparison and anomaly tracing include content similarity comparison: comparing the content similarity between the target image (such as a fault sample input by the user) and historical images in the returned set; qualitative analysis: based on the comparison results, it is used to quickly identify anomalies, trace faults, or verify whether the current state is similar to a specific state (normal or abnormal) in the past; result output: outputting the most relevant inspection results at the final determination for operation and maintenance personnel to make decisions or further process.

[0076] Spatio-temporal segmentation is a key technology that combines temporal and spatial attributes to structure and organize data. This method divides data into shards with spatial proximity and temporal continuity by analyzing spatial location (e.g., GPS coordinates), timestamps (e.g., capture time), and object identifiers (e.g., device ID). In UAV inspection scenarios, spatio-temporal segmentation can classify images based on factors such as inspection flight segment, equipment unit, and time interval, ensuring that images from the same spatial region or adjacent time periods are grouped into the same shard. This significantly reduces the scope of irrelevant data retrieval, providing a foundation for efficient subsequent storage and retrieval.

[0077] This invention employs a unified storage and functional sharding mapping mechanism for vector databases. Milvus, a high-performance vector database system oriented towards vector retrieval, supports the simultaneous storage of image feature vectors and their structured metadata (such as timestamps, latitude and longitude, device type, etc.) within the same collection or partition, and provides hybrid indexing capabilities. This overcomes the drawbacks of traditional "separation of vectors and metadata," which leads to complex cross-system joint queries and low retrieval efficiency. Milvus achieves unified storage by defining a FLOAT_VECTOR field (for storing high-dimensional feature vectors) and multiple structured fields (such as VARCHAR, DATETIME, and INT types) within a single collection. The system can simultaneously perform structured filtering and vector search in the same query, significantly reducing latency and improving data consistency. Building upon this, this invention introduces a functional sharding mapping mechanism: based on the metadata information of spatiotemporal shards (such as inspection target ID and time range), the system maps image data from different spatiotemporal ranges to different Collections or Partitions in Milvus through a defined calculation function, achieving a deterministic correspondence between spatiotemporal shards and physical database shards. Through this computable mapping mechanism, when a retrieval request with spatiotemporal constraints is received, the system can directly calculate the target shard address and achieve a match, without needing to perform a vector search across the entire database. This strategy effectively reduces the computational scope and memory overhead, enabling fast retrieval and shard-level parallel processing under spatiotemporal constraints.

[0078] Image feature extraction is the process of representing image content in high dimension using deep convolutional neural networks (CNNs). Commonly used network models include ResNet, VGG, and EfficientNet, which can map input images into fixed-length, high-dimensional feature vectors to describe the image's texture, shape, lighting, and structural information. In UAV inspection scenarios, after image preprocessing (such as denoising, normalization, and contrast enhancement), feature vectors are extracted using pre-trained or transfer learning models, and dimensionality reduction optimization can be performed using methods such as principal component analysis (PCA). Finally, the extracted feature vectors and their corresponding spatiotemporal metadata are written into Milvus slices, achieving unified management of content features and spatiotemporal semantics, providing reliable data support for subsequent similarity search, change detection, target comparison, and anomaly identification.

[0079] Example 1: For power line inspection scenarios, a UAV inspection image vector retrieval method based on spatiotemporal segmentation includes:

[0080] Step 1: Spatiotemporal partitioning, the specific steps are as follows:

[0081] (1) Obtain metadata of UAV inspection images, including tower identification, patrol shift code, original shooting timestamp, equipment unit identification and image type (visible light / infrared).

[0082] (2) Spatial division: A circular spatial unit with a radius of 30 meters is defined with the center point of the tower as the center. Images falling in the same spatial unit are classified into the same spatiotemporal segment.

[0083] (3) Time division: Based on the inspection shift, for example, 9:00-12:00 in the morning and 14:00-17:00 in the afternoon, images taken in the same spatial unit that fall within the same shift are grouped into the same spatiotemporal segment.

[0084] (4) Generate a unique ID (shard_id) for each spatiotemporal slice and record the spatial range, time interval and image type.

[0085] Step 2: Feature vector extraction, the specific steps are as follows:

[0086] (1) Image preprocessing: noise reduction, Gaussian filtering, size normalization; In order to highlight the characteristics of power facilities: visible light images are histogram equalization to enhance the surface texture of the tower; infrared images are segmented using adaptive thresholding to highlight the abnormal temperature areas of the equipment.

[0087] (2) Use ResNet50 to extract 2048-dimensional feature vectors, and reduce the dimensionality to 512-dimensional using PCA.

[0088] Step 3: Vector database construction and functional mapping, the specific steps are as follows:

[0089] (1) A deterministic coding function is used to realize the functional mapping between spatiotemporal slicing and physical slicing to ensure direct addressing of physical slicing. Given that the core of power inspection data lies in the discrete and symbolic tower ID and inspection shift code, this function transforms the three core metadata items determined in step one—tower identifier, shift identifier, and image type—into a unique physical slicing ID, thereby achieving accurate mapping of physical slicing.

[0090] ;

[0091] Application example:

[0092] Input: Tower ID is The flight number is T001, the cruise number code is AM, and the image type is... VL.

[0093] Output: The physical partition ID is T100_AM_VL.

[0094] (2) The physical shards store feature vectors and metadata in a unified manner, and establish IVF_SQ8 vector index and structured index for similarity search and fast filtering.

[0095] Step 4: Vector retrieval based on spatiotemporal segmentation. The user inputs the target tower ID, time interval, and image type. The system calculates and directly hits the corresponding physical segment using a functional mapping mechanism. Within the hit segment, candidate images are first filtered using a structured index, and then vector similarity search is performed to return the image set and its feature vectors.

[0096] Step 5: Perform change detection based on the returned feature vectors. Compare the feature vector of the current image with the feature vector of a historical baseline image of the same tower structure. If the similarity is lower than a preset threshold, the tower structure or component is automatically marked as a potential anomaly or change. Finally, the image information is returned for manual or automated change detection and fault tracing analysis.

[0097] Example 2: Traffic facility inspection scenario, a UAV inspection image vector retrieval method based on spatiotemporal segmentation includes:

[0098] Step 1: Spatiotemporal partitioning, the specific steps are as follows:

[0099] (1) Obtain metadata of UAV inspection images, including GPS coordinates, shooting timestamp, inspection segment ID, equipment unit identifier and image type (visible light / infrared). Spatial division: delineate spatial polygons / buffers with roads, bridges or signs as units, and classify images falling in the same unit into the same spatiotemporal slice.

[0100] (2) Time division: The rolling window is used every hour or half hour to group images taken in the same spatial unit into the same spatiotemporal slice.

[0101] (3) Generate a unique ID (shard_id) for each spatiotemporal slice and record the spatial range, time interval and image type.

[0102] Step 2, Feature Vector Extraction, the specific steps are as follows:

[0103] (1) Image preprocessing: noise reduction, Gaussian filtering, size normalization; In order to highlight the features of traffic facilities: visible light images are textured to highlight road surface cracks or potholes; multispectral images can enhance vegetation cover or facility surface conditions.

[0104] (2) Use ResNet50 to extract 2048-dimensional feature vectors, and reduce the dimensionality to 512-dimensional using PCA.

[0105] Step 3, Vector Database Construction and Functional Mapping, the specific steps are as follows:

[0106] (1) For the continuous spatial coordinates and timestamps involved in traffic inspection data, a spatiotemporal hash function is adopted. The core of this function is quantization encoding, which transforms continuous latitude and longitude and timestamps into discrete hash buckets, and then connects them to achieve fragmentation hit and spatiotemporal proximity.

[0107] ;

[0108] Input Example: Input parameter location A: (Lat 30.220, Lon 120.100), time A: 11:35 AM, calculate Output ID = wtdz668_H11; Input parameter location B: (Lat30.221, Lon 120.101), time A: 11:45 AM, calculation Output ID = wtdz668_H11.

[0109] (2) The physical shards store feature vectors and metadata in a unified manner, and establish IVF_SQ8 vector index and structured index for similarity search and fast filtering.

[0110] Step 4: Based on the inspection segment ID, time interval, and other information, the system directly calculates and matches the corresponding physical segment ID using a functional mapping mechanism. Within the matched segment, candidate images are quickly filtered using a structured index, and then vector similarity search is performed to return the image set and its feature vectors.

[0111] Step 5: Change Detection and Comparison Analysis of Retrieval Results. Based on the returned feature vectors, change detection is performed. The feature vectors of the current road surface image are compared with the feature vectors of historical baseline images. If the similarity is lower than a preset threshold, it is automatically marked as a potential defect change such as road surface cracks or potholes, which is used for defect comparison and anomaly analysis of traffic facilities, effectively supporting subsequent maintenance decisions.

[0112] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for retrieving UAV inspection image vectors based on spatiotemporal segmentation, characterized in that, include: Acquire drone inspection images and image metadata, and perform spatiotemporal segmentation of the drone inspection images based on the image metadata to obtain spatiotemporal segments of the drone inspection images; Feature vectors are extracted from the UAV inspection images to obtain image feature vectors; The spatiotemporal slices of the UAV inspection images are mapped to physical slices of the vector database, and the image feature vectors and image metadata are stored in the vector database according to the mapping relationship; Based on the search criteria, physical segments of the vector database are obtained, and collaborative retrieval is performed within the physical segments using a hybrid index of content and spatiotemporal information to obtain a set of target images.

2. The UAV inspection image vector retrieval method based on spatiotemporal segmentation according to claim 1, characterized in that, Spatiotemporal segmentation of the UAV inspection images includes: The spatial units of the UAV inspection images are defined with the inspected target object as the center. The time intervals for the drone inspection images are defined by inspection shifts or rolling time windows. Based on an image set that conforms to the spatiotemporal range and image type, the spatiotemporal slices of the UAV inspection images are obtained.

3. The UAV inspection image vector retrieval method based on spatiotemporal segmentation according to claim 1, characterized in that, Feature vector extraction of the UAV inspection images includes: The UAV inspection images are preprocessed, and a pre-trained deep convolutional neural network model is used to extract feature vectors from the preprocessed images to obtain high-dimensional feature vectors. The high-dimensional feature vector is reduced in dimensionality to obtain the reduced feature vector.

4. The UAV inspection image vector retrieval method based on spatiotemporal segmentation according to claim 1, characterized in that, Storing the image feature vectors and image metadata in a vector database includes: Using a deterministic coding function or a spatiotemporal hash function, the image metadata corresponding to the spatiotemporal segment is used as input parameters to calculate and generate the physical segment ID; After generating the physical fragment ID, the image feature vector and image metadata are stored in the vector database.

5. The UAV inspection image vector retrieval method based on spatiotemporal segmentation according to claim 3, characterized in that, Collaborative retrieval of the physical slices based on a hybrid index of content and spatiotemporal information includes: filtering by associating metadata fields with a structured index established on the image metadata to obtain a candidate image set; For the candidate image set, a vector similarity search is performed based on the vector index established on the high-dimensional feature vector to obtain the target image set.

6. The UAV inspection image vector retrieval method based on spatiotemporal segmentation according to claim 1, characterized in that, After obtaining the target image set, the process includes: performing feature similarity comparison on the target image set to achieve change detection or target comparison analysis.

7. The UAV inspection image vector retrieval method based on spatiotemporal segmentation according to claim 1, characterized in that, The feature similarity comparison of the target image set includes: The feature vectors and contents of the images in the target image set are compared with the feature vectors and contents of historical benchmark images. If the calculated similarity is lower than a preset similarity threshold, potential anomalies or changes are automatically marked.

Images