A real scene photo instance segmentation classification verification method

Abstract

This invention provides a method for segmenting and classifying real-world photo instances, relating to the field of computer vision technology. The method includes: constructing a real-world verification dataset based on the spatial distribution of the vector data to be verified; establishing an interpretation feature library containing texture and geometric features; constructing a cascaded network architecture consisting of an object detection subnetwork and a segmentation subnetwork, using bounding boxes to guide segmentation operations, freezing the segmentation encoder parameters during training, embedding interpretation feature data into the network backbone to obtain a lightweight cascaded instance segmentation model; projecting the vector data onto the image coordinate system using a spatial mapping transformation algorithm to obtain a theoretical projection area; comparing the spatial overlap and attribute consistency between the model segmentation results and the theoretical projection area to generate classification verification correction information. This invention solves the problem of the lack of an effective correlation verification mechanism between heterogeneous data, achieving automation and efficiency in ground feature verification.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a method for segmenting, classifying, and verifying real-world photograph instances. Background Technology

[0002] In the fields of natural resource surveys and geographic national condition monitoring, the verification of nearshore surface vector data has long relied on manual field inspections or manual visual interpretation based on high-resolution remote sensing imagery. With the rapid development of deep learning technology, the industry has begun to try to introduce computer vision technologies such as convolutional neural networks to perform semantic segmentation or target detection on collected real-world photos in an attempt to achieve automated identification of ground features such as roads, buildings, and water bodies, replacing traditional manual verification methods.

[0003] To further improve recognition accuracy in complex scenes, current technological trends are evolving towards large models and combined algorithms. A mainstream research focus is combining efficient object detection algorithms with large segmentation models possessing strong generalization capabilities. For example, using the YOLO series of algorithms for object localization, and then combining it with the SAM segmentation model for pixel-level fine segmentation. This combination of detection and segmentation aims to leverage the feature understanding capabilities of large models under zero-sample or few-sample conditions to address the insufficient generalization ability of traditional small models when facing objects with significant appearance differences or irregular shapes.

[0004] However, applying these technologies to vector verification of large-scale real-world photographs still faces significant challenges. First, there is the conflict between computational resources and efficiency. Taking YOLO version 10 combined with the SAM model as an example, the latter has as many as 600 million parameters, resulting in extremely high hardware memory requirements and long processing times for model training and inference, making it difficult to meet the timeliness requirements of batch processing massive amounts of data. Second, there is insufficient ability to distinguish fine-grained features. From the perspective of near-shore roads, it is difficult to distinguish similar features such as woodlands and orchards, and phenomena such as roadside trees obscuring ponds frequently occur, making conventional models prone to missed detections or misjudgments. Finally, there is a gap in spatial matching of heterogeneous data. Existing instance segmentation results are based on image pixel coordinates, while the vector data to be verified is based on geospatial coordinates. Current technologies lack an automated semantic space mapping mechanism to accurately compare the attribute consistency between the two. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a real-scene photo instance segmentation and classification verification method. This invention solves the problems of existing technologies, such as excessive model parameters leading to high computational costs and low efficiency, difficulty in extracting fine-grained features of ground objects in complex scenes leading to poor recognition accuracy, and lack of an effective correlation verification mechanism between instance segmentation results and vector data spatial attributes.

[0006] To achieve the above objectives, the present invention provides the following solution: A method for segmenting and classifying real-world photo instances, comprising: Based on the spatial distribution information of the vector data to be verified, obtain the corresponding real-world image; The real-scene images are cleaned for effectiveness, and images that meet the imaging scale of ground features are selected according to preset spatial constraints to obtain a real-scene verification dataset containing geographic location information. Extract key identification features of preset land cover categories in the real scene image, and store the key identification features in the interpretation mark feature library; Build a cascaded network architecture; The signal transmission path of the cascaded network architecture is determined based on preset transmission rules; Based on the signal transmission path, the cascaded network architecture is trained using the real-world verification dataset, and key identification features from the interpretation marker feature library are embedded into the cascaded network architecture to obtain a trained lightweight cascaded instance segmentation model. The vector data to be verified is projected onto the pixel coordinate system of the real scene image using a spatial mapping transformation algorithm to obtain the theoretical projection area; The segmentation results output by the lightweight cascaded instance segmentation model are compared with the theoretical projection region to calculate spatial overlap and attribute consistency; if the overlap and attribute consistency do not meet the preset requirements, classification verification correction information is generated.

[0007] The present invention discloses the following technical effects: This invention provides a method for segmenting and classifying real-scene photos for verification. By constructing a lightweight cascaded network architecture, this invention utilizes bounding boxes to guide segmentation and freeze the encoder parameters of large models, significantly reducing the memory usage and time cost of model training and inference, effectively solving the problems of high computational overhead and low efficiency of existing large model combinations. By constructing a multi-scale interpretation feature library and embedding it into the network backbone, this invention enhances the model's ability to extract fine-grained features of highly similar land features such as woodlands and orchards, as well as occluded targets, overcoming the deficiency of poor recognition accuracy in complex scenes. Furthermore, by using spatial projection transformation to establish a mapping relationship between vector data and real-scene images, and through dual comparison of spatial overlap and attribute consistency, this invention achieves automated closed-loop verification between heterogeneous data, solving the problem of false alarms and misreports caused by the lack of effective association mechanisms in traditional methods, and greatly improving operational efficiency. Attached Figure Description

[0008] 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.

[0009] Figure 1 This is a flowchart of a real-scene photo instance segmentation and classification verification method provided in an embodiment of the present invention. Detailed Implementation

[0010] 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.

[0011] 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.

[0012] like Figure 1 As shown, the present invention provides a method for segmenting and classifying real-world photo instances, including: Step 100: Obtain the corresponding real-world image based on the spatial distribution information of the vector data to be verified; Step 200: Perform effective cleaning on the real-scene images and filter out images that meet the imaging scale of ground features according to preset spatial constraints to obtain a real-scene verification dataset containing geographic location information; Step 300: Extract the key identification features of the preset land cover categories in the real scene image, and store the key identification features in the interpretation mark feature library; Step 400: Construct a cascaded network architecture; Step 500: Determine the signal transmission path of the cascaded network architecture based on preset transmission rules; Step 600: Based on the signal transmission path, train the cascaded network architecture using the real-world verification dataset, and embed the key identification features in the interpretation marker feature library into the cascaded network architecture to obtain a trained lightweight cascaded instance segmentation model; Step 700: Project the vector data to be verified onto the pixel coordinate system of the real scene image using a spatial mapping transformation algorithm to obtain the theoretical projection area; Step 800: Compare the segmentation result output by the lightweight cascaded instance segmentation model with the theoretical projection region to calculate the spatial overlap and attribute consistency; if the overlap and attribute consistency do not meet the preset requirements, generate classification verification correction information.

[0013] Furthermore, the specific implementation process of step 100 is as follows: This embodiment first analyzes the geometric attributes of the vector data to be verified, reads the geometric coordinate sequence that constitutes the outline of the ground features, and uses this geometric coordinate sequence as a central reference to expand the spatial morphology outward according to a preset distance threshold, thereby constructing a spatial retrieval buffer covering the vicinity of near-shore roads. Based on the geographical range of this spatial retrieval buffer, this embodiment discretizes it and generates multiple image acquisition tasks containing precise latitude and longitude parameters and shooting angle requirements. These image acquisition tasks are then mapped to a preset distributed task scheduling queue according to priority or spatial proximity to achieve load balancing and orderly management of acquisition requests.

[0014] To address the issue of network platforms restricting high-frequency access from a single source, this embodiment pre-builds and maintains an account pool containing multiple different network access credentials, and monitors the activity status and remaining access quota of each credential in the account pool in real time. When executing tasks in the scheduling queue, this embodiment uses a round-robin or random weighted algorithm to dynamically retrieve currently available network access accounts from the account pool, and distributes the image acquisition tasks to be executed to different logical nodes. By using a multi-account concurrency strategy to distribute request pressure, it effectively avoids the risk of server access denial due to short-term high-concurrency requests.

[0015] This embodiment further establishes a long-term communication channel with the target network platform, utilizes multi-threading technology to concurrently execute the aforementioned allocated image acquisition tasks, and reduces network handshake overhead and improves data throughput efficiency by reusing the underlying transmission control protocol connection. This embodiment sends data requests carrying geographic location parameters to the network platform through this long-term connection channel and receives the returned data stream in real time. After verifying the integrity of the data packets, the acquired real-scene images are stored in local storage media according to geospatial indexing rules, thereby completing the automated acquisition of real-scene data for a specific vector area.

[0016] Furthermore, the specific implementation process of step 200 is as follows: This embodiment first performs pixel-level quality inspection on the acquired real-world images, converting the image data into grayscale space and calculating the statistical characteristics of the pixel grayscale distribution, such as the average grayscale value or grayscale variance. Considering the possibility of data corruption or completely black images during network download, this embodiment pre-defines a criterion for judging non-completely black images. An image is considered to contain valid visual information only if its pixel grayscale distribution statistical characteristics exceed this criterion. This embodiment marks such images as valid candidate images and removes invalid data that fails the inspection, thereby cleaning up low-quality or meaningless samples during the data preprocessing stage.

[0017] This embodiment reads the metadata information attached to valid candidate images and parses out the latitude and longitude coordinates of the recorded shooting time, using this as the camera's viewpoint position. Simultaneously, this embodiment acquires the spatial geometric information of the vector data to be verified and determines the geometric center coordinates of the vector ground object through geometric calculations. To quantify the spatial relationship between the observation point and the target, this embodiment projects the shooting position coordinates and the geometric center coordinates of the vector data onto the same plane coordinate system and calculates the spatial Euclidean distance between them. This distance directly reflects the physical distance between the camera and the ground object.

[0018] This embodiment establishes an inverse relationship model between spatial Euclidean distance and ground feature imaging scale based on optical imaging principles. That is, the farther the shooting distance, the fewer pixels the ground feature occupies in the image, and the more blurred its features become. To ensure that subsequent algorithms can extract clear texture and contour features, this embodiment sets a preset sharpness threshold, which represents the maximum shooting distance that meets the minimum resolution requirement for feature extraction. This embodiment filters out valid candidate images whose spatial Euclidean distance is less than the preset sharpness threshold, and establishes a one-to-one index association between these images and their corresponding shooting location coordinates. Finally, this is stored as a real-scene verification dataset containing geographic location information, ensuring that the data used for verification possesses both image quality validity and the visibility of ground feature details.

[0019] Furthermore, the specific implementation process of step 300 is as follows: This embodiment first introduces semantic saliency detection technology from computer vision to preprocess the acquired real-world image. By analyzing the color contrast, spatial distance, and global distribution characteristics of image pixels, it automatically locates the region in the image with the highest visual saliency that semantically matches the preset land cover category. Based on the mask boundary generated by saliency detection, this embodiment crops an image patch containing a single land cover target from the original complex real-world background and defines it as the target region of interest. This shields the subsequent feature extraction process from interference from roads, sky, or other irrelevant background features, ensuring the purity of feature analysis.

[0020] For different types of ground features, this embodiment employs differentiated feature calculation strategies. When the identified region of interest (ROI) belongs to the vegetation category, this embodiment uses a frequency domain analysis algorithm to calculate the periodicity parameters of the image texture to distinguish between the regularity of rows and columns in artificial orchards and the random distribution of natural forests. Simultaneously, it statistically analyzes the distribution parameters of the color space to capture the color characteristics of fruits or specific tree species, thereby obtaining biological texture features. When the ROI belongs to the artificial structure category, this embodiment uses the image moment principle to calculate the geometric moments of the target contour to describe the shape eccentricity and orientation of the object. It also calculates the area ratio of the target pixels within the region to characterize the physical scale of the object, thus obtaining geometric morphological features used to distinguish structures of different scales, such as large bridges and small cemeteries. This embodiment inputs the calculated biological texture features and geometric morphological features into a feature encoder, and transforms the multi-dimensional physical statistics into a unified high-dimensional feature vector through normalization. Furthermore, this embodiment employs a feature fusion algorithm to concatenate or weightedly combine feature vectors from different modalities to generate robust key identification features that not only preserve the microscopic texture details of the land cover but also integrate macroscopic morphological information. Finally, this embodiment establishes an index mapping relationship between land cover categories and key identification features, and serializes and stores this feature vector in a pre-set interpretation marker feature library, providing standard reference samples for subsequent instance segmentation model training.

[0021] Furthermore, the specific implementation process of step 400 is as follows: This embodiment constructs a cascaded network architecture consisting of a target detection subnetwork and a segmentation subnetwork, aiming to achieve a leap from coarse-grained localization to fine-grained segmentation through staged processing. First, the target detection subnetwork is built. This subnetwork is based on a deep convolutional neural network architecture and includes a backbone layer for feature extraction, capable of abstracting deep feature maps containing edge, texture, and semantic information from the input image layer by layer. Simultaneously, this embodiment configures a regression prediction layer at the end of this backbone layer. Through specific convolutional operations or fully connected layers, the extracted high-dimensional features are mapped to specific geometric parameters, thereby enabling the output of the bounding box coordinates of ground objects, providing accurate spatial location priors for subsequent segmentation tasks.

[0022] Following the detection subnetwork, this embodiment configures a segmentation subnetwork, which employs a more adaptive Transformer architecture to handle long-range dependencies. In this embodiment, an encoder for image encoding is deployed within the segmentation subnetwork, responsible for converting the input real-world image into a high-dimensional image embedding vector to fully preserve the global semantics and local details of the image. More importantly, this embodiment introduces a cue encoder specifically designed to receive external cues. This component is designed to parse sparse geometric cue signals and encode them into cue feature vectors aligned with the dimensions of the image embedding vector, enabling the network to understand spatial location instructions.

[0023] Finally, this embodiment includes a decoder in the segmentation sub-network for mask generation. This decoder, as a core component of multimodal information fusion, is responsible for simultaneously receiving image embedding vectors from the image encoder and cue feature vectors from the cue encoder. This embodiment utilizes a cross-attention mechanism within the decoder to calculate the semantic correlation between image features and cue features, dynamically focusing the attention of the segmentation operation on the spatial range indicated by the cue information. Through this structural design, the decoder can upsample the fused features and restore them to a binary pixel mask corresponding to the original image resolution, thereby completing pixel-level generation of the ground feature outline.

[0024] Furthermore, the specific implementation process of step 500 is as follows: This embodiment establishes a clear signal cascading communication protocol to define the data flow logic between different neural network modules. Based on this protocol, this embodiment determines a unidirectional signal transmission path from the output of the object detection sub-network to the input of the segmentation sub-network, ensuring that the inference results of preceding nodes can be losslessly transmitted to subsequent nodes. Along this path, this embodiment does not transmit the full image feature map, but instead selects to transmit geometric location information with high spatial generalization, thereby constructing an efficient computational mode for detection-guided segmentation and avoiding the segmentation sub-network blindly searching the entire image without prior knowledge.

[0025] This embodiment specifically performs the process of converting the bounding box coordinates output by the object detection subnetwork into cue information. When the object detection subnetwork identifies a ground object in the image and regresses the rectangular bounding box surrounding the object, this embodiment extracts the coordinates of the four corner points or the center point of the bounding box, as well as its width and height parameters. This embodiment formats these discrete geometric coordinate values ​​into a sparse cue vector conforming to the segmentation subnetwork interface specification, or a positional encoding sequence, so that it can be directly read and parsed by the subsequent cue encoder, thereby transforming pure geometric position data into semantic guidance signals that the neural network can understand.

[0026] This embodiment utilizes the generated prompt information to dynamically constrain the attention computation range of the segmentation sub-network. After inputting the prompt information containing the bounding box coordinates into the segmentation sub-network, this embodiment forces the decoder of the segmentation sub-network to focus only on image features within and around the bounding box, automatically ignoring background noise areas outside the bounding box during computation. In this way, this embodiment implements a spatial hard attention mechanism in the feature interaction stage, allowing the segmentation sub-network to perform refined foreground-background binary classification within a limited pixel range, rather than classifying all pixels in the entire image one by one. This significantly reduces computational load and improves the accuracy of target edge segmentation.

[0027] Furthermore, the specific implementation process of step 600 is as follows: This embodiment first implements a parameter freezing strategy for the segmentation sub-network in the cascaded network architecture. Considering that segmentation models typically contain image encoders with a large number of parameters, in order to reduce training costs and prevent overfitting on small sample data, this embodiment locks all weight parameters in the encoder by cutting off the gradient backpropagation path, keeping their values ​​constant during subsequent training. Simultaneously, this embodiment explicitly designates the cue encoder and decoder in the segmentation sub-network as trainable, allowing these two lightweight components to fine-tune their parameters based on specific real-world verification task data. This retains the powerful general feature extraction capabilities learned during large model pre-training while requiring only a very small number of adaptation layer parameters to complete domain transfer, achieving lightweight and efficient model training.

[0028] This embodiment establishes a deep interaction mechanism between the neural network model and an external expert knowledge base. Based on the land cover category labels in the current input real-world verification dataset, it accurately retrieves and calls the corresponding key identification features from a pre-set interpretation marker feature library. This embodiment uses a linear projection layer or a multilayer perceptron to map these physical-level key identification features into high-dimensional feature embedding vectors aligned with the dimensions of the network feature map. To achieve feature enhancement guided by prior knowledge, this embodiment embeds a feature attention fusion module in the backbone layer of the object detection sub-network. This module calculates the correlation matrix between the high-dimensional feature embedding vector and the current real-world image features in the channel dimension, thereby generating channel attention weights. External key identification features are then dynamically injected into the feature extraction stream of the object detection sub-network through weighted dot multiplication, enabling the network to specifically focus on key details such as forest texture or structure outlines.

[0029] This embodiment initiates the end-to-end joint training process of the model based on a prepared real-world validation dataset. Forward propagation computation is strictly performed according to the preset signal transmission path; that is, the object detection sub-network first generates localization box cues, which are then passed to the segmentation sub-network to generate the final mask. During this process, this embodiment constructs a joint objective function containing both detection and segmentation loss functions, quantifying the overall error by calculating the difference between the model's prediction results and the real-world feature annotations. This embodiment utilizes the gradient descent optimization algorithm, backpropagating the gradient based on the calculated total error to synchronously update all parameters of the object detection sub-network and the network parameters of the unfrozen layers in the segmentation sub-network. Through multiple rounds of iterative optimization, the joint loss value is continuously reduced until the model performance stabilizes and converges, ultimately resulting in a well-trained, lightweight cascaded instance segmentation model with enhanced interpretive knowledge capabilities.

[0030] Specifically, in this embodiment, the training loss of the object detection sub-network is first defined, and the loss value is composed of a weighted sum of the classification error and the localization error. Among them, classification error is calculated based on the probability of a candidate target being predicted as a specific category and the log-likelihood loss between the true category label and the actual category label of all candidate targets participating in the detection supervision. It aims to measure the accuracy of the model in judging the semantics of ground objects. Localization error is calculated by selecting the candidate target index set that is judged as positive samples, calculating the L1 norm distance between the predicted bounding box parameter vectors of these positive samples and the corresponding true bounding box parameter vectors, which is the sum of absolute errors. This measures the deviation between the predicted box and the true box in geometric position.

[0031] Meanwhile, this embodiment defines the training loss of the segmentation subnetwork, which is calculated on the set of pixels covered by the segmentation supervision. It combines pixel-level binary cross-entropy loss with Dice loss based on set overlap. Specifically, the system calculates the logarithmic difference between the true foreground label and the predicted foreground probability for each pixel. It also calculates the ratio of twice the intersection of the predicted map and the true map to the sum of the two. A very small positive constant is added to the denominator to prevent the denominator from being zero to ensure numerical stability. The ability of the model to segment irregular ground edges is optimized by minimizing this loss.

[0032] During the inference phase, the overall output process of the lightweight cascaded instance segmentation model is as follows: First, the input real-world image is input into the object detection sub-network, and the detection sub-network parameters and mapping function are used to output a series of candidate instances containing class probabilities and bounding boxes; for each retained instance, the system determines its predicted class according to the principle of maximizing probability, and triggers the retrieval mapping mechanism of the interpretation feature library, retrieving the key identification feature embedding vector of the corresponding class from the library. This vector is a high-dimensional feature vector that is pre-extracted and stored to characterize the unique visual attributes of specific land features (such as woodland texture or building form); Finally, the system inputs the real-world image, the predicted bounding box parameters, and the retrieved key identification feature embedding vectors into the segmentation sub-network. Using the parameters of the segmentation sub-network and the mapping function, under the semantic guidance of the feature vectors and the spatial constraints of the bounding box, the system generates an accurate predicted pixel mask for the i-th instance, thus completing the entire process from image input to instance segmentation result output.

[0033] To intuitively demonstrate the logic behind the model parameter values, it is assumed that the total number of samples in the batch of real-world images input to the system is set to 16, and the number of candidate targets generated by the model on a single image for detection supervision is 100, of which 10 are judged as positive samples after ground truth bounding box matching; the system has a preset total of 5 land cover categories (e.g., including woodland, buildings, water, etc.). For a positive sample labeled as "building", its ground truth category label is 1 at the corresponding building category index; in the segmentation stage, the size of the pixel set covered by the segmentation supervision is set to 64 x 64, i.e., 4096 pixels. For a pixel located at the edge of a building in this set, its ground truth foreground label is 1, while the model predicts that the probability of this pixel belonging to the foreground of a building may be 0.85. To prevent the denominator from being zero in the division operation, the focus factor of the sample weight is adjusted to 2; finally, when the model outputs its inference, the number of valid instances retained after screening is 5.

[0034] Furthermore, the specific implementation process of step 700 is as follows: This embodiment first analyzes the geospatial attribute structure of the vector data to be verified, identifies and extracts the key node sequences that constitute the external contour of the target feature, and combines the planar coordinates of these nodes in the geographic information system with the regional digital elevation model data to reconstruct a set of three-dimensional spatial vertex coordinates that can accurately describe the spatial morphology of the feature. Simultaneously, this embodiment reads the metadata information recorded in the header of the real-scene image file, decodes the precise three-dimensional position coordinates of the camera at the time of shooting as translation parameters, extracts electronic compass data or attitude angle data reflecting the orientation of the camera lens optical axis as the shooting azimuth angle, and obtains the focal length value recording the optical physical properties of the lens, preparing precise physical parameter inputs for the subsequent establishment of a spatial geometric transformation model.

[0035] This embodiment constructs a core operator for spatial transformation based on the pinhole camera imaging principle of computer graphics. Specifically, it converts the camera's shooting position coordinates into translation vectors and decomposes the shooting azimuth angle into rotational components around the spatial coordinate axes to construct a rotation matrix. These components are then combined to generate an extrinsic parameter matrix describing the relative pose relationship between the world coordinate system and the camera coordinate system. Furthermore, this embodiment combines lens focal length information with the physical dimensions and pixel resolution parameters of the image sensor to construct an intrinsic parameter matrix describing the internal optical imaging geometry of the camera. Finally, this intrinsic parameter matrix is ​​concatenated and multiplied with the extrinsic parameter matrix to generate a perspective projection transformation matrix capable of mapping three-dimensional geographic points to a two-dimensional image plane.

[0036] This embodiment converts the coordinates of each extracted 3D spatial vertex of a ground feature into homogeneous coordinates. It then performs matrix multiplication on these coordinates using the constructed perspective projection transformation matrix to calculate the projection position of the vertex on the camera's imaging plane. Perspective division is then used to normalize the result to 2D planar pixel coordinates in the real-world image. This embodiment strictly adheres to the predetermined topological connection order between vertices in the original vector data. The calculated planar pixel coordinates are sequentially connected in the pixel coordinate system of the real-world image, and the area enclosed by the connecting lines is closed using a scanline filling algorithm or a region growing algorithm, thereby generating a theoretical projection area that is strictly aligned with the viewpoint of the real-world image.

[0037] Furthermore, the specific implementation process of step 800 is as follows: This embodiment first decodes and analyzes the data stream output by the lightweight cascaded instance segmentation model, separating the predicted pixel mask representing the fine contours of ground features and the predicted category label indicating the semantic attributes of the ground features. To quantitatively evaluate the degree of spatial agreement between the real-world ground features identified by the model and the vector ground features in the database, this embodiment uses the intersection-union algorithm in image processing to perform pixel-level superposition calculations on the predicted pixel mask and the theoretical projection area generated in the previous step. This embodiment counts the number of pixels overlapping these two regions on the two-dimensional image plane as the intersection area, and the total number of pixels covered by the combined area as the union area. Then, the spatial overlap coefficient is obtained by calculating the ratio of the intersection area to the union area. This coefficient, as a normalized quantitative indicator, intuitively reflects the level of consistency in spatial distribution between the ground feature boundaries described by the vector data and the actual ground feature boundaries presented in the real-world photograph.

[0038] This embodiment further retrieves the attribute field information of the vector data to be verified, and reads the original feature attribute code registered in the original geographic information system for the feature object. To address potential encoding differences between different data sources, this embodiment pre-defines a set of semantic mapping rules to uniformly convert the predicted category labels output by the model and the read original feature attribute codes into a standardized feature semantic category space. This embodiment compares whether the two point to the same type of feature entity within the unified semantic space. For example, it determines whether the forest land label identified by the model matches the forest land attribute recorded in the vector data, thereby obtaining a clear attribute consistency judgment result, achieving a technological leap from simple geometric location comparison to deep semantic attribute verification.

[0039] This embodiment establishes a verification logic based on dual constraints of space and attributes, setting a preset overlap threshold as the baseline for measuring spatial location accuracy. This embodiment monitors the above calculation process in real time. Once the spatial overlap coefficient is found to be less than the preset overlap threshold, it indicates that the vector data has a significant positional offset or shape distortion. Alternatively, if the attribute consistency judgment result shows "no," it indicates that the attribute records of the vector data do not match the actual scene. This embodiment immediately triggers a correction suggestion generation program. This embodiment calculates the geometric center coordinates of the theoretical projection area as the precise location index and uses the high-confidence predicted category labels identified by the model as attribute modification suggestions. These are combined to generate classification verification correction information containing erroneous location and correction content, providing a reliable basis for subsequent data updates and maintenance.

[0040] This embodiment also provides a real-scene photo instance segmentation and classification verification system, including: The real-scene image acquisition module is used to acquire the corresponding real-scene image based on the spatial distribution information of the vector data to be verified; The real-scene verification dataset construction module is used to perform effective cleaning on the real-scene images and filter out images that meet the imaging scale of ground features according to preset spatial constraints to obtain a real-scene verification dataset containing geographic location information. The interpretation feature library establishment module is used to extract key identification features of preset land cover categories in the real scene image and store the key features into the interpretation feature library; Cascading network architecture building module, used to build cascading network architecture; A signal transmission path determination module is used to determine the signal transmission path of the cascaded network architecture based on preset transmission rules. The model training and embedding module is used to train the cascaded network architecture based on the signal transmission path using the real-world verification dataset, and to embed the key identification features in the interpretation feature library into the cascaded network architecture to obtain a trained lightweight cascaded instance segmentation model. The theoretical projection region generation module is used to project the vector data to be verified onto the pixel coordinate system of the real scene image using a spatial mapping transformation algorithm to obtain the theoretical projection region. The classification verification and correction module is used to compare the segmentation results output by the lightweight cascaded instance segmentation model with the theoretical projection region to calculate the spatial overlap and attribute consistency; if the overlap and attribute consistency do not meet the preset requirements, classification verification and correction information is generated.

[0041] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0042] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for segmenting and classifying real-world photo instances for verification, characterized in that, include: Based on the spatial distribution information of the vector data to be verified, obtain the corresponding real-world image; The real-scene images are cleaned for effectiveness, and images that meet the imaging scale of ground features are selected according to preset spatial constraints to obtain a real-scene verification dataset containing geographic location information. Extract key identification features of preset land cover categories in the real scene image, and store the key identification features in the interpretation mark feature library; Build a cascaded network architecture; The signal transmission path of the cascaded network architecture is determined based on preset transmission rules; Based on the signal transmission path, the cascaded network architecture is trained using the real-world verification dataset, and key identification features from the interpretation marker feature library are embedded into the cascaded network architecture to obtain a trained lightweight cascaded instance segmentation model. The vector data to be verified is projected onto the pixel coordinate system of the real scene image using a spatial mapping transformation algorithm to obtain the theoretical projection area; The segmentation results output by the lightweight cascaded instance segmentation model are compared with the theoretical projection region to calculate spatial overlap and attribute consistency; if the overlap and attribute consistency do not meet the preset requirements, classification verification correction information is generated.

2. The method for segmenting and classifying real-scene photos according to claim 1, characterized in that, The step of obtaining the corresponding real-world image based on the spatial distribution information of the vector data to be verified includes: The geometric coordinate sequence of the vector data to be verified is parsed, and a spatial retrieval buffer covering a preset distance range is constructed based on the geometric coordinate sequence; Based on the spatial retrieval buffer, an image acquisition task containing geographic location parameters is determined, and the image acquisition task is mapped to a preset distributed task scheduling queue; The system retrieves a pre-set pool of network access accounts and concurrently executes image acquisition tasks from the distributed task scheduling queue via multi-threaded long-connection communication to download real-world images from the network platform.

3. The method for segmenting and classifying real-scene photograph instances according to claim 1, characterized in that, The process involves effective cleaning of the real-world images and filtering out images that meet the ground feature imaging scale based on preset spatial constraints to obtain a real-world verification dataset containing geographic location information, including: Calculate the pixel grayscale distribution statistical features of the real scene image. If the pixel grayscale distribution statistical features meet the preset non-completely black image standard, then mark the corresponding real scene image as a valid candidate image. The shooting location coordinates are parsed from the metadata of the valid candidate images, and the spatial Euclidean distance between the shooting location coordinates and the geometric center of the vector data to be verified is calculated. Based on the inverse relationship between the spatial Euclidean distance and the imaging scale of ground features, valid candidate images with a spatial Euclidean distance less than a preset sharpness threshold are selected. The selected valid candidate images and their corresponding shooting location coordinates are associated and stored to obtain the real-scene verification dataset.

4. The method for segmenting and classifying real-scene photo instances according to claim 1, characterized in that, The step of extracting key identification features of preset land cover categories in the real-world image and storing the key identification features in the interpretation marker feature library includes: Semantic saliency detection is performed on the real-scene image to crop out the target region of interest corresponding to the preset land cover category; Based on the target region of interest belonging to the vegetation category, the texture periodicity and color distribution parameters are calculated to obtain biological texture features; Based on the target region of interest belonging to the category of artificial structures, the contour geometric moments and pixel area ratio are calculated to obtain the geometric morphological features; The biological texture features and the geometric morphological features are vectorized, encoded, and fused to generate the key identification features, and the key identification features are mapped and stored in the interpretation marker feature library.

5. The method for segmenting and classifying real-scene photo instances according to claim 1, characterized in that, The cascaded network architecture includes: Object detection subnetwork and segmentation subnetwork; The target detection subnetwork includes a backbone layer for feature extraction and a regression prediction layer for outputting bounding box coordinates. The segmentation subnetwork includes an encoder for image encoding, a cue encoder for receiving external cue information, and a decoder for generating a mask.

6. The method for segmenting and classifying real-scene photo instances according to claim 5, characterized in that, The preset transmission rules include: The coordinates of the bounding box output by the target detection subnetwork are used as prompt information and input into the segmentation subnetwork to limit the range of pixels that the segmentation subnetwork can operate on.

7. The method for segmenting and classifying real-scene photo instances according to claim 5, characterized in that, Based on the signal transmission path, the cascaded network architecture is trained using the real-world verification dataset, and key identification features from the interpretation marker feature library are embedded into the cascaded network architecture to obtain a trained lightweight cascaded instance segmentation model, including: Perform a parameter freeze operation on the encoder in the segmentation sub-network to determine that the cue encoder and the decoder are in a trainable state. The key identification features corresponding to the labels of the real-world verification dataset are retrieved from the interpretation feature library, and the key identification features are mapped into high-dimensional feature embedding vectors. A feature attention fusion module is constructed in the backbone layer of the target detection sub-network. The feature attention fusion module is used to calculate the channel attention weights of the high-dimensional feature embedding vector and the image features in the real-scene verification dataset. The key identification features are then injected into the feature extraction process of the target detection sub-network. Based on the real-world verification dataset, forward propagation calculation is performed according to the signal transmission path. The detection loss function of the target detection sub-network and the segmentation loss function of the segmentation sub-network are jointly optimized. The network parameters of the non-frozen layer are updated until the model converges, and the trained lightweight cascaded instance segmentation model is obtained. The expression for the detection loss function is: ； The expression for the segmentation loss function is: ； The expression for the lightweight cascaded instance segmentation model is: ； in, The training loss for the target detection subnetwork; The training loss for segmenting the subnetwork; The number of candidate targets to participate in the testing and supervision; Index the candidate targets; For the first The predicted candidate target is the _th The probability of a class (by (Given) For the first The true category label of each candidate target; This is the set of candidate target indices that are judged as positive samples; The number of positive samples; For the first The predicted bounding box parameter vector of each candidate target; This is the corresponding true bounding box parameter vector; For vectors Norm; To segment the set of pixels covered by the supervision; Size of the pixel set; For pixel index; For pixels True foreground markers; For pixels The predicted prospect probability; To prevent extremely small positive constants with a denominator of zero; I represents the input real-world image; The overall output of I for the lightweight cascaded instance segmentation model; To detect and retain the number of instances output by the sub-network; For the target detection subnetwork mapping function; For subnetwork mapping functions; To detect subnetwork parameters; For the parameters of the segmented sub-network; This is the total set of parameters for the cascaded model; For the reason from The first selected Predict the category for each instance; To interpret the retrieval mapping of the feature database; To extract from the interpretable feature library by category The extracted key identification feature embedding vector; For the first Predicted pixel mask for each instance.

8. The method for segmenting and classifying real-scene photo instances according to claim 1, characterized in that, The vector data to be verified is projected onto the pixel coordinate system of the real-scene image using a spatial mapping transformation algorithm to obtain the theoretical projection area, including: The geospatial attributes of the vector data to be verified are analyzed to extract the three-dimensional spatial vertex coordinates of the ground feature outlines; Camera imaging parameters are extracted from the metadata of the real-scene image. The camera imaging parameters include: shooting position coordinates, shooting azimuth angle and lens focal length information. Based on the camera imaging parameters, a perspective projection transformation matrix is ​​constructed that connects the geographic world coordinate system and the image pixel coordinate system; The perspective projection transformation matrix is ​​used to perform matrix multiplication on the coordinates of the three-dimensional space vertex to calculate the planar pixel coordinates of the three-dimensional space vertex on the real scene image, and a closed theoretical projection region is generated based on the topological connection relationship of the planar pixel coordinates.

9. The method for segmenting and classifying real-scene photograph instances according to claim 1, characterized in that, The segmentation results output by the lightweight cascaded instance segmentation model are compared with the theoretical projection region to calculate spatial overlap and attribute consistency. If the overlap and attribute consistency do not meet the preset requirements, classification verification correction information is generated, including: The predicted pixel mask and predicted category label are parsed from the segmentation results output by the lightweight cascaded instance segmentation model. The spatial overlap coefficient is obtained by calculating the ratio of the intersection area to the union area between the predicted pixel mask and the theoretical projection region. The attribute fields of the vector data to be verified are called to obtain the original land feature attribute code, and it is determined whether the predicted category label and the original land feature attribute code belong to the same land feature semantic category to obtain the attribute consistency determination result. If the spatial overlap coefficient is less than the preset overlap threshold, or if the attribute consistency determination result is otherwise, the center coordinates of the theoretical projection area are extracted, and the classification verification correction information containing the location index and attribute modification suggestions is generated in combination with the predicted category label.

10. A real-scene photo instance segmentation and classification verification system, characterized in that, include: The real-scene image acquisition module is used to acquire the corresponding real-scene image based on the spatial distribution information of the vector data to be verified; The real-scene verification dataset construction module is used to perform effective cleaning on the real-scene images and filter out images that meet the imaging scale of ground features according to preset spatial constraints to obtain a real-scene verification dataset containing geographic location information. The interpretation feature library establishment module is used to extract key identification features of preset land cover categories in the real scene image and store the key features into the interpretation feature library; Cascading network architecture building module, used to build cascading network architecture; A signal transmission path determination module is used to determine the signal transmission path of the cascaded network architecture based on preset transmission rules. The model training and embedding module is used to train the cascaded network architecture based on the signal transmission path using the real-world verification dataset, and to embed the key identification features in the interpretation feature library into the cascaded network architecture to obtain a trained lightweight cascaded instance segmentation model. The theoretical projection region generation module is used to project the vector data to be verified onto the pixel coordinate system of the real scene image using a spatial mapping transformation algorithm to obtain the theoretical projection region. The classification verification and correction module is used to compare the segmentation results output by the lightweight cascaded instance segmentation model with the theoretical projection region to calculate the spatial overlap and attribute consistency; if the overlap and attribute consistency do not meet the preset requirements, classification verification and correction information is generated.

Images