A smart measurement system based on immigration indicator surveys

By using a deep learning-based intelligent measurement system, the problems of low efficiency, poor data quality, and data silos in traditional immigration indicator surveys have been solved, achieving an efficient and automated immigration indicator survey process and ensuring data consistency and comparability.

CN123118328APending Publication Date: 2026-11-17POWER CHINA KUNMING ENG CORP LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611261052.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-19
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Traditional immigration indicator surveys are inefficient, have inconsistent data quality, suffer from severe data silos, lack standardized information recording methods, and lack automated processing of image data, resulting in long survey cycles, large errors, and insufficient data utilization.

Method used

An intelligent measurement system based on deep learning is adopted, including an image acquisition module, a human and housing survey intelligent recognition module, a land survey intelligent recognition module, an indicator calculation and report generation module, and a manual review and correction module. Through the organic integration of image acquisition, intelligent recognition, indicator calculation, and manual review, an automated survey process is achieved.

Benefits of technology

Significantly improve survey efficiency, ensure data consistency and quality, achieve effective correlation between images and indicator data, shorten the survey cycle, and improve data comparability and traceability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118328A_ABST
    Figure CN123118328A_ABST
Patent Text Reader

Abstract

This invention relates to an intelligent measurement system based on immigration indicator surveys, belonging to the fields of immigration survey and image recognition technology. The system includes: an image acquisition module for acquiring images at the survey site and evaluating their quality in real time; a housing and dwelling survey intelligent recognition module for identifying building structures, decoration details, and facilities and equipment using a deep learning model; a land survey intelligent recognition module for identifying land types, above-ground attachments, and trees; an indicator calculation and report generation module for integrating indicator data and generating survey reports; and a manual review and correction module for graded review based on confidence level and recording correction logs. This invention achieves full-process intelligentization of immigration indicator surveys from image acquisition to report generation, significantly improving survey efficiency and data consistency, and ensuring the traceability of survey results.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of immigration survey and image recognition technology, specifically relating to an intelligent measurement system based on immigration indicator surveys. Background Technology

[0002] Infrastructure construction projects such as water conservancy and hydropower projects are often accompanied by large-scale land acquisition and resettlement work. Resettlement indicator surveys are a fundamental task in the preparation of resettlement plans, covering dozens of indicators including population, housing structure and area, detailed renovations, facilities and equipment, land use types, above-ground attachments, and tree species and quantities. The accuracy and completeness of the survey results directly affect the formulation and implementation of resettlement compensation plans, and are crucial to the vital interests of the resettled population and the smooth progress of project construction.

[0003] Currently, resettlement indicator surveys primarily rely on traditional manual door-to-door surveys. Surveyors carry paper forms, measuring tapes, cameras, and other tools, conducting on-site inspections, measurements, and records for each household and plot of land. This traditional method suffers from the following prominent problems: First, it is inefficient; a resettlement survey for a medium-sized reservoir project often requires dozens of surveyors and several months to complete, resulting in high manpower costs and long cycles. Second, data quality is inconsistent; manual operation is prone to subjective errors, and different surveyors may use inconsistent standards for judging the same indicator. Third, data silos are severe; image data is only used for archiving, and paper records require subsequent manual entry and summarization, which is not only labor-intensive but also prone to errors. There is a lack of effective correlation and automated processing methods between images and indicator data. Fourth, information recording methods are inconsistent, leading to a high error rate in statistical summarization. Although some surveys have introduced digital tools such as digital cameras and spreadsheets in recent years, the core model of manual measurement and recording remains largely unchanged, and the rich information contained in image data has not been fully explored and utilized.

[0004] With the rapid development of artificial intelligence technologies such as deep learning, semantic segmentation, object detection, and fine-grained image classification have demonstrated powerful automated processing capabilities in fields such as architectural surveying and land use surveys, providing a technological foundation for the intelligent transformation of immigration indicator surveys. However, there is currently a lack of a complete system solution that organically integrates image acquisition, intelligent recognition, indicator calculation, report generation, and manual review to systematically address the problems of low efficiency, large errors, and data silos in traditional immigration indicator surveys. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, this invention provides an intelligent measurement system based on immigration indicator surveys.

[0006] The objective of this invention can be achieved through the following technical solutions: An intelligent measurement system based on immigration indicator surveys, comprising: The image acquisition module collects image data of the survey subjects at the immigration survey site through mobile terminal devices, performs real-time quality assessment of the collected images, and associates and binds qualified images with the survey subject information and geographical coordinates. The intelligent identification module for human and housing surveys preprocesses and registers the house images acquired by the image acquisition module, and uses a pre-trained deep learning model to identify the house structure, decoration details, and facilities and equipment, and outputs human and housing survey index data and its identification confidence level. The land survey intelligent identification module preprocesses the land and surface attachment images acquired by the image acquisition module, and uses a pre-trained deep learning model to identify land type, surface attachments, and trees, and outputs land survey index data and its identification confidence level. The indicator calculation and report generation module standardizes and integrates the indicator data output by the intelligent identification module for housing and population survey and the intelligent identification module for land survey according to a unified survey indicator system, performs indicator calculation and summarization, and generates a structured immigration survey report. The manual review and correction module classifies the intelligent recognition results according to the recognition confidence level, generates a manual review task queue, provides a correction interface for reviewers to review and correct the recognition results, and records the correction operation log.

[0007] In a further embodiment of the present invention, the image acquisition module includes: The data collection planning unit provides differentiated shooting guidelines based on the type of survey task. In the human and housing survey mode, it guides the shooting in the order of the overall appearance of the house, each functional room, decoration details, and facilities and equipment. In the land survey mode, it guides the shooting in the order of the plot panorama, above-ground attachments, and individual trees. The image quality assessment unit performs real-time sharpness and illumination uniformity detection on the acquired images using the Laplacian variance or BRISQUE no-reference image quality assessment method, and prompts the user to retake images that do not meet the quality standards. The spatial positioning and association unit records the geographic coordinates, shooting time, and shooting angle of each image, and associates and binds the images with the corresponding survey objects.

[0008] In a further embodiment of the present invention, the intelligent identification module for human-room surveys includes: The image preprocessing and registration unit performs grayscale conversion, noise reduction, and enhancement processing on the building images, and uses an image registration algorithm to spatially align images taken from different angles or under different conditions. The building structure recognition unit uses a pre-trained semantic segmentation network to perform pixel-level semantic segmentation on building images, identify the building structure type, the boundaries and areas of each functional room, and use standard size reference objects in the image to calibrate the pixels with the actual size, and estimate the building's span, depth and floor height. The decoration detail identification unit uses a fine-grained image classification model to automatically identify and evaluate the material type and decoration level of the floor, walls, and ceiling inside the house. The material type includes tiles, wood flooring, latex paint, wallpaper, and suspended ceiling, and the decoration level includes high-end, mid-range, ordinary, and simple. The facility and equipment identification unit uses target detection algorithms to detect, locate, and classify doors, windows, kitchen and bathroom facilities, and water and electricity facilities in building images, and estimates their quantity and size.

[0009] In a further embodiment of the present invention, the land survey intelligent identification module includes: The land type identification unit classifies land use types in the land parcel image using a pre-trained deep learning classification model. The land use types include cultivated land, orchard, forest land, grassland, construction land, and water area. The ground attachment identification unit uses a combination of target detection and semantic segmentation to automatically identify and locate retaining walls, ditches, roads, culverts, bridges and drying yards in the image, and estimate their length, width, height, cross-sectional shape and lining material type; The tree identification and statistics unit uses a fine-grained image classification model to identify tree species, performs tree-by-tree detection and counting through a target detection algorithm, and estimates the diameter at breast height (DBH) range of trees by calibrating reference objects in the image.

[0010] In a further embodiment of the present invention, the indicator calculation and report generation module includes: The indicator data integration unit maps the original identification results output by each identification module into standardized survey indicator items through a preset survey indicator mapping table. Each standardized survey indicator item includes indicator ID, indicator name, indicator attribute, indicator value and corresponding confidence level. The distributed computing unit divides the large-scale immigration survey task into pieces according to administrative region, land parcel number or household head number, and distributes them to multiple computing nodes to perform index calculation and summarization in parallel. Each computing node receives the task through a Redis message queue and writes the calculation results to a shared database. The caching and reuse unit uses the MD5 value of the generated metric calculation as the cache key to cache the calculated metric results in Redis. For the corrected metric data, the data is stored in the correction and solidification table. Subsequent calculations will prioritize using the cached or solidified data. The report generation unit automatically generates household-specific reports for housing surveys, land-specific reports for land surveys, and summary statistical reports by administrative level based on the integrated standardized indicator data.

[0011] In a further embodiment of the present invention, in the distributed computing unit, each computing node registers node information with a preset Redis center and sets an expiration time when it starts up, and sends heartbeat data to Redis periodically through a local heartbeat mechanism to keep it alive; when allocating tasks, the actual number of tasks to be processed is modulo the number of currently alive nodes to calculate the number of tasks that each node needs to process.

[0012] In a further embodiment of the present invention, the manual review and correction module includes: The confidence level labeling and priority sorting unit divides the identification results into three levels: high confidence, medium confidence, and low confidence, based on the identification confidence level, and generates a queue of manual review tasks according to the principle of prioritizing low confidence and key indicators. The difference comparison and annotation unit compares and overlays the intelligent recognition results with the original image pixel by pixel, highlights the areas on the image where the recognition results are questionable, and displays the recognition labels for reviewers to compare. The manual correction interface unit provides a correction interface with drop-down menus and input boxes, allowing auditors to modify the values ​​of indicators that are incorrectly identified and manually add annotations of missed objects to images. It also automatically records the person making the correction, the correction time, and an audit log comparing the correction before and after.

[0013] In a further embodiment of the present invention, the pre-trained deep learning models in the intelligent identification module for housing and land surveys have training samples containing labeled immigration survey image data. The models output the probability value of the corresponding category as the identification confidence level while outputting the identification result.

[0014] In a further embodiment of the present invention, the manual review and correction module further includes a feedback learning unit, which is used to periodically add the manually corrected data to the training dataset of each recognition model, perform incremental training or fine-tuning of the model, track changes in accuracy and recall through model performance evaluation, and automatically release a new version of the model when the model performance reaches a preset release threshold.

[0015] In a further embodiment of the present invention, the image acquisition module, the intelligent identification module for human and housing surveys, the intelligent identification module for land surveys, the indicator calculation and report generation module, and the manual review and correction module interact with each other via a cloud server, wherein the image acquisition module is deployed on a mobile terminal, and the other modules are deployed on a cloud server or a local server.

[0016] The present invention has at least the following beneficial effects: 1. By using image acquisition and deep learning-based intelligent recognition to replace traditional manual measurement and recording, dozens of survey indicators can be automatically extracted and identified from an image within seconds, improving overall survey efficiency and shortening the immigration survey cycle.

[0017] 2. The deep learning model adopts a unified recognition standard and judgment logic, avoiding subjective judgment differences between different investigators; the systematic image quality assessment ensures data quality from the source, ensuring the comparability and consistency of survey data. Attached Figure Description

[0018] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0019] Figure 1 This is a system framework diagram provided in one embodiment of the present invention; Figure 2 This is a block diagram of the internal structure of the intelligent identification module for human-room surveys provided in one embodiment of the present invention; Figure 3 This is a block diagram of the internal structure of the land survey intelligent identification module provided in one embodiment of the present invention. Detailed Implementation

[0020] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.

[0021] Please refer to the details. Figure 1-3 In one embodiment, the present invention provides an intelligent measurement system based on an immigration indicator survey, comprising: The image acquisition module collects image data of the survey subjects at the immigration survey site through mobile terminal devices, performs real-time quality assessment of the collected images, and associates and binds qualified images with the survey subject information and geographical coordinates. The intelligent identification module for human and housing surveys preprocesses and registers the house images acquired by the image acquisition module, and uses a pre-trained deep learning model to identify the house structure, decoration details, and facilities and equipment, and outputs human and housing survey index data and its identification confidence level. The land survey intelligent identification module preprocesses the land and surface attachment images acquired by the image acquisition module, and uses a pre-trained deep learning model to identify land type, surface attachments, and trees, and outputs land survey index data and its identification confidence level. The indicator calculation and report generation module standardizes and integrates the indicator data output by the intelligent identification module for housing and population survey and the intelligent identification module for land survey according to a unified survey indicator system, performs indicator calculation and summarization, and generates a structured immigration survey report. The manual review and correction module classifies the intelligent recognition results according to the recognition confidence level, generates a manual review task queue, provides a correction interface for reviewers to review and correct the recognition results, and records the correction operation log.

[0022] The resettlement survey is a fundamental step in the land acquisition and resettlement work for water conservancy and hydropower projects. The accuracy and completeness of the survey results directly affect the formulation and implementation effectiveness of the resettlement compensation plan. Traditional resettlement surveys primarily rely on manual door-to-door surveys, with investigators carrying paper forms, measuring tapes, cameras, and other tools to conduct on-site inspections, measurements, and records for each household and plot of land. This method has several inherent drawbacks: firstly, manual measurement and recording are inefficient; a resettlement survey for a medium-sized reservoir project often requires dozens of investigators and several months to complete; secondly, manual operation is prone to subjective errors, with different investigators potentially using inconsistent standards to judge the same survey subject, leading to inconsistent data quality; furthermore, paper-based data requires subsequent manual entry and summarization, which is not only labor-intensive but also prone to errors. In recent years, although some surveys have begun to introduce digital tools such as digital cameras and spreadsheets, there is still a lack of effective correlation and automated processing methods between image data and indicator data. Image data is only used as archival evidence and fails to fully realize its value in intelligent identification and automated measurement. With the rapid development of artificial intelligence technologies such as deep learning, image recognition, object detection, and semantic segmentation have demonstrated powerful automated processing capabilities in fields such as architectural surveying and land use investigation, providing a technological foundation for the intelligent transformation of immigration indicator surveys. However, there is currently a lack of a complete system solution that organically integrates image acquisition, intelligent recognition, indicator calculation, report generation, and manual review to systematically address issues such as low efficiency, large errors, and data silos in immigration indicator surveys.

[0023] In this embodiment, the system consists of five core modules: an image acquisition module, a human and housing survey intelligent identification module, a land survey intelligent identification module, an indicator calculation and report generation module, and a manual review and correction module. The modules interact with each other through a cloud server.

[0024] The image acquisition module, deployed on a mobile terminal device, is responsible for collecting image data of survey subjects at the immigration survey site. During the acquisition process, the module utilizes the mobile terminal's camera hardware to capture images according to preset shooting parameters. For each captured image, the module immediately invokes a built-in image quality assessment subroutine to perform real-time detection of image sharpness and illumination uniformity by calculating the Laplacian variance or the BRISQUE no-reference image quality evaluation index. The Laplacian variance assesses the sharpness of image edges by calculating the variance value after the Laplacian transform; a higher variance value indicates a sharper image. BRISQUE extracts the spatial natural scene statistical features of the image to provide a quality score within the range of 0 to 100; a lower score indicates higher image quality. If the image quality is substandard, the module displays a prompt on the mobile terminal screen, requesting the investigator to retake the image. Images meeting quality standards are then associated with the survey subject information, current geographical coordinates, shooting time, and shooting angle, forming a structured image data record, which is then uploaded to the cloud server in real time.

[0025] The intelligent identification module for housing surveys is deployed on a cloud server, receiving housing image data from the image acquisition module. This module first preprocesses the housing images, including grayscale conversion, noise reduction, and enhancement, to eliminate the impact of uneven lighting and sensor noise on the identification results. Then, an image registration algorithm is used to spatially align multiple images of the same house taken from different angles or under different lighting conditions, providing a data foundation for subsequent joint identification. In the identification phase, the module calls a pre-trained deep learning model to complete three identification tasks: housing structure identification uses a semantic segmentation network to perform pixel-level semantic segmentation of the housing images, identifying the structural type of the house and the boundaries and areas of each functional room; detailed decoration item identification uses a fine-grained image classification model to automatically identify the material type and decoration level of the floor, walls, and ceiling; and facility and equipment identification uses a target detection algorithm to detect, locate, and classify doors, windows, kitchen and bathroom facilities, and water and electricity facilities. Each identification model outputs the probability value of the corresponding category as the identification confidence level along with the identification result. Finally, the module outputs structured housing survey index data, with each index accompanied by a corresponding confidence value.

[0026] The land survey intelligent identification module is also deployed on a cloud server, receiving images of land and surface attachments from the image acquisition module. After preprocessing the land images similarly to the building images, this module calls a pre-trained deep learning model to complete three identification tasks: land type identification, which classifies land parcels into land use types such as cultivated land, orchards, forest land, grassland, construction land, or water areas using a classification model; surface attachment identification, which uses a combination of object detection and semantic segmentation to automatically identify and locate retaining walls, ditches, roads, culverts, bridges, and drying yards in the images; and tree identification and counting, which uses a fine-grained image classification model to identify tree species and an object detection algorithm to detect and count each tree individually. Each identification result also includes a confidence score.

[0027] The indicator calculation and report generation module receives indicator data from the two identification modules: the housing and population survey and the land survey. It then standardizes and integrates this data according to a unified survey indicator system. This module maps the raw identification results from each module to standardized survey indicator items using a pre-defined indicator mapping table. Each standardized indicator item includes an indicator ID, indicator name, indicator attribute, indicator value, and corresponding confidence level. After integration, the module performs indicator calculation and summarization. For example, it calculates the total building area based on the area data of each functional room, and summarizes the total number of attachments on a plot based on the quantity and size of each type of attachment. Finally, the module automatically generates a structured immigration survey report based on the integrated standardized indicator data, including household reports, plot reports, and summary statistical reports by administrative level.

[0028] The manual review and correction module categorizes the intelligent recognition results based on their recognition confidence level. This module marks recognition results with a confidence level higher than 0.9 as high confidence, those between 0.7 and 0.9 as medium confidence, and those below 0.7 as low confidence. A queue of manual review tasks is generated based on the principle of prioritizing low confidence and key indicators, and then pushed to the reviewers' work interface. Reviewers can view the overlay comparison between the original image and the intelligent recognition result in the correction interface, modify the values ​​of incorrectly identified indicators, or manually add annotations for missed objects to the image. All correction operations are automatically recorded in an audit log, including the person making the correction, the correction time, and a comparison of the before and after corrections.

[0029] This embodiment constructs a complete automated survey chain from image acquisition to report generation through the collaborative work of five modules, realizing full-process intelligentization of immigration indicator surveys. Compared with existing technologies, this embodiment has at least the following significant technical effects: First, it greatly improves survey efficiency. Traditional manual surveys require manual measurement and recording of each household and item, while this system, through image acquisition and intelligent recognition, can automatically extract and recognize dozens of survey indicators from an image within seconds, improving overall survey efficiency by several to tens of times. Second, it significantly improves data consistency. The deep learning model uses unified recognition standards and judgment logic, avoiding subjective judgment differences between different investigators and ensuring the comparability and consistency of survey data. Third, through confidence level grading and manual review mechanisms, it ensures data quality while enjoying automation efficiency. Low-confidence results are manually reviewed, while high-confidence results are automatically approved, achieving an optimal balance between efficiency and quality. Fourth, the association and binding of image data and indicator data ensures the traceability of survey results. Any indicator can be traced back to the original collected image, facilitating subsequent auditing and quality checks. Fifth, the system adopts a modular architecture, with image acquisition deployed on mobile terminals and recognition and calculation deployed in the cloud. This ensures both the flexibility of on-site collection and full utilization of cloud computing resources, providing good scalability and deployment flexibility.

[0030] In a further embodiment of the present invention, the image acquisition module includes: The data collection planning unit provides differentiated shooting guidelines based on the type of survey task. In the human and housing survey mode, it guides the shooting in the order of the overall appearance of the house, each functional room, decoration details, and facilities and equipment. In the land survey mode, it guides the shooting in the order of the plot panorama, above-ground attachments, and individual trees. The image quality assessment unit performs real-time sharpness and illumination uniformity detection on the acquired images using the Laplacian variance or BRISQUE no-reference image quality assessment method, and prompts the user to retake images that do not meet the quality standards. The spatial positioning and association unit records the geographic coordinates, shooting time, and shooting angle of each image, and associates and binds the images with the corresponding survey objects.

[0031] In this embodiment, the data acquisition planning unit provides differentiated shooting guidelines based on the type of survey task. When the investigator selects the "Person and House Survey Mode" on the mobile terminal, the data acquisition planning unit displays a standardized shooting process guide on the screen using a combination of text and images, guiding the investigator to take photos in the order of the overall appearance of the house, each functional room, decoration details, and facilities and equipment. Specifically, the system first prompts the investigator to take photos of the overall appearance of the house from the front, side, and back, then prompts them to enter each functional room to take panoramic photos of the interior, followed by close-up shots of decoration details such as the floor, walls, and ceiling, and finally prompts them to take detailed photos of equipment such as doors, windows, kitchen and bathroom facilities, and water and electricity facilities. Each shooting step is accompanied by example pictures and text descriptions to ensure that the investigator clearly understands the content and angle requirements for the current shooting. When the investigator selects the "Land Survey Mode," the data acquisition planning unit switches to land survey shooting guidelines, guiding the investigator to take photos in the order of the overall view of the plot, above-ground attachments, and individual trees. The system first prompts you to take panoramic photos covering the entire plot at key points along the plot's boundary. Then, it prompts you to photograph each attachment within the plot, such as retaining walls, ditches, roads, culverts, bridges, and drying yards. Finally, it prompts you to photograph individual trees within the plot. The data acquisition and planning unit also includes a built-in shooting progress tracking function, displaying completed and pending shooting items in real-time as a progress bar or checklist to prevent missed shots.

[0032] The image quality assessment unit initiates quality detection immediately after each image acquisition. This unit employs two complementary no-reference image quality assessment methods for comprehensive evaluation. First, Laplacian variance detection: after converting the acquired image to grayscale, the variance of the Laplacian transform is calculated. The Laplacian transform detects edge information in the image through its second derivative; images with clear focus have higher Laplacian variance values, while blurred images have significantly lower variance values. Second, BRISQUE (Blind / Referenceless Image Spatial Quality Evaluator): this algorithm extracts statistical features of the natural scene in the spatial domain from the image and provides a quality score within the range of 0 to 100 using a trained regression model; lower scores indicate higher image quality. This unit also detects illumination uniformity, determining the presence of overexposed or underexposed areas by calculating the distribution characteristics of the image brightness histogram. When the Laplace variance is lower than the preset threshold, or the BRISQUE score exceeds the preset threshold, or the illumination uniformity does not meet the requirements, the image quality assessment unit determines that the image quality is substandard, pops up a prompt box on the mobile terminal screen and displays the specific reasons for the substandard quality, and asks the investigators to retake the picture until a qualified image is acquired.

[0033] The spatial positioning and association unit automatically records metadata upon successful image acquisition. This unit uses the mobile terminal's GPS module to obtain the current geographic coordinates, the system clock to obtain the capture time, and the mobile terminal's gyroscope and accelerometer to obtain the capture angle. This metadata, along with the image file itself, is encapsulated into a structured image record and associated with the corresponding survey object through a survey object identifier. There are two association methods: first, at the start of acquisition, the investigator manually selects or scans the survey object's QR code / barcode, and all subsequent images are automatically associated with that survey object; second, after acquisition, images are batch-associated with specified survey objects via point-and-click on the mobile terminal. All image records and their metadata are uploaded to the cloud server in real-time or in batches after acquisition for use by the subsequent recognition module.

[0034] In a further embodiment of the present invention, the intelligent identification module for human-room surveys includes: The image preprocessing and registration unit performs grayscale conversion, noise reduction, and enhancement processing on the building images, and uses an image registration algorithm to spatially align images taken from different angles or under different conditions. The building structure recognition unit uses a pre-trained semantic segmentation network to perform pixel-level semantic segmentation on building images, identify the building structure type, the boundaries and areas of each functional room, and use standard size reference objects in the image to calibrate the pixels with the actual size, and estimate the building's span, depth and floor height. The decoration detail identification unit uses a fine-grained image classification model to automatically identify and evaluate the material type and decoration level of the floor, walls, and ceiling inside the house. The material type includes tiles, wood flooring, latex paint, wallpaper, and suspended ceiling, and the decoration level includes high-end, mid-range, ordinary, and simple. The facility and equipment identification unit uses target detection algorithms to detect, locate, and classify doors, windows, kitchen and bathroom facilities, and water and electricity facilities in building images, and estimates their quantity and size.

[0035] In this embodiment, the image preprocessing and registration unit is responsible for optimizing the quality and spatially aligning the acquired house images. In the preprocessing stage, this unit performs the following operations sequentially: converting color images to grayscale to reduce computational complexity; removing noise from the images using Gaussian filtering or median filtering; and improving image contrast using histogram equalization or adaptive contrast enhancement algorithms to better present details in both dark and bright areas. In the registration stage, this unit employs a feature-based image registration algorithm. First, it extracts SIFT or ORB feature points from each image. Then, it selects reliable matching point pairs through feature point matching and the RANSAC algorithm. Finally, it calculates the homography transformation matrix between images, transforming multiple images of the same house taken from different angles or under different lighting conditions to the same coordinate space, achieving spatial alignment. The registered images can be stitched together to form a more complete panoramic view of the house, or they can be jointly identified and fused within the same coordinate frame.

[0036] The building structure recognition unit employs a pre-trained semantic segmentation network to perform pixel-level semantic segmentation on the registered building images. The semantic segmentation network used in this unit is based on DeepLabV3+ or UNet architecture. The encoder uses ResNet or MobileNet as the backbone network to extract multi-scale features, while the decoder restores spatial resolution through upsampling and skip connections. The final output is a semantic segmentation map of the same size as the input image, with each pixel classified into one of the following categories: wall, floor, doors and windows, living room, bedroom, kitchen, or bathroom. By statistically analyzing the regional range of pixels for each functional room category and combining this with standard-sized reference objects in the image, the pixel-to-actual-size is calibrated. Specifically, the pixel size of the reference object in the image is identified, and the pixel / actual-size conversion ratio is calculated based on the actual physical size of the reference object to estimate the area of ​​each functional room. Simultaneously, the building structure type is identified by analyzing the building outline and the layout of internal partitions. This unit also estimates the span, depth, and floor height of a house using the triangulation principle of multi-view images: selecting the pixel coordinates of the same feature point in images taken from different angles, combining the camera's intrinsic parameters and the geometric relationship between images, and calculating the three-dimensional spatial coordinates of the feature point using the forward intersection method, thereby obtaining the three-dimensional size information of the house.

[0037] The interior decoration detail recognition unit uses a fine-grained image classification model to automatically identify the interior decoration details of a house. Fine-grained image classification is a classification task for subcategories within the same broad category where differences are subtle. This unit uses a model based on ResNet or EfficientNet, employing an attention mechanism to focus on discriminative regions in the image, achieving accurate classification of floor materials, wall materials, and ceiling materials. For decoration level assessment, this unit comprehensively judges the brand characteristics of the materials, construction techniques, and overall visual effect, classifying decoration levels into four levels: high-end, mid-range, ordinary, and simple. Each recognition result is accompanied by a class probability value output by the Softmax function as a confidence score.

[0038] The facility and equipment identification unit employs object detection algorithms to detect and identify various facilities and equipment in building images. The object detection model used by this unit is based on YOLOv8 or Faster R-CNN architecture. Through end-to-end training, it can simultaneously detect multiple target objects in a single image and output bounding box coordinates and a category label for each detected target. The facilities and equipment detected by this unit include: doors and windows, kitchen and bathroom facilities, and water and electricity facilities. For each detected facility and equipment, the unit not only identifies its category but also estimates its actual size by combining the bounding box pixel size with the aforementioned pixel / actual size conversion ratio. The number of facilities is obtained directly by counting the number of detected target instances. All identification results are summarized into structured facility and equipment survey data output.

[0039] In a further embodiment of the present invention, the land survey intelligent identification module includes: The land type identification unit classifies land use types in the land parcel image using a pre-trained deep learning classification model. The land use types include cultivated land, orchard, forest land, grassland, construction land, and water area. The ground attachment identification unit uses a combination of target detection and semantic segmentation to automatically identify and locate retaining walls, ditches, roads, culverts, bridges and drying yards in the image, and estimate their length, width, height, cross-sectional shape and lining material type; The tree identification and statistics unit uses a fine-grained image classification model to identify tree species, performs tree-by-tree detection and counting through a target detection algorithm, and estimates the diameter at breast height (DBH) range of trees by calibrating reference objects in the image.

[0040] In this embodiment, the land type identification unit classifies land use types in plot images using a pre-trained deep learning classification model. The unit receives panoramic images of the plot taken by investigators at key points along the plot boundaries, as well as internal images of the plot taken from different angles. The classification model is based on a deep convolutional neural network architecture such as ResNet or EfficientNet and is pre-trained on a dataset containing hundreds of thousands of remote sensing images and ground photographs labeled with land use types. The model extracts features such as texture, color, shape, and spatial layout from the images through multi-layer convolution and pooling operations, and finally outputs probability distributions for six land use types through fully connected layers and a Softmax classifier: cultivated land, orchard, woodland, grassland, construction land, and water area. The category with the highest probability value output by the model is the identification result, and this probability value is also used as the identification confidence score. For plots with ambiguous boundaries or mixed types, the unit can comprehensively utilize the combined judgment results of multiple images to output the most likely land type and its confidence score.

[0041] The ground attachment identification unit employs a combination of object detection and semantic segmentation to identify and measure various ground attachments in images. First, the unit uses an object detection model to locate and initially classify the various attachments in the image. Then, for attachments with regular geometric shapes, such as retaining walls, ditches, and roads, the semantic segmentation model is used to perform pixel-level precise segmentation of their contours. The identified attachment types include: retaining walls, ditches, roads, culverts, bridges, and drying yards. Based on the identification, the unit estimates the size parameters of the attachments using the following methods: pixel / actual size calibration using reference objects in the image; for length parameters, calculation based on the extension range of the pixel region of the attachment category in the semantic segmentation results combined with the calibration ratio; for cross-sectional parameters such as width and height, estimation based on detecting the horizontal and vertical pixel span of the attachment in the image combined with the calibration ratio; for cross-sectional shape, judgment based on analyzing the contour morphology of the attachment's cross-section in the semantic segmentation results; and for lining material type, identification of the attachment's surface material using a fine-grained image classification model. All identification and estimation results are accompanied by confidence values.

[0042] The tree identification and statistics unit is responsible for identifying tree species, counting each tree individually, and estimating diameter at breast height (DBH) for trees within the plot. This unit first uses a fine-grained image classification model to identify tree species. The model uses crown morphology, leaf shape, bark texture, and flower and fruit characteristics as criteria, and can identify species including, but not limited to, timber species, economic forest species, and landscaping species. Based on species identification, this unit uses an object detection algorithm to detect and count each tree in the image. The detection model is specially trained to simultaneously locate the trunk or crown of each tree in an image containing multiple trees and output bounding boxes. The number of trees is directly obtained by counting the instances of the detected bounding boxes. For DBH estimation, this unit uses the following method: identifying reference objects placed next to the tree trunks in the image, and calculating the estimated DBH based on the ratio of the reference object's pixel size to the tree trunk diameter's pixel size, combined with the actual size of the reference object. For multiple images of the same tree taken from different angles, this unit averages the estimation results from multiple angles to improve accuracy. All identification results are aggregated into a structured tree survey data output.

[0043] In a further embodiment of the present invention, the indicator calculation and report generation module includes: The indicator data integration unit maps the original identification results output by each identification module into standardized survey indicator items through a preset survey indicator mapping table. Each standardized survey indicator item includes indicator ID, indicator name, indicator attribute, indicator value and corresponding confidence level. The distributed computing unit divides the large-scale immigration survey task into pieces according to administrative region, land parcel number or household head number, and distributes them to multiple computing nodes to perform index calculation and summarization in parallel. Each computing node receives the task through a Redis message queue and writes the calculation results to a shared database. The caching and reuse unit uses the MD5 value of the generated metric calculation as the cache key to cache the calculated metric results in Redis. For the corrected metric data, the data is stored in the correction and solidification table. Subsequent calculations will prioritize using the cached or solidified data. The report generation unit automatically generates household-specific reports for housing surveys, land-specific reports for land surveys, and summary statistical reports by administrative level based on the integrated standardized indicator data.

[0044] In this embodiment, the indicator data integration unit maps the raw identification results output by each identification module to standardized survey indicator items through a preset survey indicator mapping table. This mapping table is a structured configuration database containing the definitions of all standardized indicator items in the migration survey indicator system. Each indicator item includes an indicator ID, indicator name, indicator attribute, and the corresponding mapping rule for the raw identification result. When the indicator data integration unit receives the raw identification results output by the housing survey intelligent identification module and the land survey intelligent identification module, it iterates through each mapping rule in the mapping table, converting the field names and values ​​in the raw identification results into standardized indicator items one by one.

[0045] The distributed computing unit is responsible for allocating large-scale immigration survey tasks to multiple computing nodes for parallel execution of indicator calculation and aggregation. This unit first segments the entire survey task according to administrative region, land parcel number, or household head number, breaking down the large-scale task into multiple independent sub-tasks. Each sub-task contains standardized indicator data for one or more survey subjects, along with the computational operations to be performed on that data. Task allocation employs a Redis-based distributed task queue mechanism: each computing node registers its node information with the Redis central server and sets an expiration time upon startup, and periodically sends heartbeat data to Redis through a local heartbeat mechanism to maintain its liveness; during task allocation, the actual number of tasks to be processed is modulo the number of currently active nodes to calculate the number of tasks each node needs to process, and then the tasks are pushed to each computing node through the Redis message queue. Each computing node pulls tasks from the message queue, performs indicator calculation and aggregation, and writes the calculation results to a shared database. After the calculation is complete, each node reports the task completion status to Redis.

[0046] The caching and reuse unit avoids redundant calculations through a caching mechanism. This unit generates a unique MD5 value for each metric calculation as the cache key. The MD5 value calculation is based on factors such as the identifier of the survey object, the list of metrics to be calculated, and the version number of the original data it relies on. Once a metric calculation is complete, the result is stored in a Redis cache as a key-value pair. Subsequent identical calculation requests directly return the cached result without recalculation. For metric data that has been manually reviewed and corrected, this unit stores it in a dedicated correction-stabilized table and marks it as "stabilized." Subsequent metric calculations and report generation prioritize using cached data; if the data is not present in the cache, it uses the data from the correction-stabilized table; recalculation is only triggered when neither is present. This mechanism ensures that corrected data can be correctly referenced in subsequent calculations.

[0047] The report generation unit automatically generates multi-level immigration survey reports based on integrated standardized indicator data. Report types include three categories: Household-by-household survey reports, listing all survey indicators such as house structure, room area, decoration details, and facilities; Plot-by-plot land survey reports, listing all survey indicators such as land type, list and dimensions of attachments, list and diameter at breast height of trees, etc.; and summary statistical reports by administrative level, summarizing key indicators such as total building area, land area of ​​various types, number of attachments of various types, and total number of trees at each administrative level (province, city, county, township, village). The report generation unit includes multiple report templates, allowing users to select the output format according to their needs. All reports automatically add page numbers, generation date, report number, and other metadata to ensure standardization and traceability.

[0048] In a further embodiment of the present invention, in the distributed computing unit, each computing node registers node information with a preset Redis center and sets an expiration time when it starts up, and sends heartbeat data to Redis periodically through a local heartbeat mechanism to keep it alive; when allocating tasks, the actual number of tasks to be processed is modulo the number of currently alive nodes to calculate the number of tasks that each node needs to process.

[0049] In this embodiment, the node registration and discovery mechanism is as follows: each computing node automatically performs a registration operation with a preset Redis center upon startup. If a node stops sending heartbeats due to a failure, network interruption, or normal shutdown, its registration key will be automatically deleted by Redis after the expiration time, and the node will be automatically removed from the "live nodes" list.

[0050] Task Allocation Algorithm: When task allocation is required, the task scheduling center first retrieves a list of all currently active nodes from Redis, and then counts the actual number of tasks to be processed. Task allocation uses a modulo algorithm: the list of tasks to be processed is sorted by task ID or task index, and then modulo N is taken. Specifically, for the i-th task, it is assigned to the node with the specified number. The advantages of this allocation method are: first, it eliminates the need for complex load balancing algorithms, resulting in minimal computational overhead; second, when the number of tasks is much greater than the number of nodes, the number of tasks allocated to each node is basically balanced; third, when the number of nodes changes, only the modulo allocation needs to be recalculated, giving the algorithm excellent dynamic adaptability. The specific execution method of task allocation is as follows: the task scheduling center encapsulates each task as a message and pushes it to the task queue of the corresponding node using the Redis LPUSH or RPUSH command. Each node then uses the Redis BLPOP or BRPOP command to pull tasks from its own queue in a blocking manner and execute them. After the task is completed, the node writes the result to the shared database and reports the completion status to the task scheduling center.

[0051] Fault recovery mechanism: When a computing node fails during task execution, its registration key automatically expires due to heartbeat cessation, and the node is removed from the live list. Upon periodic checks and detecting that the node is offline, the task scheduling center returns any unprocessed tasks in the node's queue to the pending task pool, awaiting reallocation. For assigned tasks that have not yet reported completion, the task scheduling center identifies and reallocates them using a timeout detection mechanism. This design ensures that even if some nodes fail, the entire computing task can still be completed eventually.

[0052] In a further embodiment of the present invention, the manual review and correction module includes: The confidence level labeling and priority sorting unit divides the identification results into three levels: high confidence, medium confidence, and low confidence, based on the identification confidence level, and generates a queue of manual review tasks according to the principle of prioritizing low confidence and key indicators. The difference comparison and annotation unit compares and overlays the intelligent recognition results with the original image pixel by pixel, highlights the areas on the image where the recognition results are questionable, and displays the recognition labels for reviewers to compare. The manual correction interface unit provides a correction interface with drop-down menus and input boxes, allowing auditors to modify the values ​​of indicators that are incorrectly identified and manually add annotations of missed objects to images. It also automatically records the person making the correction, the correction time, and an audit log comparing the correction before and after.

[0053] In this embodiment, the confidence level labeling and priority sorting unit is responsible for classifying the intelligent recognition results and generating a review task queue. This unit reads the confidence level value associated with each recognition result and categorizes them into three levels according to a preset threshold: results with a confidence level ≥ 0.9 are labeled "high confidence," indicating that the model has a high degree of certainty about the recognition result and it is automatically archived without manual review; results with a confidence level ≤ 0.7 and < 0.9 are labeled "medium confidence," indicating that the model has some certainty but there is uncertainty, and a sampling review method is used; results with a confidence level < 0.7 are labeled "low confidence," indicating that the model has a low degree of certainty and each result requires manual review. When generating the manual review task queue, this unit sorts the results according to the principle of "low confidence priority, key indicators priority," meaning that all low-confidence recognition results are placed at the top of the queue, and within the low-confidence category, they are further sorted according to the importance of the indicators. Through this classification and sorting strategy, the reviewers' workload is prioritized for the recognition results that most require manual judgment, achieving optimal allocation of review resources.

[0054] The difference comparison and annotation unit provides auditors with an intuitive review interface. When an auditor opens a recognition result to be reviewed, this unit displays two views side-by-side in the review interface: the original captured image on the left and a visual overlay of the intelligent recognition result on the right. In the visual overlay, the unit overlays the recognition result onto the original image as a semi-transparent color block or bounding box, and uses different colors to distinguish different recognition categories. For recognition areas with low confidence, the unit highlights them on the image to prompt auditors to pay close attention. Simultaneously, the unit displays all recognition labels in a list format in the sidebar. Auditors can click on any label, and the interface automatically locates and zooms in on the corresponding recognition area in the image. This intuitive comparison and location method allows auditors to quickly determine whether the recognition result is correct and whether there are any omissions or misidentifications.

[0055] The manual correction interface unit provides auditors with convenient correction tools. The correction interface includes the following interactive elements: for classification and identification results, a drop-down menu allows auditors to select the correct category from a list of standard options; for numerical identification results, an input box allows auditors to directly modify the values; for missed identifications, a tool is provided to manually draw bounding boxes or annotation points on the image, allowing auditors to use a mouse or stylus to select the missed objects and choose their category labels. All correction operations are automatically recorded in the audit log, which includes: the person making the correction, the correction time, the original identification result before correction, the result after correction, the correction type, and the basis for correction. The audit log is stored in a dedicated log table, is tamper-proof, and provides a complete record of operations for subsequent quality traceability and accountability.

[0056] In a further embodiment of the present invention, the pre-trained deep learning models in the intelligent identification module for housing and land surveys have training samples containing labeled immigration survey image data. The models output the probability value of the corresponding category as the identification confidence level while outputting the identification result.

[0057] In this embodiment, the training data sources and annotations are as follows: The training data for each recognition model comes from two sources. First, archived image data accumulated from historical immigration survey projects. These images have been manually measured and recorded using traditional survey methods and can be used as weakly labeled data. Second, image data that has been confirmed and corrected by the manual review and correction module during system operation. This data has accurate pixel-level annotations or bounding box annotations and accurate category labels, making it a high-quality training sample. All training data undergoes standardized annotation processing by a professional annotation team: For semantic segmentation tasks, annotators use pixel-level annotation tools to assign category labels to each pixel in the image; for object detection tasks, annotators use bounding box annotation tools to outline each target object and assign category labels; for image classification tasks, annotators assign category labels to the entire image. The annotated data is stored in the training dataset after quality checks.

[0058] Model Training Process: Each recognition model employs a two-stage training strategy of pre-training + fine-tuning. In the first stage, the model is pre-trained on a general large-scale dataset to learn general visual feature extraction capabilities. In the second stage, the pre-trained model is fine-tuned on an immigration survey image training dataset. Specifically, pre-trained weights are loaded, the last layer or the last few layers of the model are replaced with new layers adapted to the immigration survey task, and then several rounds of training are continued on the immigration survey dataset with a small learning rate until the model's performance on the validation set no longer improves. During training, data augmentation techniques are used to expand the diversity of training samples and improve the model's generalization ability. After training, the model performance is evaluated on an independent test set, with evaluation metrics including accuracy, precision, recall, and F1 score. Only when the model performance reaches a preset release threshold will it be deployed to the production environment.

[0059] Confidence Output Mechanism: Each recognition model outputs the probability value of the corresponding category along with the recognition result, serving as the recognition confidence score. For classification tasks, the fully connected layer of the model is followed by a Softmax activation function, outputting a probability distribution vector. Each element represents the probability that the input image belongs to the corresponding category. The category with the highest probability value is taken as the recognition result, and this maximum probability value is the confidence score. For object detection tasks, the model outputs a category probability value and a location confidence score for each detected object. The confidence score of the recognition result is obtained by combining both or by taking the category probability value. For semantic segmentation tasks, the model outputs a category probability distribution for each pixel. The recognition confidence score for that pixel is the maximum category probability value, while the confidence score of the entire recognition result is the average or minimum of the confidence scores of all pixels in that region. All confidence scores are in the range of 0 to 1, with values ​​closer to 1 indicating a higher level of confidence in the recognition result.

[0060] In a further embodiment of the present invention, the manual review and correction module further includes a feedback learning unit, which is used to periodically add the manually corrected data to the training dataset of each recognition model, perform incremental training or fine-tuning of the model, track changes in accuracy and recall through model performance evaluation, and automatically release a new version of the model when the model performance reaches a preset release threshold.

[0061] In this embodiment, the data feedback and accumulation mechanism involves the feedback learning unit collaborating with other units in the manual review and correction module. After reviewers verify and correct the recognition results on the correction interface, all corrected data is stored in a dedicated correction database. The feedback learning unit periodically extracts new corrected data from the correction database, cleans and converts the data, and then adds it to the training dataset of each recognition model. For data with a correction type of "modify value," the feedback learning unit adds it to the training set as a category label correction sample, replacing the original incorrect label with the corrected one. For data with a correction type of "add annotation," the feedback learning unit adds it to the training set as a new annotation sample. To ensure the quality of the training data, the feedback learning unit also sets up a data filtering mechanism, adopting only corrected data independently confirmed by at least two reviewers, or corrected data with low confidence results below 0.7, to ensure the high reliability of the annotation data added to the training set.

[0062] Incremental Training and Model Fine-tuning: When the accumulated training data reaches a certain amount, the feedback learning unit automatically triggers the incremental training or fine-tuning process of the model. The specific method for incremental training is as follows: load the model weights deployed in the current production environment as initial weights, and perform several rounds of training on the newly added training data with a small learning rate. During training, continuous learning techniques such as elastic weight consolidation or knowledge distillation are employed to maintain the model's retention of old knowledge while it learns new knowledge, avoiding the "catastrophic forgetting" problem. After fine-tuning, the updated model's performance is evaluated on an independent validation set.

[0063] Model Performance Evaluation and Automated Deployment: The feedback learning unit incorporates a model performance evaluation and version management mechanism. After each incremental training or fine-tuning, the unit comprehensively evaluates the new model on a reserved test set. Evaluation metrics include: overall accuracy, precision and recall for each category, and the magnitude of change in each metric compared to the current production model. If all performance metrics of the new model meet or exceed the preset deployment threshold, the new model is automatically marked as a "candidate deployment version." Before deployment, the unit also performs A / B testing, running the new model and the current production model in parallel on a small portion of real-world data, comparing their recognition results and confidence distributions to ensure that the new model performs better than or equal to the current model in real-world application scenarios. After passing the A / B test, the new model is automatically deployed as the new version of the production model, replacing the currently deployed model version. The entire process, from data accumulation, incremental training, performance evaluation to automated deployment, is fully automated and requires no manual intervention.

[0064] In a further embodiment of the present invention, the image acquisition module, the intelligent identification module for human and housing surveys, the intelligent identification module for land surveys, the indicator calculation and report generation module, and the manual review and correction module interact with each other via a cloud server, wherein the image acquisition module is deployed on a mobile terminal, and the other modules are deployed on a cloud server or a local server.

[0065] In this embodiment, the mobile terminal is deployed as follows: The image acquisition module is deployed on a mobile terminal device carried by the investigator, including but not limited to smartphones and dedicated tablets. An image acquisition application runs on the mobile terminal. This application utilizes the terminal's camera hardware to acquire images, uses the GPS module to obtain geographic coordinates, and uses the gyroscope and accelerometer to obtain shooting angle information. The image acquisition application has offline working capabilities. In remote survey areas with poor network signals, the acquired images and metadata can be temporarily stored in the mobile terminal's local storage and uploaded to the cloud server in batches after the network is restored. The mobile terminal interacts with the cloud server via a mobile communication network or a WiFi network.

[0066] Deployment on cloud servers or local servers: The intelligent identification module for housing and population surveys, the intelligent identification module for land surveys, the indicator calculation and report generation module, and the manual review and correction module are deployed on either cloud servers or local servers. The cloud server deployment option is suitable for large-scale migration survey projects with high requirements for elasticity, scalability, and high availability. It can utilize GPU instances provided by cloud service providers for deep learning model inference, cloud databases for data storage, and cloud storage services for image file storage and management. The local server deployment option is suitable for projects with high requirements for data security and network independence. All modules are deployed on the survey unit's own server cluster, and data flows entirely within the internal network, never traversing the public internet. Both deployment options are functionally identical, and users can choose according to their actual needs.

[0067] Inter-module data interaction: Modules interact with each other via a cloud server. The specific data flow is as follows: The image acquisition module uploads the acquired images and associated metadata to the cloud server's image storage service via HTTPS, and simultaneously writes the image metadata into the image record table of the database. The intelligent identification modules for population and housing surveys and land surveys read the images to be identified from the image storage service, retrieve the corresponding metadata from the database, and write the identification results into the identification result table of the database after completion. The indicator calculation and report generation module reads the identification results from the database, performs indicator calculation and summarization, writes the calculation results into the indicator result table of the database, and generates a report stored in the file storage service. The manual review and correction module reads the identification results to be reviewed from the database. Reviewers access the module's web interface through a browser to complete the review and correction operations, and the correction results are written back to the database. All data interactions between modules are secured through authentication and encryption mechanisms to ensure data transmission security.

[0068] The system deployment features scalability design: the cloud server deployment scheme supports elastic scaling of computing nodes. During peak survey periods, the system can automatically or manually increase the number of GPU computing nodes to handle the increased recognition task load; during off-peak periods, the number of computing nodes can be reduced to save costs. The image storage service adopts an object storage scheme, which can automatically expand storage capacity as the data volume grows. The database service uses master-slave replication or clustering schemes to ensure high data availability and read / write performance.

[0069] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. An intelligent measurement system based on immigration index survey, characterized in that, include: The image acquisition module collects image data of the survey subjects at the immigration survey site through mobile terminal devices, performs real-time quality assessment of the collected images, and associates and binds qualified images with the survey subject information and geographical coordinates. The intelligent identification module for human and housing surveys preprocesses and registers the house images acquired by the image acquisition module, and uses a pre-trained deep learning model to identify the house structure, decoration details, and facilities and equipment, and outputs human and housing survey index data and its identification confidence level. The land survey intelligent identification module preprocesses the land and surface attachment images acquired by the image acquisition module, and uses a pre-trained deep learning model to identify land type, surface attachments, and trees, and outputs land survey index data and its identification confidence level. The indicator calculation and report generation module standardizes and integrates the indicator data output by the intelligent identification module for housing and population survey and the intelligent identification module for land survey according to a unified survey indicator system, performs indicator calculation and summarization, and generates a structured immigration survey report. The manual review and correction module classifies the intelligent recognition results according to the recognition confidence level, generates a manual review task queue, provides a correction interface for reviewers to review and correct the recognition results, and records the correction operation log.

2. The intelligent measurement system based on immigration index survey according to claim 1, characterized in that, The image acquisition module includes: The data collection planning unit provides differentiated shooting guidelines based on the type of survey task. In the human and housing survey mode, it guides the shooting in the order of the overall appearance of the house, each functional room, decoration details, and facilities and equipment. In the land survey mode, it guides the shooting in the order of the plot panorama, above-ground attachments, and individual trees. The image quality assessment unit performs real-time sharpness and illumination uniformity detection on the acquired images using the Laplacian variance or BRISQUE no-reference image quality assessment method, and prompts the user to retake images that do not meet the quality standards. The spatial positioning and association unit records the geographic coordinates, shooting time, and shooting angle of each image, and associates and binds the images with the corresponding survey objects.

3. The intelligent measurement system based on immigration index survey according to claim 1, characterized in that, The intelligent identification module for the human-room survey includes: The image preprocessing and registration unit performs grayscale conversion, noise reduction, and enhancement processing on the building images, and uses an image registration algorithm to spatially align images taken from different angles or under different conditions. The building structure recognition unit uses a pre-trained semantic segmentation network to perform pixel-level semantic segmentation on building images, identify the building structure type, the boundaries and areas of each functional room, and use standard size reference objects in the image to calibrate the pixels with the actual size, and estimate the building's span, depth and floor height. The decoration detail identification unit uses a fine-grained image classification model to automatically identify and evaluate the material type and decoration level of the floor, walls, and ceiling inside the house. The material type includes tiles, wood flooring, latex paint, wallpaper, and suspended ceiling, and the decoration level includes high-end, mid-range, ordinary, and simple. The facility and equipment identification unit uses target detection algorithms to detect, locate, and classify doors, windows, kitchen and bathroom facilities, and water and electricity facilities in building images, and estimates their quantity and size.

4. The intelligent measurement system based on immigration index survey according to claim 1, characterized in that, The land survey intelligent recognition module includes: The land type identification unit classifies land use types in the land parcel image using a pre-trained deep learning classification model. The land use types include cultivated land, orchard, forest land, grassland, construction land, and water area. The ground attachment identification unit uses a combination of target detection and semantic segmentation to automatically identify and locate retaining walls, ditches, roads, culverts, bridges and drying yards in the image, and estimate their length, width, height, cross-sectional shape and lining material type; The tree identification and statistics unit uses a fine-grained image classification model to identify tree species, performs tree-by-tree detection and counting through a target detection algorithm, and estimates the diameter at breast height (DBH) range of trees by calibrating reference objects in the image.

5. The intelligent measurement system based on immigration index survey according to claim 1, characterized in that, The indicator calculation and report generation module includes: The indicator data integration unit maps the original identification results output by each identification module into standardized survey indicator items through a preset survey indicator mapping table. Each standardized survey indicator item includes indicator ID, indicator name, indicator attribute, indicator value and corresponding confidence level. The distributed computing unit divides the large-scale immigration survey task into pieces according to administrative region, land parcel number or household head number, and distributes them to multiple computing nodes to perform index calculation and summarization in parallel. Each computing node receives the task through a Redis message queue and writes the calculation results to a shared database. The caching and reuse unit uses the MD5 value of the generated metric calculation as the cache key to cache the calculated metric results in Redis. For the corrected metric data, the data is stored in the correction and solidification table. Subsequent calculations will prioritize using the cached or solidified data. The report generation unit automatically generates household-specific reports for housing surveys, land-specific reports for land surveys, and summary statistical reports by administrative level based on the integrated standardized indicator data.

6. The intelligent measurement system based on immigration index survey according to claim 5, characterized in that, In the distributed computing unit, each computing node registers its node information with a preset Redis center and sets an expiration time when it starts up. It also sends heartbeat data to Redis periodically through a local heartbeat mechanism to keep itself alive. When allocating tasks, the actual number of tasks to be processed is modulo the number of currently alive nodes to calculate the number of tasks that each node needs to process.

7. The intelligent measurement system based on immigration index survey according to claim 1, characterized in that, The manual review and correction module includes: The confidence level labeling and priority sorting unit divides the identification results into three levels: high confidence, medium confidence, and low confidence, based on the identification confidence level, and generates a queue of manual review tasks according to the principle of prioritizing low confidence and key indicators. The difference comparison and annotation unit compares and overlays the intelligent recognition results with the original image pixel by pixel, highlights the areas on the image where the recognition results are questionable, and displays the recognition labels for reviewers to compare. The manual correction interface unit provides a correction interface with drop-down menus and input boxes, allowing auditors to modify the values ​​of indicators that are incorrectly identified and manually add annotations of missed objects to images. It also automatically records the person making the correction, the correction time, and an audit log comparing the correction before and after.

8. The intelligent measurement system based on immigration index survey of claim 1, wherein, The pre-trained deep learning models in the intelligent identification module for housing and land surveys are trained using labeled immigration survey image data. The models output the probability value of the corresponding category as the identification confidence level while outputting the identification result.

9. The intelligent measurement system based on immigration index survey according to claim 1, characterized in that, The manual review and correction module also includes a feedback learning unit, which is used to periodically add the manually corrected data to the training dataset of each recognition model, perform incremental training or fine-tuning of the model, and track changes in accuracy and recall through model performance evaluation. When the model performance reaches the preset release threshold, a new version of the model is automatically released.

10. The intelligent measurement system based on immigration index survey according to claim 1, characterized in that, The image acquisition module, the intelligent identification module for human and housing surveys, the intelligent identification module for land surveys, the indicator calculation and report generation module, and the manual review and correction module interact with each other via a cloud server. The image acquisition module is deployed on a mobile terminal, while the other modules are deployed on a cloud server or a local server.