CN112766417A

CN112766417A - Method and system for recognizing current land type of land utilization of target land block by using field photo

Info

Publication number: CN112766417A
Application number: CN202110205018.5A
Authority: CN
Inventors: 袁锦秀; 于艺海; 战立杰
Original assignee: Individual
Current assignee: Individual
Priority date: 2021-02-26
Filing date: 2021-02-26
Publication date: 2021-05-07

Abstract

The invention discloses a method and a system for recognizing the current land utilization status of a target land block by using a scene photo, wherein the method comprises the following steps: the method comprises the steps of generating identification parameters for determining the boundary of a photo target land parcel and identification parameters of land categories based on image data, extracting features to be identified and features to be identified of the image boundary of a photo based on the boundary identification parameters and the land category identification parameters, respectively matching the features to be identified and the features to be identified of the land category with a boundary feature library and a land category feature library, determining an image target land parcel area according to the matching result of the boundary features, and determining the land category of the image area to be identified as a preset land category when the matching degree of the features to be identified of the image data target land parcel area and any image features in all image features of the preset land categories is larger than or equal to a matching threshold value. And determining the land types of all the single photos of the target land parcel one by one, merging and calculating, and determining the land type of the target land parcel according to the minimum control area of the land type.

Description

Method and system for recognizing current land type of land utilization of target land block by using field photo

Technical Field

The invention relates to the technical field of image data application, in particular to a method and a system for recognizing a land type of a target land utilization status by using a scene photo.

Background

Convolutional neural networks are the most widely used deep neural network structures at present. VGG is a series of convolution neural network models proposed by Oxford Visual Geometry Group in 2014, and is characterized in that a stack of 3x3 small convolution kernels is adopted to replace a larger convolution kernel (11 x11, 7x7 and 5x 5) in a traditional AlexNet convolution network, the depth of the network is improved under the condition of the same receptive field, and therefore a more complex mode can be learned. The VGG neural network model is applied to the fields of face recognition, image classification and the like. The traditional picture classification generally needs three processes of bottom layer feature extraction, feature coding and classifier design, and the image classification method based on deep learning can replace manual design or image feature selection in the traditional picture classification through learning hierarchical feature representation, so that the accuracy rate of the complex natural scene picture classification is greatly improved. The deep learning based picture classification has wide application in the field of mobile internet, but mainly involves the identification of objects of the same type.

No relevant report exists at present for image identification and land type confirmation of land utilization status classification based on technical regulations (TD/T1055-2019) published by the ministry of natural resources for national soil survey for the third time. The third national homeland survey in the regulations is classified as follows:

the third national soil survey work classification is in contrast relation with the three categories of the national people's republic of China land management Law as follows:

the traditional algorithms for prior art target detection are: Haar/LBP/integral HOG/ACF feature + Adaboost cascade classifier, HOG + SVM, discrete trained reconstructed flexible part models (DPM), template matching, etc. The mainstream algorithms for target detection implemented by deep learning in recent years are divided into two types: a two-stage process and a one-stage process. Ssd (single Shot detector) is a one-stage method, and the main idea is to perform dense sampling uniformly at different positions of a picture, and to use different scales and aspect ratios during sampling, and then to perform classification and regression directly after using CNN to extract features. The SSD300 model is an implementation of the SSD method, with input pictures with a resolution of 300 × 300. Current SSD300 implementations that detect land boundaries and present land classes separately are less. The result of object detection typically includes category information and location information of the object. The categories are marked by letters or numbers; the position is represented by the coordinates of the center point of the circumscribed rectangular frame of the target area and the width and height.

Modern photogrammetry techniques are currently mainly applied to aerial and aerospace remote sensing images, while traditional photogrammetry is a scientific technique which utilizes an optical camera to take pictures and researches and determines the shape, size, position and mutual relation of the shot objects through the pictures, and the traditional photogrammetry techniques are mainly used for measuring topographic maps in the early stage.

The prior art has difficulty in identifying natural scene photos, and most of the current deep learning picture identification applications are based on public data sets or identification based on specific scenes and specific samples. The recognition of the scene picture of the natural scene has certain difficulty due to the fact that the type of ground objects possibly contained in the picture and the variable factors such as the illumination, the size, the resolution, the shooting posture and the like of the picture are more. In the third national homeland survey in the prior art, the photos for field evidence presentation are generally judged and read manually during the subsequent checking process, so that the checking efficiency is low, careless and misjudgment are easy to occur due to subjective judgment difference, the checking cost is high, and the data quality risk is high. In addition, in the manual interpretation process, there are some difficulties that the area cannot be accurately calculated by manually interpreting photos, and for the plot plots with too large comprehensive multiple plots, whether a certain plot exceeds the minimum upper plot area cannot be judged through the photos, so that image measurement and calculation or field operation proofing again are often needed, and the cost is increased.

Therefore, there is a need for a technique for determining a target parcel class using live photographs based on convolutional neural networks and photogrammetry.

Disclosure of Invention

The technical scheme of the invention provides a method and a system for identifying a target land parcel by using a live photo, which aim to solve the problem of how to confirm the target land parcel according to a photo image and related data.

In order to solve the above problems, the present invention provides a method for recognizing a target parcel type using a live photograph, the method comprising:

generating identification parameters for determining the boundaries of target plots of a single photo and the identification parameters of the land types based on the image data;

extracting image boundary characteristics and land characteristics of each image data in a plurality of field photo image data through a characteristic extraction model, and establishing an image boundary characteristic library comprising all image boundary characteristics of the plurality of photo image data and a land characteristic library comprising all image land characteristics;

extracting the image boundary to-be-identified features and the land to-be-identified features of the single photo image data to be identified based on the boundary identification parameters and the land identification parameters;

matching the features to be recognized of the image boundary and the features to be recognized of the land type with boundary features and land type features in a boundary feature library and a land type feature library respectively, and determining the image area as a target land area in image data when the matching degree of the features to be recognized of the image boundary and any image boundary features in all image edge features of preset categories in the image boundary feature library is greater than or equal to a matching threshold value, and the shape similarity of the image area extracted by the features to be recognized of the image boundary after being corrected and a global or local graphic image of the position and the orientation of the target land vector graphic is greater than or equal to the matching threshold value; when the matching degree of all image boundary features to be recognized of the photo and all image boundary features in the image boundary feature library is smaller than a preset threshold value, or the shape similarity between an image area extracted by the image boundary features to be recognized and a global or local graphic image of the position and the direction of a target plot vector graphic after correction is smaller than a preset threshold value, determining all areas of the image data to be recognized as image data target plot areas; when the matching degree of the features to be identified of the image data target land area land class and any image features in all image features of a preset land class is larger than or equal to a matching threshold value, determining the land class of the image area to be identified as the preset land class;

determining land types of all single photo target land areas of a target land one by one;

and merging and calculating the area of the land types of all the single photo target land areas of the target land, and determining the land type of the target land according to a preset land type minimum control area rule.

Preferably, when the matching degrees of all the image land features to be recognized of the photo and all the image land features in the land feature library are smaller than a preset threshold, the image data to be recognized is determined as image data of an irrelevant category, and the photo is determined as a photograph irrelevant to the land.

Preferably, before the generating of the identification parameters for determining the boundary of the target parcel and the feature of the land category in the single photo based on the image data and extracting the feature to be identified of the image boundary and the feature to be identified of the land category of the image data to be identified based on the identification parameters, the generating further comprises:

designing a data model, and storing an original photo, shooting information and target block vector data to a specific physical storage position according to the data model;

extracting shooting information such as a shooting azimuth angle, a shooting point geographic coordinate, a shooting pitch angle and the like from the picture, storing the shooting information into a shooting information data table which can be retrieved according to the unique identification code of the picture, and storing the shooting information data table in a storage position corresponding to a specific structure;

storing the vector graphic data of the target parcel as a vector graphic data set of the target parcel which can be retrieved according to the unique identification code of the parcel according to a preset data structure, and storing the vector graphic data of the target parcel in a storage position corresponding to a specific structure;

storing the original photo as a photo file data set which can be retrieved according to the unique identification code of the photo and the unique identification code of the land block according to a preset directory structure, and storing the photo file data set in a storage position corresponding to a specific directory;

carrying out image signature on the original photo, determining a repeated original photo by comparing the image signatures of the original photo, and removing the repeated original photo to keep a single original photo;

performing multiple quality checks on image data of a single picture, corresponding shooting information and vector graphics data of a target plot, and removing the single picture if the image data of the single picture is damaged or unclear; if the shooting information and the vector graphic data of the target plot corresponding to the single picture are incomplete, removing the single picture; if the single picture is lack of correlation with the corresponding shooting information and the vector graphic data of the target plot, correcting the correlation information; and if the image data of the single picture is not damaged and clear, and the shooting information corresponding to the single picture, the vector graphic data of the target land parcel and the incidence relation are complete, respectively storing the single picture, the shooting information and the vector graphic data of the target land parcel in the preset storage positions. Preferably, the ground class includes:

three categories: agricultural land, construction land and unused land;

first-stage classification: wetlands, cultivated lands, plantation lands, woodlands, grasslands, commercial service lands, industrial and mining lands, residential lands, public management and public service lands, special lands, transportation lands, water and water conservancy facilities lands, and other lands;

and (4) secondary classification: mangrove land, forest marsh, bush marsh, marsh grass land, salt field, coastal beach, inland beach, marsh land, paddy field, irrigated land, dry land, orchard, tea garden, rubber garden, other gardens, arbor forest land, bamboo forest land, bush forest land, other forest land, natural pasture land, artificial pasture land, other grassland, logistics storage land, commercial service facility land, industrial land, mining land, residential land for cities and towns, residential land for villages, public facility land, park and green land, office community news publishing land, scientific and cultural land, special land, railway land, rail traffic land, public road land, town road land, traffic service land, rural road, airport land, port dock land, pipeline transportation land, river water surface, lake water surface, reservoir water surface, pond water surface, ditch, water construction land, river water surface, lake water surface, river water surface, pond water surface, ditch, water conservancy construction land, artificial pasture land, other grassland, artificial pasture land, other grass, Glaciers and permanent accumulated snow, vacant land, facility farming land, field ridge, saline-alkali land, sand land, bare rock gravel land.

In accordance with another aspect of the present invention, there is provided a method of generating an identification parameter for determining a boundary of a target parcel in a single photograph based on image data, the method comprising:

determining a picture, shooting information and vector graphic data of the target parcel according to the unique identification code of the target parcel, and extracting the image data in the picture file, the shooting information in the shooting information data table and the vector graphic of the vector graphic data set of the target parcel;

selecting boundary basic image data in the photo image data, determining boundary initial identification parameters based on the boundary basic image data, training the boundary initial identification parameters through training image data in the image data, and adjusting the boundary initial identification parameters according to output results of the initial identification parameters so as to generate boundary identification parameters to be optimized;

adjusting the output result according to a preset rule, taking the adjusted output result as input data, and performing cyclic training on the boundary identification parameter to be optimized until the output result of the boundary identification parameter to be optimized reaches a stable state;

when the output result reaching the stable state meets the preset requirement, stopping training the boundary identification parameter to be optimized, and taking the optimized boundary identification parameter as an identification parameter for determining the boundary of the target block based on the image data;

and storing the image boundary recognition result meeting the preset requirement as a boundary information data set which can be retrieved according to the unique identification code of the piece according to a pre-designed data model, and storing the boundary information data set in a storage position corresponding to the specific structure.

Preferably, the preset rule for adjusting the boundary recognition output result includes:

if the boundary identification output result is the ground linear segmentation ground object, the ground object is calibrated to be a natural linear ground object or an artificial linear ground object, and the artificial linear ground object is subdivided into a temporary type and a permanent type; and if the non-ground linear segmentation ground object is obtained through the boundary identification output result, the non-ground linear segmentation ground object is calibrated as a non-boundary characteristic.

Preferably, before determining the original photograph, the shooting information and the vector graphic of the target parcel and the image data file associated with the target parcel according to the unique identification code of the target parcel, and extracting the image data in the original photograph file, the shooting information in the shooting information data table, and the vector graphic of the vector graphic data set of the target parcel, further comprises:

according to a pre-designed data model, whether original pictures, shooting information and target region vector graphic image data are stored in a specific physical storage position is confirmed;

performing multiple quality checks on a single picture, corresponding shooting information and vector graphic data of a target plot, and removing the single picture if the image data of the single picture is damaged or unclear; if the shooting information and the vector graphic data of the target plot corresponding to the single picture are incomplete, removing the single picture; if the single picture is lack of correlation with the corresponding shooting information and the vector graphic data of the target plot, correcting the correlation information; and if the image data of the single picture is not damaged and clear, and the shooting information corresponding to the single picture, the vector graphic data of the target land parcel and the incidence relation are complete, respectively storing the single picture, the shooting information and the vector graphic data of the target land parcel in the preset storage positions.

In accordance with another aspect of the present invention, there is provided a method of generating identification parameters for determining a target parcel class for a single photograph based on image data, the method comprising:

determining a photo file and target parcel vector graphic data associated with a target parcel according to a unique identification code of the target parcel, and extracting image data in the photo file and target parcel expected parcel type data in the target parcel vector graphic data; determining a corresponding boundary information record according to the unique identification code of the photo, and extracting the boundary data;

selecting land type basic image data in the boundary of the photo image data target land block, determining land type initial identification parameters based on the land type basic image data, training the land type initial identification parameters through training image data in the image data, and adjusting the land type initial identification parameters according to the output result of the initial identification parameters so as to generate land type identification parameters to be optimized;

adjusting the output result according to the expected land type of the target land parcel, taking the adjusted output result as input data, and performing cyclic training on the land type identification parameter to be optimized until the output result of the obtained land type identification parameter to be optimized reaches a stable state;

and when the output result of the land types reaching the stable state is compared with the expected land types and the matching degree is greater than or equal to the matching threshold value, stopping training the land type identification parameters to be optimized, and taking the optimized land type identification parameters as the identification parameters for determining the target land type of the land parcel based on the image data.

Preferably, the determining of the photograph file and the target parcel vector graphics data associated with the target parcel according to the target parcel unique identification code and the extracting of the image data in the photograph file and the target parcel expected parcel class data in the target parcel vector graphics data; determining a corresponding boundary information record according to the unique identification code of the photo, and before extracting the boundary data, further comprising:

according to a pre-designed data model, whether the photo file, the boundary information and the vector graphic data of the target plot are stored in a specific physical storage position is confirmed;

performing multiple quality checks on the photo file, the corresponding boundary information and the vector graphic data of the target plot, and removing a single photo if the single photo, the corresponding boundary information or the vector graphic data of the target plot of the photo file is incomplete; if the single photo of the photo file is lack of association with the corresponding boundary information and the target plot vector graphic data, correcting the associated information; and if the single photo, the corresponding boundary information, the target plot vector graphic data and the incidence relation of the photo file are complete, respectively storing the single photo, the corresponding boundary information and the target plot vector graphic data in the preset storage positions.

Preferably, the selecting the basic image data in the image data further includes selecting the basic image data in the image data by a non-maximum suppression method.

Based on another aspect of the present invention, a method for merging and calculating the area of the land type of all the target land areas of the single photo target land areas and determining the target land type according to the land type minimum control area rule of the preset type is provided, wherein the method comprises the following steps:

extracting all image data of a single photo target land area of a target land and identifying land types;

when all image data of the target land block have one identification land class, determining the target land block class as the identification land class, and setting the area of a target land block class region as the real space area recorded by the target land block vector graph;

when all the image data of the target land have multiple types of identification land, shooting, correcting, de-duplicating and splicing all the image land areas;

when the spliced image completely covers the vector graph of the target land parcel, calculating the spatial resolution of the spliced image, and calculating the real spatial area corresponding to each land area according to the spatial resolution and the non-deformation pixel area of each land area of the spliced image;

when the spliced image cannot completely cover the vector graph of the target land parcel, calculating the real space area corresponding to each land area according to the preset spatial resolution;

when the target land parcel only has one land class and the real space area is smaller than the minimum control area of the land class, marking the target land parcel as a small micro-pattern spot;

when the target land parcel has multiple land types, comparing the real space area of each land type with the minimum control area of the land type, when the real space area of the land type is smaller than the minimum control area, marking the land type area as a small micro land type area, after the comparison of all the multiple land types is finished, combining the small micro land type area with the adjacent land type area, combining the land types into the land types of the adjacent land type area, and obtaining the real space area as the sum of the areas of the combined land type areas;

and finally determining the target land block as a multi-land-type land block when the small micro land-type areas of the target land block are combined and various land types exist, and recording the image areas and the real space areas of the various land types.

Preferably, the preset minimum control area rule of the land types comprises that the minimum control area of the construction land is 200 square meters; the minimum control area of the facility agricultural land is 200 square meters; the minimum control area of the agricultural land is 400 square meters; the minimum control area of unutilized is 600 square meters.

According to another aspect of the present invention, there is provided a system for recognizing a target parcel class using a live photograph, the system including:

the preprocessing module is used for preprocessing data before identification, and comprises: data access for storing original data such as original photograph, shot information and target block vector data into a specific physical storage location according to a pre-designed data model, including extracting shot information such as shot azimuth, shot point geographic coordinate and shot pitch angle from the original photograph, storing into a shot information data table which can be retrieved according to the unique identification code of the block, and storing into a storage location corresponding to the specific structure, storing target block vector graphic data into a target block vector graphic data set which can be retrieved according to the unique identification code of the block according to a pre-set data structure, and storing into a storage location corresponding to the specific structure, storing the original photograph into a photograph file data set which can be retrieved according to the unique identification code of the block and the unique identification code of the block according to a pre-set directory structure, and storing into a storage location corresponding to the specific directory, after multiple quality check is carried out on image data of a single picture, corresponding shooting information and vector graphic data of a target land parcel, if the image data of the single picture is not damaged and clear, and the shooting information corresponding to the single picture and the vector graphic data of the target land parcel and the incidence relation are complete, the single picture, the shooting information and the vector graphic data of the target land parcel are respectively stored in the preset storage positions, and when an image boundary identification output result meets the preset requirement, the image boundary identification output result is stored as a boundary information data set which can be searched according to a unique identification code of the picture according to a preset data model and is stored in the storage position corresponding to a specific structure; the data quality verification is used for carrying out image signature on the original photo, determining a repeated original photo by comparing the image signature of the original photo, and removing the repeated original photo to keep a single original photo; performing multiple quality checks on the image data of a single picture, corresponding shooting information and vector graphic data of a target plot to determine whether the single picture is clear and damaged and whether the information and the association are complete; data correction, which is used for removing the single picture if the image data of the single picture is damaged or unclear or the shooting information and the target plot vector graphic data corresponding to the single picture are incomplete; if the single picture is lack of correlation with the corresponding shooting information and the vector graphic data of the target plot, correcting the correlation information;

the feature library generating module is used for extracting image boundary features and land features of each image data in a plurality of field photo image data through a feature extraction model, and establishing an image boundary feature library comprising all image boundary features of the plurality of photo image data and a land feature library comprising all image land features;

the feature extraction module is used for extracting the features to be identified of the image boundary and the features to be identified of the land class of the single photo image data to be identified based on the boundary identification parameters and the land class identification parameters;

the feature identification module is used for respectively matching the features to be identified of the image boundary and the features to be identified of the land type with the boundary features and the land type features in a boundary feature library and a land type feature library, and when the matching degree of the features to be identified of the image boundary and any image boundary feature in all image edge features of preset categories in the image boundary feature library is greater than or equal to a matching threshold value, and the shape similarity of an image area extracted by the features to be identified of the image boundary after being corrected and a global or local graphic image of the position and the orientation of a target land block vector graphic is greater than or equal to the matching threshold value, determining the image area as a target land block area in image data; when the matching degree of all image boundary features to be recognized of the photo and all image boundary features in the image boundary feature library is smaller than a preset threshold value, or the shape similarity between an image area extracted by the image boundary features to be recognized and a global or local graphic image of the position and the direction of a target plot vector graphic after correction is smaller than a preset threshold value, determining all areas of the image data to be recognized as image data target plot areas; when the matching degree of the features to be identified of the image data target land area land class and any image features in all image features of a preset land class is larger than or equal to a matching threshold value, determining the land class of the image area to be identified as the preset land class; determining land types of all single photo target land areas of a target land one by one;

a merging calculation module, configured to merge and calculate the area of all single-photo target parcel areas of a target parcel, extract image data and identification parcel areas of all single-photo target parcel areas of the target parcel, determine the target parcel as the identification parcel when all the image data of the target parcel has only one identification parcel, set the area of the target parcel area as a real space area recorded by a target parcel vector graph, perform shooting correction and deduplication stitching on each image parcel area when all the image data of the target parcel has multiple identification parcels, calculate the spatial resolution of the stitched image when the stitched image completely covers the target parcel vector graph, and calculate the real space area corresponding to each parcel area according to the spatial resolution and the area of non-deformed pixels of each parcel area of the stitched image, when the spliced image cannot completely cover the vector graph of the target land parcel, calculating the real space area corresponding to each land area according to the preset spatial resolution;

the land type determining module is used for determining the land type of a target land block according to the minimum control area of the land type, determining the land type of the target land block as the land type when the target land block only has one land type, and marking the target land block as a small micro-pattern spot if the real space area is smaller than the minimum control area of the land type; when the target land parcel has multiple land types, comparing the real space area of each land type with the minimum control area of the land type, when the real space area of the land type is smaller than the minimum control area, marking the land type area as a small micro land type area, after the comparison of all the multiple land types is finished, combining the small micro land type area with the adjacent land type area, combining the land types into the land types of the adjacent land type area, and obtaining the real space area as the sum of the areas of the combined land type areas; and finally determining the target land block as a multi-land-type land block when the small micro land-type areas of the target land block are combined and various land types exist, and recording the image areas and the real space areas of the various land types.

Preferably, the feature recognition module is further configured to determine the image data to be recognized as image data of an unrelated category and determine the photo as a photograph unrelated to the land category when matching degrees of all the image land features to be recognized of the photograph and all the image land features in the land feature library are smaller than a preset threshold.

Preferably, the ground class includes:

three categories: agricultural land, construction land and unused land;

In accordance with another aspect of the present invention, there is provided a system for generating identification parameters for determining boundaries of target parcel in a single photograph based on image data, the system comprising:

the preprocessing module is used for preprocessing training data before training and comprises: data access, which is used for storing original data such as original photos, shooting information and target plot vector data and the like to a specific physical storage position according to a pre-designed data model, and comprises a data access module which is used for confirming whether the original photos, the shooting information and the target plot vector graphic image data are stored to the specific physical storage position according to the pre-designed data model, after multiple quality verification is carried out on single photo image data, corresponding shooting information and target plot vector graphic data, if the image data of the single photo is not damaged and clear, and the shooting information and the target plot vector graphic data corresponding to the single photo and the correlation relationship are complete, the single photo, the shooting information and the target plot vector graphic data are respectively stored to the preset storage position, and photos, pictures, shooting information and target plot vector graphic data associated with the target plot are determined according to a target plot unique identification code, Shooting information and a vector graphic data file of a target land block, extracting image data in the picture file, shooting information in a shooting information data table and vector graphic data of a vector graphic data set of the target land block, and storing the image boundary recognition output result as a boundary information data set which can be retrieved according to a unique identification code of a piece according to a pre-designed data model and a storage position corresponding to a specific structure when the image boundary recognition output result meets the preset requirement; the data quality verification is used for carrying out image signature on the original photo, determining a repeated original photo by comparing the image signature of the original photo, and removing the repeated original photo to keep a single original photo; performing multiple quality checks on the image data of a single picture, corresponding shooting information and vector graphic data of a target plot to determine whether the single picture is clear and damaged and whether the information and the association are complete; data correction, which is used for removing the single picture if the image data of the single picture is damaged or unclear or the shooting information and the target plot vector graphic data corresponding to the single picture are incomplete; if the single picture is lack of correlation with the corresponding shooting information and the vector graphic data of the target plot, correcting the correlation information;

the initial parameter setting module is used for selecting boundary basic image data in the photo image data, determining boundary initial identification parameters based on the boundary basic image data, training the boundary initial identification parameters through training image data in the image data, and adjusting the boundary initial identification parameters according to output results of the initial identification parameters so as to generate boundary identification parameters to be optimized;

the optimization training module is used for adjusting the output result according to a preset rule, taking the adjusted output result as input data, and performing cyclic training on the boundary identification parameter to be optimized until the output result of the acquired boundary identification parameter to be optimized reaches a stable state;

and the final parameter module is used for stopping training the boundary identification parameter to be optimized when the output result reaching the stable state meets the preset requirement, and taking the optimized boundary identification parameter as the identification parameter for determining the boundary of the target block based on the image data.

Preferably, the adjusting the preset rule of the boundary output result includes:

In accordance with another aspect of the present invention, there is provided a system for generating identification parameters for determining a target parcel of a single photograph based on image data, the system comprising:

the preprocessing module is used for preprocessing training data before training and comprises: data access for storing data such as photos, boundary information and target parcel vector to a specific physical storage location according to a pre-designed data model, including for confirming whether photo files, boundary information and target parcel vector graphics data are stored to the specific physical storage location according to the pre-designed data model, performing multiple quality checks on the photo files, corresponding boundary information and target parcel vector graphics data, respectively storing a single photo, corresponding boundary information and target parcel vector graphics data of the photo files to the preset storage location if the single photo, corresponding boundary information and target parcel vector graphics data and the association relationship are complete, determining photo files and target parcel vector graphics data associated with a target parcel according to a unique identification code of the target parcel, and extracting image data in the photo files and target parcel vector graphics data in the target parcel vector graphics data Block expected land type data; determining corresponding boundary information records according to the unique identification codes of the photos, and extracting the number of the boundaries; the data quality check is used for carrying out multiple quality checks on the image data of a single photo, the corresponding boundary information and the vector graphic data of the target plot so as to determine whether the information and the association are complete; data correction, which is used for removing a single photo of the photo file if the single photo, corresponding boundary information or target plot vector graphic data is incomplete; if the single photo of the photo file is lack of association with the corresponding boundary information and the target plot vector graphic data, correcting the associated information;

the initial parameter setting module is used for selecting the land type basic image data in the boundary of the photo image data target land block, determining a land type initial identification parameter based on the land type basic image data, training the land type initial identification parameter through the training image data in the image data, and adjusting the land type initial identification parameter according to the output result of the initial identification parameter so as to generate a land type identification parameter to be optimized;

the optimization training module is used for adjusting the output result according to the expected land type of the target land parcel, taking the adjusted output result as input data, and performing cycle training on the land type identification parameters to be optimized until the output result of the obtained land type identification parameters to be optimized reaches a stable state;

and the final parameter module is used for comparing the output result of the land types reaching the stable state with the expected land types, stopping training the land type identification parameters to be optimized when the matching degree is greater than or equal to the matching threshold value, and taking the optimized land type identification parameters as the identification parameters for determining the target land type based on the image data.

Preferably, the initial parameter module selects boundary basic image data and ground-class basic image data in the image data, and further selects basic image data in the image data by a non-maximum suppression method.

The technical scheme of the invention provides a method and a system for recognizing a target land parcel by using a live photo, wherein the method comprises the following steps: generating identification parameters for determining the boundaries of target plots of a single photo and the identification parameters of the land types based on the image data; extracting image boundary characteristics and land characteristics of each image data in a plurality of field photo image data through a characteristic extraction model, and establishing an image boundary characteristic library comprising all image boundary characteristics of the plurality of photo image data and a land characteristic library comprising all image land characteristics; extracting the image boundary to-be-identified features and the land to-be-identified features of the single photo image data to be identified based on the boundary identification parameters and the land identification parameters; matching the features to be recognized of the image boundary and the features to be recognized of the land type with the boundary features and the land type features in a boundary feature library and a land type feature library respectively, determining an image target land area according to the matching result of the boundary features, and determining the land type of the image target land area according to the matching result of the land type features; determining land types of all single photo target land areas of a target land one by one; and merging and calculating the area of the land types of all the single photo target land areas of the target land, and determining the land type of the target land according to a preset land type minimum control area rule.

Drawings

A more complete understanding of exemplary embodiments of the present invention may be had by reference to the following drawings in which:

FIG. 1 is a flow chart of a method for identifying a target parcel class using a live photograph in accordance with a preferred embodiment of the present invention;

FIG. 2 is a flow diagram of a method for generating identification parameters for determining boundaries of target parcel in a single photograph based on image data, in accordance with a preferred embodiment of the present invention;

FIG. 3 is a flowchart of a method for generating identification parameters for determining a target parcel of a single photograph based on image data, in accordance with a preferred embodiment of the present invention;

FIG. 4 is a diagram of a data model architecture in accordance with a preferred embodiment of the present invention;

FIG. 5 is a flowchart illustrating the overall process of a system for identifying a target parcel class using live photographs, in accordance with a preferred embodiment of the present invention;

fig. 6 is a block diagram of a system for recognizing a target parcel type using a live photograph in accordance with a preferred embodiment of the present invention;

FIG. 7 is a block diagram of a system for generating identification parameters for determining the boundaries of a target parcel in a single photograph based on image data, in accordance with a preferred embodiment of the present invention;

FIG. 8 is a block diagram of a system for generating identification parameters for determining a single photograph target parcel class based on image data in accordance with a preferred embodiment of the present invention;

FIG. 9-1 is a diagram illustrating the result of image boundary detection according to a preferred embodiment of the present invention; fig. 9-2 is a diagram illustrating the result of image terrain detection according to the preferred embodiment of the present invention.

Detailed Description

The exemplary embodiments of the present invention will now be described with reference to the accompanying drawings, however, the present invention may be embodied in many different forms and is not limited to the embodiments described herein, which are provided for complete and complete disclosure of the present invention and to fully convey the scope of the present invention to those skilled in the art. The terminology used in the exemplary embodiments illustrated in the accompanying drawings is not intended to be limiting of the invention. In the drawings, the same reference numerals are used for the same modules/elements.

Unless otherwise defined, terms (including land industry terms and scientific terms) used herein have the ordinary meaning as understood by those skilled in the art. Further, it will be understood that terms, such as those defined in commonly used dictionaries, should be interpreted as having a meaning that is consistent with their meaning in the context of the relevant art and will not be interpreted in an idealized or overly formal sense.

Fig. 1 is a flowchart of a method for recognizing a target parcel class using a live photograph according to a preferred embodiment of the present invention. The embodiment of the application provides a method for recognizing a target land parcel by using a live photo, which comprises the following steps:

preferably, in step 101: identification parameters are generated that determine the boundaries of the target parcel of a single photograph based on the image data.

Preferably, before extracting the feature to be recognized of the boundary by using the boundary recognition parameter, the method further includes: and extracting the image boundary characteristics of each image data in a plurality of live photo image data through a characteristic extraction model, and establishing an image boundary characteristic library comprising all the image boundary characteristics of the plurality of photo image data.

Preferably, at step 102: and extracting the image boundary to-be-identified characteristics of the single photo image data to be identified based on the boundary identification parameters.

Preferably, in step 103: matching the features to be identified of the image boundary with all boundary features of a boundary feature library, identifying the image boundary according to the boundary feature matching result, and determining an image target plot area: when the matching degree of the features to be recognized of the image boundary and any image boundary features in all image edge features of preset categories in the boundary feature library is greater than or equal to a matching threshold value, and the shape similarity between an image area extracted by the features to be recognized of the image boundary and a global or local graphic image of the position and the direction of a target plot vector graphic after correction is greater than or equal to the matching threshold value, determining the image boundary area as an image target plot boundary, and determining an internal area of the image boundary as a target plot area in image data; when the matching degree of all the image boundary features to be recognized and all the image boundary features in the image boundary feature library is smaller than a preset threshold value, or the shape similarity of an image area extracted by the image boundary features to be recognized and a global or local image of the position and the direction of a target block vector graph after correction is smaller than a preset threshold value, determining an image outer frame as the boundary of the image data, and determining all the areas of the image data to be recognized as the target block area of the image data.

In the present application, the boundary recognition result is stored in a specific storage location of the recognition database according to a pre-designed data model, as shown in fig. 4, for subsequent class recognition.

Preferably, at step 104: identification parameters are generated that determine the target parcel land class for a single photograph based on the image data.

Preferably, before extracting the feature to be recognized of the land category by using the land category recognition parameter, the method further comprises: and extracting the image land class characteristics of each image data in the plurality of live photo image data through a characteristic extraction model, and establishing an image land class characteristic library comprising all the image land class characteristics of the plurality of photo image data.

Preferably, at step 105: and extracting the image land class to-be-identified features of the single photo image data to be identified based on the land class identification parameters.

Preferably, at step 106: and matching the image land to-be-identified features with all land features of a land feature library, and determining the land of the image area to be identified as a preset land when the matching degree of the image data target land area land to-be-identified features and any image features in all image features of the preset land is greater than or equal to a matching threshold value.

Preferably, when the matching degree of all the image land features to be recognized of the photo and all the image land features in the land feature library is smaller than a preset threshold, the image data to be recognized is determined as image data of an irrelevant type, and the photo is determined as a photo irrelevant to the land.

Repeating 101 to 106, and determining the land types of all the target land area of the single photo;

preferably, the ground classes in the present application include:

three categories: agricultural land, construction land and unused land;

Preferably, in step 107: and merging and calculating the area of the land types of all the single photo target land areas of the target land, and determining the land type of the target land according to a preset land type minimum control area rule.

According to the method, by constructing a SpatialAnalysis program, the land type identification results of the target land area region of all the single photos are analyzed, shot and corrected, spliced and area-calculated, and finally the land type is determined. Extracting all single-photo land type identification results including land types and image areas aiming at each target land block, determining the target land type as the identification land type when one identification land type exists and only one identification land type exists, and determining the area of the target land type area as the real space area recorded by the target land block vector graph; when there are multiple types of ground identification, the real space area of each ground type needs to be calculated, and then the final ground type is determined according to the minimum control area.

The method and the device adopt NNdifuse of ENVI to realize data fusion and splicing after shooting correction of various image land areas. When the spliced image completely covers the vector graph of the target plot, calculating the spatial resolution of the image and the real space area corresponding to each regional area; and when the target land block vector graphics cannot be completely covered, calculating the real space area of each land type area according to the preset spatial resolution.

The final land type is determined by comparing the real space area of each land area of the multi-land target land block with the preset minimum land type control area. When the real space area of the land type area is smaller than the minimum control area, the land type area is marked as a small micro land type area, all the small micro land type areas are combined with the adjacent land type area, the land type is the land type of the adjacent large land type area, and iteration comparison and combination are carried out until the real space area of all the land type areas is larger than the preset land type minimum control area. And finally determining the target land parcel as a multi-land parcel when the reserved land parcel still has multiple land parcels, and recording the image areas and the real space areas of the multiple land parcels.

Preferably, the preset ground class minimum control area rule in the present application includes: the minimum control area of the construction land is 200 square meters; the minimum control area of the facility agricultural land is 200 square meters; the minimum control area of the agricultural land is 400 square meters; the minimum control area of unutilized is 600 square meters.

Fig. 2 is a flow chart of a method for generating identification parameters for determining the boundaries of target parcel in a single photograph based on image data according to a preferred embodiment of the present invention. The embodiment of the application provides a method for generating identification parameters for determining the boundary of a target parcel in a single photo based on image data, which comprises the following steps:

preferably, in step 201: and confirming whether the original picture, the shooting information and the vector graphic data of the target plot are stored in a specific physical storage position according to a pre-designed data model. The pre-designed data model, as shown in fig. 4, checks whether the original photograph, the shot information, and the vector graphics data of the target parcel have been stored in the recognition database, and is stored in the file dataset, the data table, and the FileGeoDatabase spatial database format and data structure, respectively. And when the data storage meets the requirements, the next step is carried out.

Preferably, at step 202: and determining the picture, the shooting information and the vector graphic data of the target land block associated with the target land block according to the unique identification code of the target land block, and extracting the image data in the picture file, the shooting information in the shooting information data table and the vector graphic of the vector graphic data set of the target land block.

The image data applied by the embodiment of the application is composed of the on-site evidence demonstrating photos of the real scene, and the shooting azimuth angle, the shooting point geographic coordinates and other corresponding shooting information are required to be complete, and the incidence relation among the data is complete. The embodiment of the application provides a method for assisting in identifying image boundaries by using a target block vector graph. The target plot vector graph is a plot distinguishing boundary line with high accuracy formed by the field based on remote sensing images and historical investigation results, after the method is applied, the interference of pure natural boundaries in image boundary feature recognition can be effectively reduced, the recognition area can be quickly and accurately locked in the recognition boundary, and the problem that the shooting range of a field proof picture generally exceeds the real space of the target plot is solved.

Preferably, before extracting the image data in the photo file, the shooting information in the shooting information data table and the vector graphics of the vector graphics data set of the target land, the method further comprises the steps of carrying out image signature on the original photo, determining a repeated original photo by comparing the image signatures of the original photo, and removing the repeated original photo to reserve a single original photo; performing multiple quality checks on a single picture, corresponding shooting information and vector graphic data of a target plot, and removing the single picture if the image data of the single picture is damaged or unclear; if the shooting information and the vector graphic data of the target plot corresponding to the single picture are incomplete, removing the single picture; if the single picture is lack of correlation with the corresponding shooting information and the vector graphic data of the target plot, correcting the correlation information; and if the image data of the single picture is not damaged and clear, and the shooting information corresponding to the single picture, the vector graphic data of the target land parcel and the incidence relation are complete, respectively storing the single picture, the shooting information and the vector graphic data of the target land parcel in the preset storage positions.

The method comprises the steps of preprocessing image data, extracting, removing duplication, verifying, performing model auxiliary labeling and other preprocessing on various types of data (photos, vector data, metadata files and the like) by constructing a pipeline program, and extracting data such as good images and shooting information of the photos. The method comprises the steps of extracting vector data by using an arcpy module, constructing an incidence relation between a vector and a photo, checking a target plot vector graph, and removing a plot with incomplete and inaccurate attributes and an associated photo due to graph damage. The method and the device apply the md5 hash algorithm to sign the extracted pictures, and delete the repeated pictures by comparing the signatures, wherein only one of the repeated pictures is reserved. The picture is verified by using the Pilow module, damaged pictures with blurred images and unsatisfactory shooting angles are checked and removed by using methods such as open, Verify and load, and pictures with incomplete shooting information are removed.

Preferably, at step 2031: and selecting boundary basic image data in the photo image data.

Preferably, at step 2032: boundary initial identification parameters are determined based on the boundary base image data.

Preferably, at step 2033: and training the boundary initial identification parameter through training image data in the image data so as to adjust the boundary initial identification parameter according to an output result of the initial identification parameter, thereby generating a boundary identification parameter to be optimized.

Preferably, at step 2034: and adjusting the output result according to a preset rule, taking the adjusted output result as input data, and training the boundary identification parameter to be optimized.

Preferably, at step 2035: and judging whether the output result reaches a stable state.

Steps 2031 to 2035 are a loop training process until the output result of obtaining the boundary identification parameter to be optimized reaches a steady state. Wherein 2032 and 2033 are in an or relationship, step 2032 is used for the first cycle, and step 2033 is used for the subsequent training.

This application is through utilizing a small amount of photos to confirm initial identification parameter, and the purpose of confirming initial identification parameter is the work load that reduces artifical mark, screens out the picture that contains the effectual landmass boundary in each place from the sample to further utilize initial identification parameter to do the classification, follow-up only needs a small amount of manual works to rectify, can use as training sample, and the quality is very high. The accuracy of sample and parameter is promoted in step in this application at the training process. By constructing an initial model, a small amount of manually marked samples are used, initial identification parameters with general performance are obtained through training, a large amount of unknown samples are subjected to primary marking through the initial identification parameters, a small amount of manually corrected deviation is carried out on marking results, samples with higher quality are obtained, the initial identification parameters are used for training, optimized identification parameters with better effect are obtained, optimized identification parameters with better performance are used, a large amount of samples are marked again, a small amount of manually corrected deviation is carried out on the marking results, samples with better quality are obtained, the optimized identification parameters are used for training, and the training process is repeated in a circulating mode until the marking results reach a stable state. The method utilizes the initial identification parameters with common effects to carry out the initial processing of the sample, thereby effectively reducing a large amount of repeated manual labels in the training of the common deep learning parameters.

The initial identification parameters are trained through training image data in the image data, namely the initial identification parameters are used for detecting and identifying the sample picture, and the picture comprising the image plot boundary is extracted. And the method also comprises the step of carrying out primary screening on the sample picture to obtain the picture classified according to the image plot boundary.

In the present application, the above two methods for training the recognition parameters may be used alone or in combination. If the method is combined, after target detection, only pictures containing image land boundaries are subjected to primary screening, and the result is 3 possibilities (natural linear ground objects, temporary artificial linear ground objects and permanent artificial linear ground objects). If only one is used, the result is 4 possibilities (no obvious image plot boundary, natural linear ground feature, temporary artificial linear ground feature, permanent artificial linear ground feature) after the preliminary screening.

The method comprises the steps of selecting a sample, marking the boundary position of a land parcel in a picture by using labelImg, making a data set in a Pascal VOC format, and converting the data set into a training sample set of a boundary detection model in a Tfuture format. When the data set is manufactured, a pascalloc _ common file is modified, and output categories are adjusted to be (including no boundary and including boundary). The SSD boundary detection model is trained by utilizing a boundary detection model training sample set.

SSD300 model structure: the SSD adopts VGG16 as a basic model, and convolution layers are added to obtain more feature maps for detection. The four convolutional blocks, Conv _1 to Conv _4, including 10 convolutional layers, are VGG16 networks for extracting picture features.

The method for solving the target detection result external rectangle by the SSD300 model comprises the following steps: first, featuremas of different scales are defined, similar to dividing the artwork into n × n meshes. For each grid unit, defining a plurality of rectangles (envelops) with different sizes and different length-width ratios by taking the grid unit as a center as a candidate region, taking the coordinates (x and y) of the central point of the grid unit and the width and height change of the circumscribed rectangle corresponding to the actual target as the output of a model, and performing regression solution.

The convolution blocks Conv _5 to Conv _7 and block _8 to block _11 sequentially extract Feature maps with different sizes, block _4_ box, block _7_ box, block _8_ box, block _9_ box, block _10_ box and block _11_ box, and solve the boundary position and category of the target block by convolution regression. And finally, selecting an effective target block boundary region from the candidate boundary region by a non-maximum suppression method.

General procedure for non-maxima suppression: a. and (3) regarding the detection result as a candidate set, sequencing the candidate set according to the confidence degrees aiming at each type of target, selecting the target with the highest confidence degree, deleting the target from the candidate set, and adding the target into the result set. b. And calculating the IOU between the elements in the candidate set and the target obtained in the last step, and deleting the elements corresponding to the elements in the candidate set with the IOU larger than a given threshold value. c. And repeating the process until the candidate set is empty, and taking the result set as a final output result.

The breakpoint file of VGG _ VOC0712_ SSD _300x300_ ft _ iter _120000.ckpt is used when the SSD300 model is trained, trained summary data is stored at an interval of 60 seconds in the training process, the model is stored at an interval of 600 seconds, the used weight attenuation factor is 0.0005, an Adam optimization method is used, the learning rate is 0.003, and the batch size is 20. The results of the detection are shown in FIG. 9-1.

The image data is detected and segmented through the SSD house detection method, the obtained target plot boundary is stored, the image can be further segmented by utilizing the boundary, and the image data sample positioned in the target plot boundary is extracted. Ensuring that the sample data subsequently used for land type detection and identification is located in the target block region.

In the actual operation process, after the boundary detection of the target land parcel is finished, the boundary segmentation is not needed, the land category identification is directly carried out by using the preprocessed picture, and the land category identification model can also be processed. These two methods are superior and inferior respectively:

the method has the advantages that the boundary detection and the image segmentation of the target land are firstly carried out, and then the land type recognition is carried out, so that the efficiency of the model can be effectively improved, the robustness is enhanced, and the accuracy is improved. The boundary detection of the village target plot fully utilizes the accurate spatial orientation of the vector data, and a large number of noise areas in the sample picture can be effectively removed; after the sample image is segmented, the size is smaller, the calculated amount of a land type identification model is greatly reduced, and the land type identification efficiency is improved; the accuracy of the land type identification result is improved because the image after boundary detection segmentation is locked with the land type more accurately. The boundary detection and the land type identification of the target land are carried out separately, so that the model can be optimized independently, the accuracy of the boundary detection and the accuracy of the land type identification are improved respectively, and the result can be analyzed and improved conveniently; and the reusability of the model is better.

The application builds a boundary category identification model based on a VGG16 model.

This model has migrated the VGG16 basic model of training on the ImageNet data set to adjust, remove the top layer, add lower floor:

flatten, flattening the input, i.e. converting the multidimensional feature matrix into one-dimensional feature vectors for use by the next layer.

Dense: i.e., fully connected layer, and Dense (n) is a fully connected layer with n hidden neurons.

Batch Normalization: at each SGD, the corresponding activation is normalized by a mini-batch, so that the mean of the result (output signal dimensions) is 0 and the variance is 1. The BN can speed up convergence and control overfitting, reducing the insensitivity of the network to the initialization weights, allowing the use of a larger learning rate.

Dropout, in the training of the neural network, randomly inactivating part of neurons, and blocking cooperation among part of neurons, so that one neuron and the randomly picked neurons are forcibly required to work together, the joint dependency among part of neurons is reduced, and the Dropout reduces the risk of overfitting.

Preferably, at step 2036: and judging whether the boundary output result reaching the stable state meets the preset requirement or not. And when the matching degree of the output result and any image boundary feature in the feature library is greater than or equal to the matching threshold, stopping training, and taking the optimized boundary identification parameter as an identification parameter for determining the boundary of the target plot based on the image data.

Carrying out shape similarity matching on the corrected boundary region and a global or local graphic image of the position and the direction of the target plot vector graphic, determining an output result as an image boundary when the shape similarity matching is greater than or equal to a matching threshold, and determining all regions of the image data as image data target plot regions when the shape similarity matching is less than a preset threshold; and when the matching degree of the output result and any image boundary feature in the feature library is smaller than the matching threshold value and the image data has no obvious boundary, determining the whole area of the image data as the target area of the image data for taking a picture in a large area.

The method comprises the steps of correcting a photo image by using an oblique photography method, designing an oblique photography posture recovery model by using external orientation elements and coplanar conditions, designing a ground-free control point based on a scene constraint method by using a main optical axis inclination angle of most photos larger than 30 degrees, realizing conversion between an image side coordinate system and an object side coordinate system, realizing horizontal correction of a photographic image, performing gross error detection by using a Randac method, and performing adjustment by using a common equation of multiple cameras.

According to the method, a photo shooting space position, a photo shooting angle and a photo shooting direction are utilized to construct a photo visual field surface, the photo visual field surface is intersected with a target plot vector graph, an overlapped part vector is extracted, the actual length is 1 meter, the overlapped part vector is sequentially cut off and converted into an image, an external rectangle is solved, the similarity after correction of an image boundary area is calculated, the similarity is sequenced, the maximum value of the similarity is compared with a preset value, if the similarity is larger than or equal to the preset value, an image boundary, namely a vector graph boundary corresponding to the similarity is matched with the similarity. If the image boundary is smaller than the preset value, the image boundary and the vector graph boundary can not correspond to each other, and the similarity is not matched.

Preferably, at step 2037: and storing the determined image boundary as a boundary information data set which can be retrieved according to the unique identification code of the slice according to a pre-designed data model, and storing the boundary information data set in a storage position corresponding to the specific structure.

Fig. 3 is a flow chart of a method for generating identification parameters for determining the target parcel class for a single photograph based on image data in accordance with a preferred embodiment of the present invention. The embodiment of the application provides a method for generating identification parameters for determining the target parcel land type of a single photo based on image data, which comprises the following steps:

preferably, in step 301: according to the pre-designed data model, whether the photo file, the boundary information and the target block vector graphics data are stored in a specific physical storage location is confirmed. The pre-designed data model is as shown in fig. 4, and checks whether the original photo, the boundary information and the vector graphics data of the target parcel are stored in the boundary training database, and is stored in the File data set and File GeoDatabase spatial database format and data structure respectively. And when the data storage meets the requirements, the next step is carried out.

Preferably, at step 302: determining a photo file and target parcel vector graphic data associated with a target parcel according to a unique identification code of the target parcel, and extracting image data in the photo file and target parcel expected parcel type data in the target parcel vector graphic data; and determining a corresponding boundary information record according to the unique identification code of the photo, and extracting the boundary data.

The image data applied by the embodiment of the application is composed of the on-site evidence demonstrating photos of the real scene, and the corresponding shooting information such as the geographic coordinates of the shooting points and the shooting visual angles is required to be complete, and the incidence relation among the data is complete.

Preferably, before extracting the image data in the photo file, the expected land types in the target land block vectors and the boundary data, performing multiple quality checks on a single photo, the corresponding boundary information and the target land block vector graphic data, and if the single photo, the corresponding boundary information or the target land block vector graphic data of the photo file is incomplete, removing the single photo; if the single photo of the photo file is lack of association with the corresponding boundary information and the target plot vector graphic data, correcting the associated information; and if the single photo, the corresponding boundary information, the target plot vector graphic data and the incidence relation of the photo file are complete, respectively storing the single photo, the corresponding boundary information and the target plot vector graphic data in the preset storage positions.

The method comprises the steps of preprocessing image data, extracting, verifying, carrying out model auxiliary labeling and other preprocessing on various types of data (photos, vector data, metadata files and the like) by constructing a pipeline program, and extracting data such as good images, shooting information, boundary information and the like of the photos. The method comprises the steps of extracting vector data by using an arcpy module, checking vector attributes of a target plot, and removing the plot with incomplete and inaccurate attributes and associated photos.

Preferably, at step 303: based on the boundary information, an image area within the boundary in the image data is extracted.

Preferably, at step 304: and providing expected land type information corresponding to the photos according to the corresponding relation between the photos and the target land blocks.

Preferably, at step 3051: and selecting land base image data in the photo image data.

Preferably, at step 3052: and determining a land initial identification parameter based on the land base image data.

Preferably, at step 3053: and training the initial recognition parameters of the land types through training image data in the image data so as to adjust the initial recognition parameters of the land types according to the output result of the initial recognition parameters, thereby generating the recognition parameters of the land types to be optimized.

Preferably, at step 3054: and matching the expected land types of the target land parcels with the output results.

Preferably, at step 3055: and when the output land type result is not matched with the expected land type and the output result is inaccurate, adjusting the output result land type to the expected land type, taking the adjusted output result as input data, and training the land type identification parameters to be optimized.

Preferably, at step 3056: when the output land type result does not match the expected land type and the expected land type is inaccurate, the expected land type is modified. This step is in an or relationship with 3055.

Preferably, at step 3057: and judging whether the output result reaches a stable state.

Steps 3051 to 3057 are a loop training process until the output result of obtaining the to-be-optimized land type identification parameter reaches a stable state. Wherein 3052 and 3053 are in an OR relationship, step 3052 is recycled for the first time, and step 3053 is used for the subsequent training.

The method comprises the following steps of constructing a ground class identification parameter training model based on an image boundary identification parameter training model by using a migration and multiplexing method, and adjusting the ground class identification parameter training model as follows:

the boundary of the object of the SSD300 target detection model is adjusted to be the region of the ground class, and the segmentation step is removed after the detection, and the detection result is shown in fig. 9-2:

adjusting the classification of the sample primary screen from the image plot boundary classification to the image plot classification;

constructing a ground class identification model based on the VGG16 model;

and removing the similarity matching part and adjusting the similarity matching part into the ground class matching.

The training process and the optimization principle are basically consistent and are not described in detail.

The ground class identification model constructed based on the VGG16 model realizes multi-level cascade classification. And the following innovative method is adopted when training the ground class recognition classification model:

and carrying out weak supervision training by using a small amount of roughly marked samples, and then carrying out gradually stronger supervision training by using a part of finely marked large samples for multiple times. The rough marking is to mark a small number of classified and selected field typical photos, and the thick and thin marking is to manually and carefully judge, mark and correct more complex field photos. Although the roughly labeled sample cannot completely match all photos in the real scene, the roughly labeled sample has larger similarity and certain discrimination, and the finely labeled sample has stronger classification adaptability. By doing so, the cost of model training can be effectively reduced, and the accuracy can be rapidly improved to a relatively good level. If 3-5 sample photos of arbor woodland (0301) are roughly marked, the model obtained after training is used for identifying unknown new sample photos (about 100 photos), the identification degree can reach 30%, the identification result is finely marked to form (about 100) new sample photos, the model obtained after retraining is used for identifying unknown new sample photos (about 1000 photos), the identification degree can reach 50%, and the process is repeated, so that the identification rate of the model can be improved to a reasonable level in a short time, such as about 85%. In the process, the feasibility of the model and the direction of the optimization improvement of the model can be timely and quickly checked, and the effectiveness of the improvement measures can be evaluated.

The ground class identification model based on the VGG16 built by the method is trained by the application by using the samples. And a cyclic lifting training method is used, the accuracy of the model is continuously improved, after the accuracy reaches 90%, the difficulty of optimization is gradually increased, at the moment, more samples are greatly increased, and meanwhile, the identification accuracy is improved by aiming at the special condition of manual result marking. Even so, in the process of the current location type identification, many complex investigation technical rules cannot be simply realized through image identification, and the problems are taken as the need to combine various data and specified special rules, and the subsequent conditions are further improved.

Preferably, at step 306: and comparing the output result of the land types reaching the stable state with the expected land types, stopping training the land type identification parameters to be optimized when the matching degree is greater than or equal to a matching threshold value, and taking the optimized land type identification parameters as identification parameters for determining the target land type of the land blocks based on the image data.

The application is based on the target detection model of the convolutional neural network, and the image data marked by the information is used as the input data to obtain the result set of the image, and the method further comprises the following steps: a target detection model based on a convolutional neural network selects a result set of image data from input data through a non-maximum suppression method.

FIG. 4 is a diagram of a data model architecture in accordance with a preferred embodiment of the present invention. An embodiment of the present application provides a data model, including: an identification database, a boundary training database, and a geo-training database.

Preferably, the identification data database in the present application comprises raw data, intermediate data and result data. The original data mainly comprises an original photo library and a target plot vector library, and the original photo library extracts shooting information to form a shooting information library. The intermediate data comprises a boundary information library, an identification area photo library, a boundary characteristic library and a land type characteristic library. And the result data is a target land parcel type identification result database. The input data of the boundary training database comprises a boundary training photo library, a target land mass vector library, a shooting information library and a target land mass vector graphic image database, and the output data is a boundary information database, wherein the boundary training photo library is obtained by removing duplicates and defects from an original photo library and is separately stored. The input data of the land training database comprises a land training photo library, a target land vector expected land database and a boundary information library, wherein the land training photo library is obtained by removing the defects from the boundary training photo library and is separately stored, and the boundary information library is output data of the boundary training database.

Preferably, in the present application, the photos and the intermediate image data are stored and managed in a File form, the vector data are stored and managed by using arcsis File geocatabase, the table data of the shooting information, the expected land types and the like are managed by using an arcsis File geocatabase attribute table, and the model training parameters are managed by using a configuration File form.

Fig. 5 is a flowchart illustrating a system for recognizing a target parcel class using a live photograph according to a preferred embodiment of the present invention. According to the method, a boundary identification parameter training system, a land type identification parameter training system and a field photo land type identification system are built, target land block vector data are combined, photos for investigation and evidence-making of field workers are investigated, and land types are identified and determined. The system comprehensively utilizes two architectures of B/S and C/S, and the technologies of Keras framework, HTML5, javascript, CSS, GIS, Java and the like are used, so that system design and development are realized. The processing sequence between systems is:

for the sample photo, training the sample photo in a boundary recognition parameter training system, determining a boundary recognition parameter, and recognizing the boundary of the photo plot; training in a land type recognition parameter training system by using the sample photo, determining a land type recognition parameter, extracting image data and expected land types in the boundary by using the photo boundary, and recognizing the land types in the photo land block boundary; and establishing a boundary feature library and a land class feature library in a field photo land class recognition system by using the sample, performing feature extraction, feature recognition, merging calculation and finally determining the land class.

And for the photo to be tested, utilizing the trained boundary identification parameters and the ground class use parameters, and utilizing the test photo to perform feature extraction, feature identification and merging calculation on the field photo and ground class identification system to finally determine the ground class.

The method and the device realize the land type identification of the land survey evidence-demonstrating photo, and can be used for the internal work check after the land survey evidence-demonstrating.

Fig. 6 is a block diagram of a system for recognizing a target parcel class using a live photograph according to a preferred embodiment of the present invention. As shown, the system includes:

the system comprises a preprocessing module (601) used for extracting shooting information such as shooting azimuth angles, shooting point geographic coordinates and shooting pitch angles from original photos, storing the shooting information into a shooting information data table which can be retrieved according to a piece unique identification code, storing target block vector graphic data into a target block vector graphic data set which can be retrieved according to a block unique identification code, storing the original photos into a photo file data set which can be retrieved according to the piece unique identification code and the block unique identification code, and storing the photo file data set in an identification database according to a data model shown in figure 5. The system is also used for carrying out image signature on the original photo, determining a repeated original photo by comparing the signatures, and removing the repeated original photo to keep a single original photo; performing multiple quality checks on the image data of a single picture, the corresponding shooting information and the vector graphic data of the target plot; if the photo image data is damaged or unclear or the corresponding data is incomplete, removing the single photo; if the photo is lack of correlation with the corresponding data, correcting the correlation information; and if the image data of the single picture is not damaged and clear and the data and the incidence relation are complete, respectively storing the single picture, the shooting information and the vector graphic data of the target plot in the preset storage positions. And when the image boundary identification output result meets the preset requirement, the image boundary identification output result is stored as a boundary information data set which can be retrieved according to the unique identification code of the image and is stored in a corresponding storage position of the identification database.

And the feature library generating module (602) is used for extracting the image boundary features and the land features of each image data in the plurality of live photo image data through the feature extraction model, and establishing an image boundary feature library comprising all the image boundary features of the plurality of photo image data and a land feature library comprising all the image land features.

And the feature extraction module (602) is used for extracting the features to be identified of the image boundary and the features to be identified of the land class of the single photo image data to be identified based on the boundary identification parameters and the land class identification parameters.

The characteristic identification module (603) is used for matching the characteristic to be identified of the image boundary with all the characteristics in the boundary characteristic library, and when the matching degree of the characteristic to be identified of the image boundary and any image boundary characteristic in all the image edge characteristics of the preset category in the boundary characteristic library is greater than or equal to a matching threshold value, and the shape similarity of an image area extracted by the characteristic to be identified of the image boundary after being corrected and a global or local graphic image of the position and the direction of a target block vector graphic is greater than or equal to the matching threshold value, determining the image area as the target block area in the image data; and when the matching degree of all image boundary features to be recognized of the photo and all image boundary features in the boundary feature library is smaller than a preset threshold value, or the shape similarity of an image area extracted by the image boundary features to be recognized and a global or local image of the position and the direction of the target block vector graph after correction is smaller than the preset threshold value, determining all areas of the image data to be recognized as image data target block areas. The image feature recognition system is also used for matching the features to be recognized of the image land types with all the features in the land type feature library, and when the matching degree of the features to be recognized of the image data target land area land types and any image feature in all the image features of the preset land types is larger than or equal to a matching threshold value, determining the land types of the image areas to be recognized as the preset land types; and when the matching degrees of the features to be recognized of all image types of the photo and all image type features in the type feature library are smaller than a preset threshold value, determining the image data to be recognized as image data of irrelevant types, and determining the photo as a photo irrelevant to the type.

In the application, a feature extraction module (602) and a feature recognition module (603) are repeatedly executed to determine the land types of all the target land areas of the single photo one by one.

The merging calculation module (604) is used for extracting image data and identification land types of target land areas of all single photos of each target land, determining the target land types as the identification land types when one identification land type exists and only one identification land type exists, and setting the area of the target land type areas as the real space area recorded by the target land vector graphics; when multiple types of recognized land types exist, shooting correction and de-duplication splicing are carried out on land type areas of all images, when the spliced images completely cover the vector graphics of the target land block, the spatial resolution of the spliced images is calculated, and the real spatial areas corresponding to the land type areas are further calculated; and when the spliced image cannot completely cover the vector graph of the target land block, calculating the real space area corresponding to each land area according to the preset spatial resolution.

The land type determining module (605) is used for determining the land type of the target land block according to the minimum control area of the land type, when only one land type exists, the land type of the target land block is determined as the land type, and if the real space area is smaller than the minimum control area of the land type, the land type is marked as a small micro-map spot; when multiple types of land exist, the real space area of each type of land is compared with the minimum control area of the land, when the real space area is smaller than the minimum control area, the real space area is marked as a small micro land area, all the small micro land areas are combined with the adjacent land areas, the land is the land of the adjacent large land area, and iteration is carried out for comparison and combination until the real space areas of all the land areas are all larger than the preset land minimum control area. And finally determining the target land parcel as a multi-land parcel when the reserved land parcel still has multiple land parcels, and recording the image areas and the real space areas of the multiple land parcels.

Preferably, the preset ground class minimum control area rule comprises the following steps: the minimum control area of the construction land is 200 square meters; the minimum control area of the facility agricultural land is 200 square meters; the minimum control area of the agricultural land is 400 square meters; the minimum control area of unutilized is 600 square meters.

Preferably, the ground category includes:

three categories: agricultural land, construction land and unused land;

The system 600 for recognizing a target parcel type using a live photograph according to a preferred embodiment of the present invention corresponds to the method 100 for recognizing a target parcel type using a live photograph according to another preferred embodiment of the present invention, and thus, a detailed description thereof will be omitted.

Fig. 7 is a block diagram of a system for generating identification parameters for determining the boundaries of a target parcel in a single photograph based on image data, in accordance with a preferred embodiment of the present invention. As shown, the system includes:

and the preprocessing module (701) is used for confirming whether the original picture, the shooting information and the target block vector graphic image data are stored to a specific physical storage position according to a pre-designed data model, determining a picture, the shooting information and a target block vector graphic data file associated with the target block according to the target block unique identification code, and extracting the image data in the picture file, the shooting information in the shooting information data table and the vector graphic data of the target block vector graphic data set. The system is also used for carrying out image signature on the original photo, determining a repeated original photo by comparing the image signature of the original photo, and removing the repeated original photo to keep a single original photo; performing multiple quality checks on the image data of a single picture, the corresponding shooting information and the vector graphic data of the target plot, and removing the single picture if the image data of the single picture is damaged or unclear or the corresponding shooting information and the vector graphic data of the target plot are incomplete; if the single picture is lack of correlation with the corresponding shooting information and the vector graphic data of the target plot, correcting the correlation information; and if the image data is not damaged and clear, and the corresponding shooting information and the vector graphic data of the target land parcel and the incidence relation are complete, respectively storing the single picture, the shooting information and the vector graphic data of the target land parcel in preset storage positions. And when the image boundary training recognition output result meets the preset requirement, the image boundary training recognition output result is stored as a boundary information data set which can be retrieved according to the unique identification code of the image and is stored in a corresponding storage position of the boundary training database.

An initial parameter setting module (702) for selecting boundary basic image data in the photo image data, determining a boundary initial identification parameter based on the boundary basic image data, training the boundary initial identification parameter through training image data in the image data, and adjusting the boundary initial identification parameter according to an output result of the initial identification parameter, thereby generating a boundary identification parameter to be optimized;

the optimization training module (703) is used for adjusting the output result according to a preset rule, taking the adjusted output result as input data, and performing cyclic training on the boundary identification parameter to be optimized until the output result of the acquired boundary identification parameter to be optimized reaches a stable state;

a final parameter module (704) for stopping training of the boundary identification parameter to be optimized when the output result reaching the stable state meets a preset requirement, and taking the optimized boundary identification parameter as an identification parameter for determining the boundary of the target block based on the image data;

if the boundary identification output result is the ground linear segmentation ground object, the ground object is calibrated to be a natural linear ground object or an artificial linear ground object, and the artificial linear ground object is subdivided into a temporary type and a permanent type;

and if the non-ground linear segmentation ground object is obtained through the boundary identification output result, the non-ground linear segmentation ground object is calibrated as a non-boundary characteristic.

The system 700 for generating identification parameters for determining the boundary of a target parcel in a single photo based on image data according to the preferred embodiment of the present invention corresponds to the method 200 for generating identification parameters for determining the boundary of a target parcel in a single photo based on image data according to the preferred embodiment of the present invention according to another preferred embodiment of the present invention, and will not be described herein again.

Fig. 8 is a block diagram of a system for generating identification parameters for determining a target parcel of a single photograph based on image data, in accordance with a preferred embodiment of the present invention. As shown, the system includes:

a preprocessing module (801) for confirming whether the photo file, the boundary information and the target parcel vector graphics data are saved to a specific physical storage location according to a pre-designed data model, determining the photo file, the target parcel vector graphics data and the corresponding boundary information record associated with the target parcel according to the target parcel unique identification code, and extracting the image data in the photo file, the target parcel expected land type data and the boundary data in the target parcel vector graphics data. The system is also used for carrying out multiple quality checks on the photo file, the corresponding boundary information and the vector graphic data of the target plot, and if a single photo of the photo file, the corresponding boundary information or the vector graphic data of the target plot are incomplete, the single photo is removed; if the single photo is lack of association with the corresponding boundary information and the target plot vector graphic data, correcting the associated information; and if the photo, the corresponding boundary information, the target land parcel vector graphic data and the incidence relation are complete, respectively storing the photo, the corresponding boundary information and the target land parcel vector graphic data in corresponding storage positions of the land class training database.

The initial parameter setting module (802) is used for selecting the land type basic image data in the boundary of the photo image data target land block, determining the initial land type identification parameters based on the land type basic image data, training the initial land type identification parameters through the training image data in the image data, and adjusting the initial land type identification parameters according to the output result of the initial identification parameters so as to generate the land type identification parameters to be optimized;

the optimization training module (803) is used for adjusting the output result according to the expected land type of the target land parcel, taking the adjusted output result as input data, and performing cyclic training on the land type identification parameters to be optimized until the output result of the obtained land type identification parameters to be optimized reaches a stable state;

and the final parameter module (804) is used for comparing the output result of the land types reaching the stable state with the expected land types, stopping training of the land type identification parameters to be optimized when the matching degree is greater than or equal to the matching threshold value, and taking the optimized land type identification parameters as the identification parameters for determining the target land type of the land parcel based on the image data.

Preferably, the initial parameter modules (702, 802) in the boundary training system and the ground class training system select boundary base image data and ground class base image data in the image data, and further select base image data in the image data by a non-maximum suppression method.

The system 800 for generating identification parameters for determining the target parcel of a single photo based on image data according to the preferred embodiment of the present invention corresponds to the method 300 for generating identification parameters for determining the target parcel of a single photo based on image data according to the preferred embodiment of the present invention according to another preferred embodiment of the present invention, and will not be described herein again.

Until now, the land type interpretation based on the evidence-raised photos of the land survey site can only be carried out manually, and the main reason is that the following difficulties exist: firstly, because of the limitation of the shooting angle, the shooting ranges of a large number of pictures can not completely correspond to the boundary of the target plot, and the approximate corresponding relation between the picture shooting area and the real space needs to be analyzed manually; secondly, because the photos reflect continuous ground features, multiple land features are generally presented, so that target detection and position area locking cannot be performed according to the features of a single land, and possible land can be further judged only after the spatial orientation is determined; thirdly, because the target land parcel may have various land parcels, but the photo cannot accurately calculate the represented real space area through visual interpretation, and only after the position is locked, the image data capable of reflecting the land parcel characteristics is further searched, and whether the various land parcels existing in the target land parcel can be integrated or not is further determined through the image measuring area.

This application has solved through combining together degree of depth learning technique and spatial data processing technique, photogrammetry technique, has solved above-mentioned difficult point: the image boundary achievement extracted based on the target land block vector is innovatively used during training of the land type recognition classification model, the corresponding problem of the spatial position is solved, and the region position of the land type to be recognized in the photo is accurately locked; a photogrammetry method is introduced, a plurality of photos are corrected, deduplicated and spliced, the photo space resolution of similar aerial photography is provided, and the problem of area calculation is solved; and finally, determining the final land type determination problem of the multi-land type target land block according to the minimum control area of each land type. After the problem is solved, the land type interpretation based on the evidence demonstrating picture on the land survey site is automatically completed through the computer system, the investigation of common overlarge comprehensive problems in the survey is also solved, the labor input cost is greatly reduced, the survey data quality is improved, and the method has a wide application prospect in the whole field of the national survey.

The invention has been described with reference to a few embodiments. However, other embodiments of the invention than the one disclosed above are equally possible within the scope of the invention, as would be apparent to a person skilled in the art from the appended patent claims.

Generally, all terms used in the claims are to be interpreted according to their ordinary meaning in the technical field, unless explicitly defined otherwise herein. All references to "a// the [ device, component, etc. ] are to be interpreted openly as at least one instance of a device, component, etc., unless explicitly stated otherwise. The steps of any method disclosed herein do not have to be performed in the exact order disclosed, unless explicitly stated.

Claims

1. A method for recognizing a present land use class (hereinafter, referred to as a land class) of a target land by using a live photograph, the method comprising:

2. The method according to claim 1, when the matching degree of all image land features to be recognized of the photo and all image land features in the land feature library is smaller than a preset threshold, determining the image data to be recognized as image data of an irrelevant category, and determining the photo as a land-independent photo.

3. The method of claim 1, wherein the generating further comprises, before determining the identification parameters of the boundary of the target parcel and the feature of the parcel to be identified in the single photograph based on the image data, and extracting the feature of the image boundary to be identified and the feature of the parcel to be identified in the image data based on the identification parameters:

performing multiple quality checks on image data of a single picture, corresponding shooting information and vector graphics data of a target plot, and removing the single picture if the image data of the single picture is damaged or unclear; if the shooting information and the vector graphic data of the target plot corresponding to the single picture are incomplete, removing the single picture; if the single picture is lack of correlation with the corresponding shooting information and the vector graphic data of the target plot, correcting the correlation information; and if the image data of the single picture is not damaged and clear, and the shooting information corresponding to the single picture, the vector graphic data of the target land parcel and the incidence relation are complete, respectively storing the single picture, the shooting information and the vector graphic data of the target land parcel in the preset storage positions.

4. The method of claim 1, the terrain comprising:

three categories: agricultural land, construction land and unused land;

5. A method of generating identification parameters for determining boundaries of target parcel in a single photograph based on image data, the method comprising:

6. The method of claim 5, the preset rule of boundary recognition output result adjustment comprising:

7. The method of claim 5, wherein prior to determining the original photograph, the shot information, and the vector graphic of the target parcel and the image data file associated with the target parcel based on the unique identification code of the target parcel, and extracting the image data in the original photograph file, the shot information in the shot information data table, and the vector graphic of the vector graphic data set of the target parcel, further comprises:

8. A method of generating identification parameters for determining a target parcel of a single photograph based on image data, the method comprising:

9. The method of claim 8, said determining a photograph file and target parcel vector graphics data associated with a target parcel based on a target parcel unique identification code and extracting target parcel expected terrain class data in the image data in the photograph file and the target parcel vector graphics data; determining a corresponding boundary information record according to the unique identification code of the photo, and before extracting the boundary data, further comprising:

10. The method of claim 5 and claim 8, wherein selecting the base image data from the image data further comprises selecting the base image data from the image data by a non-maxima suppression method.

11. A method for merging and calculating the area of land types of target land areas of all single photo target land areas of a target land and determining the land type of the target land according to a land type minimum control area rule of a preset category, wherein the method comprises the following steps:

12. The preset geo-based minimum control area rule of claim 11, the construction land minimum control area is 200 square meters; the minimum control area of the facility agricultural land is 200 square meters; the minimum control area of the agricultural land is 400 square meters; the minimum control area of unutilized is 600 square meters.

13. A system for identifying a target parcel class using live photographs, the system comprising:

14. The system of claim 13, wherein the feature recognition module is further configured to determine the image data to be recognized as image data of an irrelevant category and determine the photo as a photo irrelevant to the ground category when all the image features to be recognized of the photo are matched with all the image features in the ground category feature library by less than a preset threshold.

15. The system of claim 13, wherein the preset minimum control area rule of land type is that the minimum control area of construction land is 200 square meters; the minimum control area of the facility agricultural land is 200 square meters; the minimum control area of the agricultural land is 400 square meters; the minimum control area of unutilized is 600 square meters.

16. The system of claim 13, the terrain class comprising:

three categories: agricultural land, construction land and unused land;

17. A system for generating identification parameters for determining boundaries of target parcel in a single photograph based on image data, said system comprising:

18. The optimization training module of claim 17, wherein the boundary output result adjustment preset rule comprises:

19. A system for generating identification parameters for determining a target parcel of a single photograph based on image data, the system comprising:

20. The system of claim 17 and claim 19, wherein the initial parameter module selects boundary basis image data and ground basis image data in the image data, and further comprises selecting basis image data in the image data by a non-maxima suppression method.