CN117372669B

CN117372669B - Moving object detection device based on natural resource image

Info

Publication number: CN117372669B
Application number: CN202311666297.0A
Authority: CN
Inventors: 高大山; 袁勇; 郭毅轩; 赵明考; 刘占恒; 马超; 武霄娟; 王晓宇; 陈畅; 张子丰; 徐兆祥; 呼丽丽; 刘松楠; 王欢
Original assignee: Beijing Xinxing Keyao Information Technology Co ltd
Current assignee: Beijing Xinxing Keyao Information Technology Co ltd
Priority date: 2023-12-07
Filing date: 2023-12-07
Publication date: 2024-03-08
Anticipated expiration: 2043-12-07
Also published as: CN117372669A

Abstract

The invention relates to a moving target detection device based on natural resource images, which aims to improve the efficiency and accuracy of natural resource image recognition. The device comprises four main modules: the system comprises an image acquisition module, an image difference processing module, a motion scene reconstruction module and a motion target confirmation module. The image acquisition module is responsible for collecting data related to natural resource image recognition from a plurality of dynamic motion analysis data. The image difference processing module is connected with the image acquisition module, and is used for preprocessing the collected data and simultaneously carrying out deep analysis on the three-dimensional image by using an image processing algorithm. And the motion scene reconstruction module performs motion scene reconstruction on the processed data according to the set scene reconstruction model and algorithm. And finally, the moving object confirmation module calculates a moving scene reconstruction result through a moving object analysis model, wherein the moving object analysis model comprises a map moving range model, a moving speed model and a moving density model, and supports a user to customize display contents according to specific requirements. The application of the device is helpful for improving the decision quality of natural resource image recognition, so that the image recognition is more scientific and reasonable.

Description

Moving object detection device based on natural resource image

Technical Field

The invention relates to the technical field of image differential processing, in particular to a moving target detection device based on natural resource images.

Background

In the field of natural resource image recognition, challenges are faced in how to accurately and efficiently analyze and synthesize large amounts of multidimensional data. Such data includes, but is not limited to, geographic Information System (GIS) data, three-dimensional images, color, and ecological environment statistics. Efficient data integration and analysis is critical to developing viable image recognition strategies, especially when considering the complex interactions of various environmental and social color factors.

The real-scene three-dimensional space is taken as real, stereoscopic and time-sequential space-time information reflecting human production, living and ecological space, is a novel national important infrastructure, and realizes real-time association and intercommunication of digital space and real space through 'man-machine compatibility, internet of things perception and ubiquitous service'.

Conventional image difference processing methods often fail to take full advantage of the potential provided by modern technology, particularly when processing large volumes of complex three-dimensional images and other spatial data. The application of the existing image processing technology in the field of natural resource image recognition is generally limited to basic image analysis, such as simple feature recognition and classification, and lacks the capability of deep registration of multi-dimensional image recognition data. This results in surfacing of three-dimensional images and other geographic information data analysis during image recognition, failing to take full advantage of the potential value of image data.

Therefore, it is necessary to develop a moving object detection device based on natural resource images.

Disclosure of Invention

The application provides a moving target detection device based on natural resource images, so as to improve the efficiency and accuracy of natural resource image recognition.

The application provides a moving target detection device based on natural resource images, which comprises an image acquisition module, an image difference processing module, a moving scene reconstruction module and a moving target confirmation module;

the image acquisition module is configured to acquire natural resource image identification related data from a plurality of dynamic motion analysis data, wherein the dynamic motion analysis data comprises a geographic information system, a three-dimensional image and statistical data, and the acquired natural resource image identification related data comprises a plurality of dimensions including spatial distribution, color indexes and ecological environment;

the image difference processing module is connected with the image acquisition module and is configured to preprocess the collected natural resource image identification related data, the preprocessing comprises data cleaning, format conversion and data registration, and an image processing algorithm is used for processing the three-dimensional image to extract geographic motion characteristics including land coverage type characteristics and land utilization modes;

The motion scene reconstruction module is connected with the image differential processing module and is configured to reconstruct a motion scene according to the set scene reconstruction model and algorithm, wherein the scene reconstruction model comprehensively considers a plurality of image recognition dimensions including spatial distribution, color indexes, ecological environment and geographic motion characteristics;

the moving object confirming module is connected with the moving scene reconstructing module and is configured to calculate a moving scene reconstructing result in the form of a moving object analysis model, wherein the moving object analysis model comprises a map moving range model, a moving speed model and a moving density model.

Still further, the image differential processing module includes an image processing subsystem for processing a three-dimensional image, the image processing subsystem being specifically configured to:

performing data preprocessing on the three-dimensional image, wherein the preprocessing comprises color correction, denoising and contrast adjustment so as to improve the image quality;

separating the motion characteristic layers of the preprocessed image to obtain a plurality of motion characteristic layers including a texture layer, a color layer and a shape layer;

extracting key information on each motion characteristic layer by adopting a multidimensional image differential processing technology, wherein the key information comprises texture characteristics of different land coverage, including water, woodland and bare land, are identified on a texture layer; analyzing color features of different land types on the color layer; extracting shape information of ground buildings and roads on the shape layer;

Registering the features of different motion feature layers by using a graph-based registration algorithm, wherein the registration algorithm maximizes comprehensive feature information while maintaining the feature independence of each layer, and classifies the features after registration to identify land coverage types and land utilization modes;

the extracted key information and the registered features are mapped and checked with geographic information system data to ensure that the geographic motion features are matched with specific positions in the real world.

Still further, the image processing subsystem includes a deep learning model for decomposing the image into a plurality of motion feature layers, the deep learning model including a backbone network and a feature extraction sub-network, wherein the backbone network uses a deep convolutional neural network for extracting basic features of the image; the feature extraction sub-network comprises a texture sub-network, a color sub-network and a shape sub-network; the texture subnetwork comprises a plurality of convolution layers for capturing texture features from fine granularity to coarse granularity; the color sub-network comprises a convolution layer based on color space conversion and is used for analyzing color distribution and change of an image; the shape subnetwork comprises a sequence convolution layer and an edge detection layer, and is used for extracting outline and structure information of the ground feature.

Still further, the texture subnetwork introduces a focus mechanism to enhance key motion characteristics of the texture.

Still further, the color subnetwork uses a color histogram equalization layer to enhance color features, the color histogram equalization layer being specifically configured to:

receiving a three-dimensional image in an RGB format as an input;

converting an input RGB image into an HSV color space by adopting an image processing algorithm;

applying color histogram equalization processing to the V-channels in the HSV color space to enhance the contrast and color characteristics of the image;

and converting the HSV image subjected to the histogram equalization processing back to an RGB color space.

Still further, the shape subnetwork includes a component for executing shape descriptors and geometry analysis algorithms, the component configured to:

identifying the edges of the ground objects in the three-dimensional image by using an edge detection technology so as to highlight the outline and the boundary of the ground objects;

applying shape descriptors, including fourier descriptors or Hu moments, to quantify the identified edge information to capture the basic features of the shape while ensuring that the descriptors are invariant to image rotation, scaling, and translation;

the basic characteristics of the shape are analyzed by implementing a geometric analysis algorithm, including calculating the area, perimeter, compactness and orientation of the geometric shape.

Still further, the motion scene reconstruction module is specifically configured to:

each image recognition dimension is defined as a node in the graph network,

calculating pearson correlation coefficients among the image recognition dimensions, and establishing edges between nodes with the pearson correlation coefficients higher than a preset threshold;

the importance of each image recognition dimension is analyzed using a random forest model, and the weight of the edge is determined according to the importance.

the dynamic recognition calibration value S is calculated according to the following formula:

wherein N is the total number of nodes in the graph network;representing nodes in the graph networkIs defined by the first node and the second node;is the weight of an edge in the graph network, both sides of the edge include nodesSum node；Is a nodeFor analyzing interaction patterns between nodes in the graph network,the calculation formula of (2) is as follows:

wherein,is a nodeIs a degree of (3).

The beneficial effects of this application include: the beneficial effects of this application are manifested in the following aspects: (1) By integrating a plurality of dynamic motion analysis data such as geographic information system data, three-dimensional images and statistical data, the device can provide an overall view of natural resource image recognition. This comprehensive data set provides a richer and multidimensional information basis for image recognition decisions. (2) The image difference processing module not only performs basic data cleaning and format conversion, but also combines a data registration technology and an image processing algorithm. This allows the apparatus to efficiently process and analyze complex data, particularly in terms of extracting land coverage type features and land use patterns, thereby improving the accuracy and efficiency of image differencing processing. (3) The motion scene reconstruction module combines a plurality of image recognition dimensions such as spatial distribution, color indexes, ecological environment, geographic motion characteristics and the like, so that a scene reconstruction result is more comprehensive and accurate. The method not only considers a single index, but also provides comprehensive analysis for image recognition decision by integrating scene reconstruction models and algorithms with multiple dimensions. (4) The moving object confirming module is designed to enable the result of the moving scene reconstruction to be calculated clearly through the moving object analysis model, wherein the moving object analysis model comprises a map moving range model, a moving speed model, a moving density model and the like. The visual computing mode is helpful for users to understand and analyze scene reconstruction results more easily, thereby promoting the effective transmission of information and the convenience of decision. (5) The motion scene reconstruction of multiple image recognition dimensions is comprehensively considered, and scientific decision support is provided for natural resource image recognition. By accurately analyzing and calculating the dynamic recognition degree among the dimensions, the device can help an image recognizer to recognize potential problems and improvement points, so that an image recognition scheme is optimized.

Drawings

Fig. 1 is a schematic diagram of a moving object detection device based on a natural resource image according to a first embodiment of the present application.

Detailed Description

In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present application. This application is, however, susceptible of embodiment in many other ways than those herein described and similar generalizations can be made by those skilled in the art without departing from the spirit of the application and the application is therefore not limited to the specific embodiments disclosed below.

The first embodiment of the application provides a moving object detection device based on natural resource images. Referring to fig. 1, a schematic diagram of a first embodiment of the present application is shown. A moving object detection device based on a natural resource image according to a first embodiment of the present application will be described in detail with reference to fig. 1.

The data motion scene reconstruction device comprises an image acquisition module 101, an image difference processing module 102, a motion scene reconstruction module 103 and a motion target confirmation module 104.

The image acquisition module 101 is configured to acquire natural resource image recognition related data from a plurality of dynamic motion analysis data, where the dynamic motion analysis data includes geographic information system data, three-dimensional images, and statistical data, and the acquired natural resource image recognition related data includes a plurality of dimensions including spatial distribution, color index, and ecological environment.

The image acquisition module 101 is a key component of a moving object detection device based on natural resource images. The main function of the module is to collect data related to natural resource image recognition from various dynamic motion analysis data, so as to ensure that the collected information is comprehensive, accurate and real-time. Specifically, the module has the following characteristics and functions:

multiple dynamic motion analysis data integration: the module is designed to collect information from a plurality of dynamic motion analysis data including Geographic Information System (GIS) data, three-dimensional images, and statistics. The dynamic motion analysis data are characterized by providing multidimensional information from macroscopic to microscopic, and providing comprehensive data support for natural resource image recognition.

Image acquisition technology: to ensure accuracy and timeliness of the data, the module employs the latest image acquisition technology. For example, surface images are captured using high resolution three-dimensional techniques, and environmental data is collected from various sensors and monitoring sites using real-time data streaming techniques.

Data updating and real-time monitoring: the module can update data periodically to ensure that the provided information reflects the latest natural resource image recognition state. For some key indexes, such as environmental quality data, the module can realize real-time monitoring and provide instant data support for image recognition decision.

The image acquisition module 101 is made up of a plurality of subsystems, each of which is responsible for collecting data from specific dynamic motion analysis data. These subsystems include a Geographic Information System (GIS) data collection subsystem, a three-dimensional image collection subsystem, and a statistics data collection subsystem.

Each subsystem is equipped internally with efficient image acquisition and transmission equipment. For example, the GIS data collection subsystem is connected with GIS databases in various places to acquire the latest space data in real time; the three-dimensional image collecting subsystem acquires a high-resolution surface image in a satellite or unmanned aerial vehicle mode; the statistical data collection subsystem is connected with the database interface of the mechanism and periodically updates the statistical data related to the color and the environment.

The module adopts the latest data integration technology and can process and integrate data from different dynamic motion analysis data. This includes unification of data formats, data mapping and registration between different dynamic motion analysis data.

And a plurality of image differential processing algorithms are arranged in the module and are used for analyzing and processing different types of data. For example, the spatial distribution data is processed by a special spatial analysis algorithm to extract the land utilization pattern and the land expansion trend.

The image difference processing module 102 is connected with the image acquisition module 101 and is configured to perform preprocessing on the acquired natural resource image identification related data, wherein the preprocessing comprises data cleaning, format conversion and data registration, and an image processing algorithm is used for analyzing the three-dimensional image and extracting information comprising land coverage type characteristics and land utilization modes.

The image difference processing module 102 plays a vital role in a moving object detection device based on natural resource images. It is responsible for efficient and accurate processing of the multiple dynamic motion analysis data received from the image acquisition module to ensure that the data is available for subsequent motion scene reconstruction.

The module includes a data cleansing subsystem that is responsible for removing invalid, erroneous, or irrelevant information from the raw data collected. Advanced data cleaning algorithms, such as outlier detection, de-duplication and data normalization techniques, are used to ensure data accuracy and consistency.

The module includes a format conversion subsystem for converting data in different formats (e.g., GIS data, three-dimensional image data, and statistics) into a unified format for data registration and analysis.

The module includes a data registration subsystem, which is the core of the module, that combines data from different dynamic motion analyses to create a comprehensive, multi-dimensional dataset. The subsystem adopts a data registration technology, such as entity identification and association analysis, so as to ensure that information of the same entity in different dynamic motion analysis data can be accurately matched.

The data processing module comprises a special image processing subsystem for processing the three-dimensional image. This subsystem employs advanced image recognition and analysis techniques, such as Convolutional Neural Networks (CNNs), to identify and classify surface features. After image processing, the subsystem can extract key information from the three-dimensional image, such as land cover type and land use pattern. This process involves complex image parsing techniques and geospatial data analysis.

The image processing subsystem executes an algorithm called a multi-dimensional spatial feature registration algorithm. The method combines deep learning, pattern recognition and multidimensional data registration technology and is used for efficiently and accurately processing and analyzing the three-dimensional image. The following are the execution steps of the algorithm:

step S101, data preprocessing: first, a three-dimensional image is subjected to normalization processing including color correction, denoising, and contrast adjustment to improve image quality.

Step S101 is an initial stage of a multidimensional spatial feature registration algorithm that focuses on converting the original three-dimensional image into a form more suitable for subsequent processing. The following is a detailed description of this step:

data preprocessing is used to improve the quality of three-dimensional images to facilitate more accurate feature recognition and analysis. The original three-dimensional image may be affected by various factors such as atmospheric conditions, equipment limitations, etc., resulting in poor image quality. Preprocessing is the basis for ensuring the accuracy of the subsequent steps.

The method specifically comprises the following substeps:

step S101-1, color correction:

the RGB channels of the image are adjusted using color balance techniques to eliminate color shift.

And adjusting according to the characteristics of the three-dimensional camera and the illumination condition during image acquisition.

Step S101-2, denoising:

a high pass filter is applied to remove random noise in the image.

Pixels affected by noise are identified and repaired using pixel neighborhood analysis.

Step S101-3, contrast adjustment:

the contrast of the image is enhanced by histogram equalization.

The gray scale range is adjusted so that the bright and dark details of the image are more apparent.

The preprocessing step may be performed using image processing software such as OpenCV, MATLAB or the like.

The output of this step is a three-dimensional image of improved quality, with better visual effect, less noise and more accurate color expression. These processed images will be used for motion feature hierarchy separation in step S102.

Step S102, motion feature hierarchical separation: the images are initially analyzed using a deep learning model, such as a modified Convolutional Neural Network (CNN). Unlike conventional CNNs, the model decomposes the image into multiple motion feature layers, such as texture, color, and shape layers.

This step is a key part of the multidimensional spatial feature registration algorithm that uses a deep learning model to process and analyze three-dimensional images. This step breaks down the three-dimensional image into different layers of motion characteristics in order to more deeply analyze and understand the image content. Different motion feature layers provide different information perspectives, helping to more fully understand geographic and land utilization features in an image.

The design of the deep learning model aims at optimizing the characteristics of the three-dimensional image and realizing effective separation and analysis of different motion feature layers.

The structure and components of the deep learning model include:

(1) Infrastructure architecture

And (3) architecture design: the model adopts a multi-branch architecture, which includes a backbone network and three specialized feature extraction subnetworks.

Backbone network: the backbone network uses a deep convolutional neural network, such as modified ResNet or DenseNet, for extracting the fundamental features of the image.

(2) Feature extraction subnetwork

(2a) Texture subnetwork:

the structure is as follows: multiple multi-scale convolution layers are included, each layer using a different size convolution kernel.

The functions are as follows: texture features ranging from fine granularity to coarse granularity can be captured.

Special components: attention mechanisms are introduced to enhance the critical motion features of textures.

(2b) Color subnetwork:

the structure is as follows: including convolutional layers based on color space conversion, such as RGB to HSV conversion layers.

The functions are as follows: focus is on analyzing the color distribution and variation of the image.

Special components: a color histogram equalization layer is used to enhance the color features. The layer is used as a component in a deep learning model for enhancing the color characteristics of an input three-dimensional image;

the layer receives as input a three-dimensional image in RGB format, wherein the image is represented as a multi-dimensional tensor; converting an input RGB image into an HSV color space by adopting an image processing algorithm, particularly into three channels of H (hue), S (saturation) and V (brightness) in an HSV model; applying a color histogram equalization process to the V-channels in the HSV color space to enhance the contrast and color characteristics of the image, wherein the histogram equalization is intended to improve the global contrast of the image, especially in case of backlight or low contrast; converting the HSV image subjected to histogram equalization processing back to an RGB color space for use in a subsequent step of the deep learning model; the layer implements the color space conversion and histogram equalization processes described above using an image processing library such as OpenCV, and then converts the processed image back into a multidimensional tensor format suitable for a deep learning model.

(2c) Shape subnetwork:

the structure is as follows: comprising a sequence convolution layer and an edge detection layer.

The functions are as follows: is used for extracting the outline and structure information of the ground object.

Special components: in combination with shape descriptors and geometry analysis algorithms. The shape descriptor and the geometric analysis algorithm can accurately extract the shape characteristics of the ground objects in the three-dimensional image, such as building outlines, road lines, water boundaries and the like. Shape information is crucial for understanding land use patterns and identifying different features. It helps to more accurately classify and understand the geographic motion features in an image.

The implementation steps of the algorithm comprise:

edge detection:

advanced edge detection techniques, such as gradient-based methods (e.g., sobel or Canny operators), are used to identify edge information in the image. Edge detection aims at highlighting the outline and boundary of a feature.

Shape descriptor extraction:

shape descriptors, such as fourier descriptors or Hu moments, are applied to quantify the detected edges. These descriptors can capture basic features of shape, such as size, orientation, and symmetry, and have invariance to image rotation, scaling, and translation.

Geometric analysis:

geometric analysis algorithms are implemented to further parse the shape. This may include calculating the area, perimeter, compactness and orientation of the geometry.

For complex shapes, such as irregular natural terrain, advanced geometric metrics such as fractal dimension are used.

In the shape subnetwork, the use of the sequence convolution layer, edge detection layer, and shape descriptor and geometry analysis algorithms can be divided into the following steps:

(2c_1) application of sequential convolution layer: the sequential convolution layer is used mainly to process shape information in the image. These layers extract and highlight shape features in the image, such as the outline of a building, lines of a road, or boundaries of a body of water, etc., through a series of convolution operations.

(2c_2) combination with edge detection: in shape subnetworks, edge detection is typically a step after the sequence convolution layer. The edge detection layer, such as with Sobel or Canny operators, focuses on identifying contours and boundaries of features in the image. These identified edge information are then fed into the sequence convolution layer for further processing and analysis.

(2c_3) after edge detection is completed, the next step is to apply shape descriptors (e.g. fourier descriptors or Hu moments) to quantify the edges. This step helps capture basic features of the shape such as size, orientation, and symmetry. These descriptors can allow the model to better understand and classify features in the image.

(2c_4) geometric analysis: next, geometric analysis is performed to further analyze the characteristics of the shape. This may include calculation of features of area, perimeter, compactness, etc. of the shape. For more complex shapes, such as natural terrain, it is also possible to use more advanced geometric metrics, such as fractal dimension.

Shape classification:

the extracted shape features are input into a classifier, such as a Support Vector Machine (SVM) or random forest, for feature classification.

Combining shape features with other features (such as texture and color) improves classification accuracy.

Network architecture relationship:

the original image first enters the backbone network for basic feature extraction, after which the basic features are passed to three feature extraction sub-networks. Each sub-network performs further specialized processing on the basic features, focusing on specific aspects such as texture, color, or shape. The backbone network works in concert with the three sub-networks to ensure that the image is analyzed in its entirety from multiple angles. The generic features provided by the backbone network provide the necessary inputs for the specific analysis of the subnetworks.

(3) Model training and tuning

Data set: three-dimensional image datasets with geographic markers are used, including different terrain, land cover, and land use patterns.

The training strategy comprises the following steps:

pre-training: the backbone network is pre-trained on a large-scale generic image dataset.

Fine tuning: each sub-network is trained and trimmed separately on the three-dimensional image dataset.

The optimization technology comprises the following steps:

data enhancement: the generalization capability of training is enhanced by applying random rotation, scaling, flipping and other techniques.

Regularization: overfitting is avoided using Dropout and weight decay.

To implement such a deep learning model, a deep learning framework such as pyrerch or TensorFlow needs to be employed. The following is an example code that uses the PyTorch framework to build such a model. Note that this is just a conceptual example and further adjustments and optimizations are required in practical applications.

import torch

import torch.nn as nn

import torchvision.models as models

import cv2

import numpy as np

class RGBtoHSVLayer(nn.Module):

def __init__(self):

super(RGBtoHSVLayer, self).__init__()

def forward(self, x):

Note #: here, the cv2 library is used for color space conversion

The # OpenCV function requires that the input be a numpy array

x_np = x.detach().cpu().numpy()

x_hsv = np.array([cv2.cvtColor(img, cv2.COLOR_RGB2HSV) for img in x_np])

Convert results back to PyTorch tensor #

x_hsv_tensor = torch.from_numpy(x_hsv).to(x.device)

return x_hsv_tensor

# color histogram equalization layer

class ColorHistogramEqualizationLayer(nn.Module):

def __init__(self):

super(ColorHistogramEqualizationLayer, self).__init__()

def forward(self, x):

# histogram equalization Using cv2

x_np = x.detach().cpu().numpy()

x_eq = np.array([cv2.equalizeHist(np.uint8(img)) for img in x_np])

x_eq_tensor = torch.from_numpy(x_eq).to(x.device).float()

return x_eq_tensor

# shape descriptor and geometry analysis

class ShapeDescriptorLayer(nn.Module):

def __init__(self):

super(ShapeDescriptorLayer, self).__init__()

def forward(self, x):

# use the Hu moment of OpenCV as shape descriptor

x_np = x.detach().cpu().numpy()

hu_moments = np.array([cv2.HuMoments(cv2.moments(img)).flatten() for img

in x_np])

hu_moments_tensor = torch.from_numpy(hu_moments).to(x.device)

return hu_moments_tensor

class MultiBranchModel(nn.Module):

def __init__(self, num_classes):

super(MultiBranchModel, self).__init__()

# backbone network

self.backbone = models.resnet50(pretrained=True)

# texture subnetwork

self.texture_branch = nn.Sequential(

nn.Conv2d(2048, 512, kernel_size=3, stride=1, padding=1),

nn.ReLU(),

Attention adding mechanism

nn.Conv2d(512, 512, kernel_size=3, stride=1, padding=1),

nn.ReLU(),

nn.AdaptiveAvgPool2d((1,1))

)

# color subnetwork

self.color_branch = nn.Sequential(

RGBtoHSVLayer(),

nn.Conv2d(2048, 512, kernel_size=3, stride=1, padding=1),

nn.ReLU(),

ColorHistogramEqualizationLayer(),

nn.AdaptiveAvgPool2d((1,1))

)

# shape subnetwork

self.shape_branch = nn.Sequential(

nn.Conv2d(2048, 512, kernel_size=3, stride=1, padding=1),

nn.ReLU(),

ShapeDescriptorLayer(),

nn.AdaptiveAvgPool2d((1,1))

)

# Classification layer

self.classifier = nn.Linear(512 * 3, num_classes)

def forward(self, x):

# extraction of backbone network features

x = self.backbone(x)

# extracting features of different motion feature layers

texture_features = self.texture_branch(x)

color_features = self.color_branch(x)

shape_features = self.shape_branch(x)

# merging feature

combined_features = torch.cat((texture_features, color_features, shape_features),

dim=1)

combined_features = combined_features.view(combined_features.size(0), -1)

Class #

out = self.classifier(combined_features)

return out

Use example #

# model = MultiBranchModel(num_classes=10)

# output = model(input_tensor)

Note #: this code provides a conceptual example and may need to be adjusted and optimized according to specific needs.

Step S103, extracting multidimensional space features: at each motion feature level, multi-dimensional image differential processing techniques are employed to extract key information. For example, identifying texture features of different land coverage, such as bodies of water, woodland, and bare land, on a texture layer; analyzing color features of different land types on the color layer; shape information of ground buildings, roads, etc. is extracted on the shape layer.

This step extracts meaningful information from each motion feature hierarchy (texture, color, shape) to facilitate more in depth analysis of the three-dimensional image. This step helps to obtain rich geographic and land use information from the three-dimensional image, which is critical to accurately identifying and classifying land cover types and land use patterns.

(1) Texture feature extraction:

the method comprises the following steps:

adaptive filter design: an adaptive filter network based on deep learning is designed, and the filter parameters of the network can be dynamically adjusted according to different characteristics (such as illumination conditions and ground object types) of an input image.

Local response normalization: local contrast of texture features in an image is enhanced using local response normalization techniques, making subtle texture differences more pronounced.

The realization is as follows:

an adaptive filter network is constructed using a deep learning framework such as PyTorch or TensorFlow.

In a network, multiple convolution layers are used to achieve different scale texture feature capture.

A local response normalization layer is applied after the convolution layer, such as a torch.nn.localresponsenram or corresponding TensorFlow implementation.

(2) Color feature extraction:

the method comprises the following steps:

color coding techniques: a color coding technique is designed that captures color information comprehensively by analyzing the spatial relationship between the color histogram and the pixels.

Clustering algorithm: different color groups are identified and encoded using an adaptive K-means clustering algorithm.

The realization is as follows:

the input image is subjected to color histogram analysis using an image processing library (e.g., openCV).

An adaptive K-means clustering algorithm is implemented, or clustering algorithms in existing machine learning libraries, such as scikit-learn kmens, are used.

And combining the histogram and the clustering result to generate the color feature code.

(3) And (3) extracting shape features:

the method comprises the following steps:

image segmentation and shape recognition: the method comprises the steps of accurately separating a target ground object from a complex background and extracting shape characteristics of the target ground object by using a Full Convolution Network (FCN) based on deep learning and combining a traditional image segmentation algorithm (such as GrabCt).

The realization is as follows:

the FCN model is implemented using a deep learning framework (e.g., pyrerch or TensorFlow) for the image segmentation task.

The output of the FCN is post-processed using the GrabCut algorithm (which may be implemented using OpenCV) to improve the accuracy of the segmentation results.

Shape information is extracted from the segmentation result, for example, using a contour detection algorithm to extract boundaries.

The reference implementation code for this step is as follows.

import torch

import torch.nn as nn

import cv2

import numpy as np

import torchvision.models as models

from sklearn.cluster import KMeans

# adaptive filter network

class AdaptiveFilterNetwork(nn.Module):

def __init__(self):

super(AdaptiveFilterNetwork, self).__init__()

self.conv1 = nn.Conv2d(3, 64, kernel_size=3, stride=1, padding=1)

self.lrn = nn.LocalResponseNorm(size=5, alpha=0.0001, beta=0.75, k=2)

Adding more layers to achieve multi-scale texture feature capture

def forward(self, x):

x = self.conv1(x)

x = self.lrn(x)

Treatment of other layers

return x

# color feature extraction

def extract_color_features(image):

Conversion of # to HSV color space

hsv_image = cv2.cvtColor(image, cv2.COLOR_RGB2HSV)

# calculate color histogram

hist = cv2.calcHist([hsv_image], [0, 1, 2], None, [256, 256, 256], [0, 256, 0, 256,

0, 256])

hist = hist.flatten()

# apply KMeans clustering

kmeans = KMeans(n_clusters=10)

kmeans.fit(hist.reshape(-1, 1))

return kmeans.cluster_centers_

# shape feature extraction

def extract_shape_features(image, model):

Image segmentation using FCN model #)

segmented = model(image)

Improved by # application of GrabCut algorithm

mask = np.zeros(segmented.shape[:2], np.uint8)

bgdModel = np.zeros((1,65), np.float64)

fgdModel = np.zeros((1,65), np.float64)

cv2.grabCut(segmented, mask, None, bgdModel, fgdModel, 5, cv2.GC_

INIT_WITH_MASK)

# extraction contour

contours, _ = cv2.findContours(mask, cv2.RETR_TREE, cv2.CHAIN_

APPROX_SIMPLE)

return contours

Example usage #

Creation of adaptive filter network instances #

adaptive_filter_net = AdaptiveFilterNetwork()

# assume that image is an input three-dimensional image

# image = ...

# extraction of texture features

texture_features = adaptive_filter_net(image)

# extraction of color features

color_features = extract_color_features(image.numpy())

# extract shape features (requiring a pre-trained FCN model)

# fcn_model = ...

# shape_features = extract_shape_features(image.numpy(), fcn_model)

Note #: these functions need to be adjusted and perfected according to actual conditions

In the above code, an adaptive filter network is implemented to extract texture features, a function is defined to extract color features, and a functional framework is provided to extract shape features. These implementations involve knowledge in the fields of deep learning, traditional image processing, and machine learning. In practical applications, the network structure, parameters and algorithms need to be adjusted according to specific data and requirements.

Step S104, feature registration and classification: the features of different layers are intelligently registered, and a graph-based registration algorithm is used, so that the algorithm can maximize comprehensive feature information while keeping the feature independence of each layer. The registered features are used to classify and identify land cover types and land use patterns.

This step effectively registers the multi-dimensional features extracted from the texture, color and shape hierarchy and uses these registered features for accurate classification. Proper registration of multi-dimensional features is critical to improving classification accuracy, especially when processing three-dimensional images with complex geographic and land-use features.

The feature registration framework includes:

(1) Multidimensional feature vector generation:

the features (texture, color, shape) extracted from each motion feature hierarchy are converted into a consistent feature vector format.

The dimensions of the different features are normalized using a fully connected or pooled layer in deep learning.

(2) Graph-based registration algorithm:

a graph-based feature registration algorithm is developed that treats the different motion feature layers as nodes in the graph and represents the correlation between features by edge weights.

Using a Graph Neural Network (GNN) to achieve this registration allows the model to dynamically adjust the registration intensity between different features.

To implement a graph-based feature registration algorithm and use a Graph Neural Network (GNN) for feature registration, the following steps may be performed:

(2a) Definition map structure

Creation of a node: in the graph network, each motion feature layer (e.g., texture layer, color layer, shape layer) is considered a node.

Edge establishment: and establishing edges among the nodes according to the correlation among the motion characteristic layers. The weights of the edges reflect the degree of correlation or influence between the different motion characteristics. These weights may be derived based on a priori knowledge, data analysis, or learning.

(2b) Application of Graph Neural Network (GNN)

GNN architecture selection: an appropriate GNN architecture is selected, such as Graph Convolutional Network (GCN) or Graph Attention Network (GAT).

The characteristic is represented as follows: a feature vector is input for each node (motion feature layer). This vector may be the feature extraction result from the previous deep learning stage.

Adjacency matrix: an adjacency matrix is created to represent the connection relationships between nodes in the graph and their weights.

(2c) Feature registration

Graph convolution layer: graph roll stacking is used in GNN to handle the characteristics of nodes. The graph convolution can aggregate and transform features of neighboring nodes based on information in the adjacency matrix.

Information aggregation: the feature information from neighboring nodes is combined using an aggregation function (e.g., summation, averaging, or maximization) in the GNN.

Feature update: the feature representation of each node is updated by the graph convolution layer to reflect the information of the neighboring nodes.

(2d) Output of feature registration

Registration feature extraction: the registered feature representation is extracted from the last layer of GNN. These features will collectively reflect the information in the original motion feature layer.

Applied to downstream tasks: the registered features are used for downstream tasks such as classification, regression or other analysis.

(3) Classification algorithm

Deep learning classifier:

a deep neural network is designed as a classifier, which takes as input the registered multidimensional features.

A common classification network structure, such as a multi-layer perceptron (MLP), is used and a nonlinear activation function is applied thereto.

(4) Training strategies:

the classifier is trained using cross entropy loss functions for different land cover types and land use patterns.

Data enhancement and regularization techniques (e.g., dropout) are applied to enhance the generalization ability of the classifier.

The following is the reference implementation code of this step:

import torch

import torch.nn as nn

import torch.nn.functional as F

import torch.optim as optim

from torch_geometric.nn import GCNConv

assume the dimensions of texture, color and shape features are 256

feature_dimension = 256

# defines a simple full-join layer to normalize the dimension of feature vectors

class FeatureTransformation(nn.Module):

def __init__(self, input_dim, output_dim):

super(FeatureTransformation, self).__init__()

self.fc = nn.Linear(input_dim, output_dim)

def forward(self, x):

x = F.relu(self.fc(x))

return x

# definition map neural network model

class GNN(nn.Module):

def __init__(self, feature_dim, num_classes):

super(GNN, self).__init__()

self.conv1 = GCNConv(feature_dim, 128)

self.conv2 = GCNConv(128, 64)

self.fc = nn.Linear(64, num_classes)

def forward(self, x, edge_index):

x = F.relu(self.conv1(x, edge_index))

x = F.dropout(x, training=self.training)

x = self.conv2(x, edge_index)

x = F.dropout(x, training=self.training)

x = self.fc(x)

return F.log_softmax(x, dim=1)

Example #: texture, color and shape features

texture_features = torch.rand(10, feature_dimension)

color_features = torch.rand(10, feature_dimension)

shape_features = torch.rand(10, feature_dimension)

# initialization feature translation layer

feature_transform = FeatureTransformation(feature_dimension, feature_dimension)

# conversion feature vector

transformed_texture_features = feature_transform(texture_features)

transformed_color_features = feature_transform(color_features)

transformed_shape_features = feature_transform(shape_features)

# merging feature vector

combined_features = torch.cat([transformed_texture_features,

transformed_color_features,

transformed_shape_features], dim=1)

Picture structure of# hypothesis

num_nodes = combined_features.shape[0]

edge = torch. Randint (0, num_nodes, (2, 2 x num_nodes)) # randomly generated edges

# initializing GNN model

GNN =gnn (feature_dimension, num_categories=5) # assume 5 categories

Forward propagation of #

output = gnn(combined_features, edges)

Model output #

print("Model Output:", output)

In practical use of the reference code, the network also requires an appropriate training process, including selecting optimizers, defining loss functions, and training loops.

Step S105, geospatial data mapping: the extracted information is mapped and collated with Geographic Information System (GIS) data to ensure that the geographic motion characteristics match specific locations in the real world.

In this embodiment, the geographic motion features further include key information extracted by using a multidimensional image differential processing technique on each motion feature layer, for example, texture features of different land coverage identified on texture layers, including water, woodland and bare land; color features of different land types identified on the color layer; and the shape information of the ground building and the road is extracted from the shape layer.

This step is the last step of the multidimensional spatial feature registration algorithm. In this step, efficient mapping and verification of the extracted feature information with Geographic Information System (GIS) data is required. The following is a detailed description of this step and its embodiments:

Geospatial data mapping is used to ensure that the geographic motion features extracted by the algorithm match specific locations in the real world. Accurate geospatial data mapping is critical for applications of geographic information systems such as earth image recognition, environmental monitoring, and the like.

The technical implementation includes:

(1) Features correspond to GIS data:

each feature (e.g., building, road, body of water, etc.) in the classification result is assigned a unique identifier.

These identifiers are associated with geographic location information in a GIS database.

(2) Spatial collation and mapping:

the extracted features are converted to geographic coordinates (e.g., latitude and longitude) using geocoding techniques.

Spatial correction methods, such as coordinate conversion and projection correction, are applied to ensure spatial consistency of features with the GIS data.

(3) Data integration:

and integrating the extracted characteristic information with the existing GIS data to construct a comprehensive geospatial database.

The characteristic information (such as texture, color, shape) and geographic location are recorded for each feature in the database.

The following is a simplified example of a geospatial data mapping using Python and GIS tools:

import geopandas as gpd

from shapely.geometry import Point

def map_features_to_gis(feature_data, gis_data):

"""

the extracted features are mapped to GIS data.

Param feature_data, dataFrame containing feature information.

Paramgis_data GeoDataFrame containing GIS information.

Return: mapped GeoDataFrame.

"""

mapped_data = gpd.GeoDataFrame()

for index, row in feature_data.iterrows():

# assume feature_data contains latitude and longitude information

point = Point(row['longitude'], row['latitude'])

# find the ground object matching with this point in the GIS data

matched_feature = gis_data[gis_data.geometry.contains(point)]

if not matched_feature.empty:

Combining feature information and geographic information #

combined_data = {**row.to_dict(), **matched_feature.iloc[0].to_dict()}

mapped_data = mapped_data.append(combined_data, ignore_index=True)

return mapped_data

Example usage #

Feature_data=. The feature information extracted by # feature_data=

# GIS _data=. The data of the gis# GIS

# mapped_data = map_features_to_gis(feature_data, gis_data)

This example uses the geopladas library to process geospatial data, which is a common tool for processing geographic data in Python. In practical applications, this step may involve complex spatial analysis and extensive image differential processing work, requiring GIS expertise. This example is only a conceptual implementation, and the mapping process in practical application may be more complex, involving the problems of conversion of different coordinate systems, integration of multiple dynamic motion analysis data, etc.

An efficient data transmission interface is arranged between the image difference processing module and the motion scene reconstruction module. The interface ensures that the processed data can be seamlessly transmitted to a scene reconstruction module for further analysis and scene reconstruction.

The processed data will be formatted into the structure required by the scene reconstruction module to ensure that the data can be properly interpreted and analyzed by the scene reconstruction algorithm.

The motion scene reconstruction module 103 is connected with the image differential processing module 102, and is configured to reconstruct a motion scene of the data provided by the image differential processing module according to a set scene reconstruction model and algorithm, wherein the scene reconstruction model comprehensively considers a plurality of dimensions including spatial distribution, color indexes, ecological environment, geographic motion characteristics and land utilization modes.

The motion scene reconstruction module specifically realizes motion scene reconstruction by the following steps:

step S301, data preparation:

the data of this step is derived from the processed data obtained from the image difference processing module 102, which may include spatial coordinates, color data, environmental indicators, etc., covering the various dimensions of the image recognition. These data are preprocessed to ensure that the data format is suitable for the input of the graph network model. This may include normalization, etc. of the data.

Step S302, constructing a graph network model:

this step automatically determines the relationships and interactions (weights of edges) between the image recognition dimensions (nodes) based on the data analysis.

The input of this step comes from step S301, including spatial distribution, color index, ecological environment, geographic motion characteristics, land utilization pattern, and the like.

The implementation method specifically comprises the following steps:

(1) And (3) node establishment:

each image recognition dimension is defined as a node in the graph network. The image recognition dimensions may include spatial distribution, color index, ecological environment, land cover type features, land use patterns, and the like.

For example, a node may be a "spatial distribution", "color index", "ecological environment", or the like.

(2) Automatic construction of edges:

correlation analysis:

the pearson correlation coefficient between the dimensions is calculated.

Edges are established between image recognition dimensions for which the correlation coefficient is above a predetermined threshold (e.g., 0.5).

Analysis of the degree of influence:

the importance of each image recognition dimension is analyzed using a machine learning model such as random forest. Since each node corresponds to one image recognition dimension, the importance of the image recognition dimension also corresponds to the importance of the node.

The weight of an edge may be determined based on the importance calibration values of the two nodes. One approach is to take the average of the importance calibration values of two nodes of an edge as the weight of the edge.

(3) Dynamic adjustment of edge weights:

the dataset is analyzed using a supervised learning algorithm (e.g., logistic regression) to determine the optimal edge weights. The datasets here typically originate from several aspects:

historical image identification data: this may include data collected in past natural resource image identification projects. These data reflect the behavior and interactions of different image recognition dimensions in past projects.

And (3) real-time data collection: if the item relates to currently ongoing image recognition, information may be collected from real-time dynamic motion analysis data, such as ground sensor networks, color activity monitoring systems, and the like.

Expert investigation and opinion: data obtained by the survey and interview field experts can be used to supplement and verify data obtained from historical and public databases.

In particular, this dataset should contain information in multiple dimensions, each dimension representing an aspect of image recognition. Each record in the dataset may contain specific values for these dimensions, as well as the performance of these dimensions in the actual image recognition case (e.g., cases of success or failure). Using this data, a supervised learning algorithm can learn interactions between different dimensions to help determine optimal weights for edges in the graph network model.

The edge weights are adjusted and verified by cross-validation or the like.

(4) Model verification:

and comparing and verifying the constructed graph network model with actual cases or historical data.

The prediction accuracy and reliability of the model are analyzed.

The reference code for constructing the graph network model is as follows:

import numpy as np

import pandas as pd

from sklearn.ensemble import RandomForestRegressor

from sklearn.model_selection import train_test_split, cross_val_score

import networkx as nx

import matplotlib.pyplot as plt

example data #

Suppose there is a set of image identification data including indices of spatial distribution, ecological environment, geographic motion characteristics, etc

data = {

'spatial distribution' [0.9, 0.8, 0.7, 0.6, 0.9],

'ecological environment' [0.8, 0.7, 0.6, 0.5, 0.4],

'geographic motion characteristics': 0.3, 0.4, 0.5, 0.6, 0.7

}

df = pd.DataFrame(data)

# calculate correlation matrix

correlation_matrix = df.corr()

# creation map

G = nx.Graph()

# adding node

for column in df.columns:

G.add_node(column)

# added edge

threshold=0.5# threshold

for i in range(len(correlation_matrix.columns)):

for j in range(i):

if abs(correlation_matrix.iloc[i, j])>threshold:

G.add_edge(correlation_matrix.columns[i], correlation_matrix.columns[j],

weight=correlation_matrix.iloc[i, j])

# estimation of feature importance Using random forest

X = df.values

y=np.random.rand (x.shape [0 ])# hypothetical target variable

rf = RandomForestRegressor()

rf.fit(X, y)

importances = rf.feature_importances_

# importance as node attribute

for i, column in enumerate(df.columns):

G.nodes[column]['importance']= importances[i]

Adjusting edge weights using supervised learning

In the example # only the training process is calculated, which in practice needs to be implemented according to a specific data set and predicted targets

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.3)

model = RandomForestRegressor()

model.fit(X_train, y_train)

Cross validation #

scores = cross_val_score(model, X, y, cv=5)

# display figure

pos = nx.spring_layout(G)

edge_labels = nx.get_edge_attributes(G, 'weight')

nx.draw(G, pos, with_labels=True, node_color='skyblue', node_size=2000)

nx.draw_networkx_edge_labels(G, pos, edge_labels=edge_labels)

plt.show()

# output model verification results

scores.mean(), scores.std()

Step S303, playground Jing Jiao standard:

Motion scene calibration values are calculated based on interaction strength and pattern between nodes in the graph network. The interaction strength is determined by analyzing the weights of the inter-node edges in the graph network. The higher the weight of an edge, the greater the interaction between two nodes, and the stronger the interaction strength. The interaction pattern is determined according to the connection pattern between nodes, for example, if a node is highly connected to a plurality of other nodes, its centrality in the network is stronger and its importance in image recognition is also higher.

Considering both the interaction strength and the interaction pattern, an algorithm is used to calculate the sports field Jing Jiao guideline. This calibration value reflects the overall level of dynamic recognition between different dimensions in the image recognition scheme.

This calibration value reflects the overall dynamic recognition of the image recognition scheme. The output playfield Jing Jiao standard value is a quantization index for analyzing the comprehensive effect of the image recognition scheme.

The calculation formula and implementation steps are as follows:

(1) Interaction strength calculation:

weighting of edgesThe interaction strength for each edge is determined, wherein,andfor the connected first in the graph networkPersonal node and the firstAnd each node.

(2) Identification of interaction pattern:

Interaction patterns are analyzed using the centrality of the nodes.

Degree of centralityThe calculation formula of (2) is as follows:

wherein,is a nodeI.e., the number of connected edges), N is the total number of nodes in the graph network.

(3) Motion scene calibration values:

the motion scene calibration value is calculated by combining the interaction strength and the node degree centrality. The calculation formula of the motion scene calibration value S is as follows:

wherein,representing nodesIs defined in the network, is a neighbor node to all neighboring nodes of the network.

Step S304, analyzing a scene reconstruction result:

the data analysis method is applied:

the stadium Jing Jiao quasi-values are grouped and interpreted using underlying data analysis techniques, such as simple cluster analysis.

By this analysis, a pattern of calibration values can be identified, for example which dimensions often occur together in high or low calibration values.

Potential problem identification:

by comparing the calibration values of the different dimensions, areas of conflict or incompatibility that may exist are identified.

For example, if a dimension is often associated with a low calibration value, it may be indicated that particular attention is required for that field.

Optimization suggestion formulation:

based on the analysis results, specific improvements or suggestions are made.

This may include adjusting emphasis of a particular dimension in the image recognition scheme or proposing more specific strategies to improve dynamic recognition.

Step S305: results application and iteration

And the scene reconstruction result is used for guiding the actual natural resource image recognition decision. And continuously adjusting and optimizing the scene reconstruction model according to feedback in practical application so as to improve the accuracy and applicability of the scene reconstruction model.

The moving object confirmation module 104 is connected to the moving scene reconstruction module 103, and is configured to calculate a result of the moving scene reconstruction in the form of a moving object analysis model, where the moving object analysis model includes a map moving range model, a moving speed model and a moving density model, and supports a user to customize display content according to requirements.

The moving object confirmation module 104 is a key component of the moving scene reconstruction device of the natural resource image identification data.

The module is used for calculating the exercise scene reconstruction result, so that a decision maker and an analyst can intuitively understand and analyze various aspects of natural resource image recognition. The result computes a comprehensive scene reconstruction including, but not limited to, spatial distribution, color index, ecological environment, etc.

The module also calculates the space distribution and the geographic motion characteristics by using a Geographic Information System (GIS) technology, and calculates the image recognition conditions of different areas on a map, such as land utilization modes, image recognition effects and the like.

In addition, the module calculates statistical data and color index using a motion velocity model (e.g., bar graph, line graph, pie chart). Complex data, such as distribution of motion scene calibration values, correlation between dimensions, etc., are calculated using data visualization techniques. The module also provides detailed analysis motion density models including interpretation of scene reconstruction results, recommended image recognition strategies, and the like. The motion density model supports a deriving function, and printing or electronic sharing is facilitated.

The module allows the user to customize the display content to his own needs, such as selecting a particular data dimension for calculation. Providing interactive functions, such as clicking on an area of a map, may view more detailed information.

The system not only displays the current standard value of the sports field Jing Jiao, but also calculates the calibration value of the predicted sports scene after the adjustment of different image recognition schemes, thereby helping the user to understand the long-term influence of different decisions. For example, if the user considers adding a business in the ground center, the system may predict the potential impact of this change on ground traffic flow, environmental quality, and color growth and calculate the corresponding stadium Jing Jiao guideline change.

The interactive dynamic recognition degree prediction simulation system can provide simulation prediction based on the actual situation according to the historical data and the trend of the actual ground image recognition, so that the simulation is closer to the actual image recognition scene. The specific steps for implementing the function include:

(1) Historical data collection and analysis:

historical data related to ground image identification is collected including, but not limited to, land use, color change data, environmental quality metrics, traffic flow, and the like.

The data is analyzed to identify relationships and trends between image recognition decisions and ground changes.

(2) Model building and training:

a machine learning model, such as a random forest or neural network, is constructed based on the collected historical data to predict the potential impact of different image recognition decisions.

Training the model using historical data ensures that it can accurately identify and predict the consequences of image recognition decisions.

A second embodiment of the present application provides a data motion scene reconstruction method for natural resource image recognition, including:

collecting natural resource image recognition related data from a plurality of dynamic motion analysis data, wherein the dynamic motion analysis data comprises a geographic information system, a three-dimensional image and statistical data; the acquired data comprise a plurality of dimensions such as spatial distribution, color indexes, ecological environment and the like;

preprocessing the acquired data, including data cleaning, format conversion and data registration; processing the three-dimensional image by utilizing an image processing algorithm to extract information such as geographic motion characteristics, land utilization modes and the like;

Performing motion scene reconstruction on the preprocessed data and the image processed data according to a set scene reconstruction model and algorithm; the scene reconstruction model comprehensively considers a plurality of image recognition dimensions including spatial distribution, color indexes, ecological environment, geographic motion characteristics and land utilization modes;

calculating a result of the motion scene reconstruction through a motion target analysis model, wherein the result comprises a map motion range model, a motion speed model and a motion density model; and provides the function of customizing the display content according to the requirements of the user.

A third embodiment of the present application provides an electronic device, including:

a processor;

and the memory is used for storing a program which, when being read and executed by the processor, executes the data motion scene reconstruction method for natural resource image recognition provided in the second embodiment of the application.

A fourth embodiment of the present application provides a computer-readable storage medium having stored thereon a computer program which, when executed by a processor, performs the data motion scene reconstruction method for natural resource image recognition provided in the second embodiment of the present application.

While the preferred embodiment has been described, it is not intended to limit the invention thereto, and any person skilled in the art may make variations and modifications without departing from the spirit and scope of the present invention, so that the scope of the present invention shall be defined by the claims of the present application.

Claims

1. The moving target detection device based on the natural resource image is characterized by comprising an image acquisition module, an image difference processing module, a moving scene reconstruction module and a moving target confirmation module;

the moving object confirming module is connected with the moving scene reconstructing module and is configured to calculate a moving scene reconstructing result in the form of a moving object analyzing model, wherein the moving object analyzing model comprises a map moving range model, a moving speed model and a moving density model;

the motion scene reconstruction module specifically realizes motion scene reconstruction through the following steps:

acquiring processed data from an image differential processing module, wherein the processed data comprise spatial distribution, color indexes, ecological environment, geographic motion characteristics and land utilization modes, and the data format is ensured to be suitable for the input of a graph network model after preprocessing;

constructing a graph network model: automatically determining the relationship and interaction between the image recognition dimensions, namely the weight of the edges, according to the data analysis;

defining each image recognition dimension as a node in the graph network, wherein the image recognition dimension comprises spatial distribution, color indexes, ecological environment, land coverage type characteristics and land utilization modes;

Automatically constructing edges by adopting correlation analysis and influence degree analysis;

analyzing the data set by using a supervised learning algorithm to determine an optimal edge weight;

model verification: comparing and verifying the constructed graph network model with actual cases or historical data;

the moving target confirming module utilizes a moving speed model, comprising a histogram, a line graph or a pie chart, calculates statistical data and color indexes, uses a data visualization technology to calculate complex data, comprises distribution of a moving scene calibration value and correlation among dimensions, and provides an analysis moving density model, comprising interpretation of a scene reconstruction result and a recommended image recognition strategy.

2. The apparatus of claim 1, wherein the image difference processing module comprises an image processing subsystem for processing a three-dimensional image, the image processing subsystem being specifically configured to:

3. The apparatus of claim 2, wherein the image processing subsystem comprises a deep learning model for decomposing the image into a plurality of motion feature layers, the deep learning model comprising a backbone network and a feature extraction sub-network, wherein the backbone network uses a deep convolutional neural network for extracting the fundamental features of the image; the feature extraction sub-network comprises a texture sub-network, a color sub-network and a shape sub-network; the texture subnetwork comprises a plurality of convolution layers for capturing texture features from fine granularity to coarse granularity; the color sub-network comprises a convolution layer based on color space conversion and is used for analyzing color distribution and change of an image; the shape subnetwork comprises a sequence convolution layer and an edge detection layer, and is used for extracting outline and structure information of the ground feature.

4. The apparatus of claim 3, wherein the texture subnetwork directs attention mechanisms to enhance key motion characteristics of the texture.

5. The apparatus of claim 3, wherein the color subnetwork uses a color histogram equalization layer to enhance color features, the color histogram equalization layer being specifically configured to:

receiving a three-dimensional image in an RGB format as an input;

6. The apparatus of claim 3, wherein the shape subnetwork comprises a component for performing shape descriptors and geometry analysis algorithms, the component configured to:

7. The apparatus according to claim 1, wherein the motion scene reconstruction module is specifically configured to:

each image recognition dimension is defined as a node in the graph network,

analyzing the importance of each image recognition dimension by using a random forest model, and determining the weight of the edge according to the importance;

calculating a motion field Jing Jiao standard value, representing the overall level of dynamic recognition degrees among different dimensions in the image recognition scheme and the overall dynamic recognition degree of the image recognition scheme, wherein the output motion field Jing Jiao standard value is a quantization index and is used for analyzing the comprehensive effect of the image recognition scheme;

and (3) analyzing a scene reconstruction result:

the data analysis method is applied: grouping and interpreting the stadium Jing Jiao standard values by using simple cluster analysis, and identifying a calibration value mode;

potential problem identification: identifying areas of conflict or incompatibility that exist by comparing calibration values of different dimensions;

Optimization suggestion formulation: according to the analysis result, specific improvement measures or suggestions are provided, wherein the specific improvement measures or suggestions comprise adjusting the key points of specific dimensions in the image recognition scheme or providing specific strategies to improve the dynamic recognition degree;

results application and iteration: and the scene reconstruction result is used for guiding the actual natural resource image recognition decision, and the scene reconstruction model is continuously adjusted and optimized according to feedback in the actual application.

8. The apparatus of claim 7, wherein the motion scene reconstruction module is specifically configured to:

wherein N is the aboveTotal number of nodes in the graph network;representing nodes in the graph networkIs defined by the first node and the second node;is the weight of an edge in the graph network, both sides of the edge include nodesSum node；Is a nodeFor analyzing interaction patterns between nodes in the graph network,the calculation formula of (2) is as follows:

wherein,is a nodeIs a degree of (3).