CN114037836A - Method for applying artificial intelligence recognition technology to three-dimensional power transmission and transformation engineering measurement and calculation - Google Patents

Abstract

The invention relates to power transmission and transformation project measurement and calculation, in particular to a method for applying an artificial intelligence recognition technology to three-dimensional power transmission and transformation project measurement and calculation. According to the invention, by automatically classifying the true-color DOM image and the point cloud data, elements such as houses and trees which have important influences on the power corridor of the power transmission line can be efficiently and accurately extracted, the workload of measuring personnel is greatly reduced, and the work of measuring and calculating the engineering quantity in the traditional mode can be completed only by fixed-point manual intervention. The high-precision boundary extraction result of the deep learning method is high, the target prediction result is compared with the label in a test set, the accuracy rate of the ground feature identification can reach more than 85%, the boundary segmentation can be basically accurately carried out on the classified ground features, and reliable basis is provided for the actual measurement and calculation work.

Description

Method for applying artificial intelligence recognition technology to three-dimensional power transmission and transformation engineering measurement and calculation

Technical Field

The invention relates to power transmission and transformation project measurement and calculation, in particular to a method for applying artificial intelligence recognition technology to three-dimensional power transmission and transformation project measurement and calculation.

Background

In the surveying industry, the traditional method for measuring and calculating the engineering quantity is to manually draw ground features such as houses, trees, expropriations and the like based on images, generate an ORG file, and import the ORG file into design software for measuring and calculating. There are generally four methods for estimation: the first method is based on full-waveform laser Detection And measurement (LiDAR, Light Detection And Ranging) data, structured information of vegetation is obtained, point cloud attributes are collected in an affected range, comprehensive waveform characteristic parameters are calculated, And measurement accuracy of a target size is obtained And improved through a single data source; the second method is that based on full-waveform LiDAR data, positioning and extraction of monomer trees, houses and land seeking are derived, and a Weber distribution and cubic spline function model are combined to be used as characteristic variables to obtain model parameters; the third method is based on multisource contemporaneous high-resolution remote sensing data, firstly, target classification is carried out based on the high-resolution data and an object-oriented segmentation method, and then, image ground objects are identified based on space details and spectral features extracted from the high-spectral data and by combining a back propagation neural network; and fourthly, recognizing the ground object based on the hyperspectral data by using a three-dimensional convolutional neural network, performing semantic segmentation on the complex ground object by using the three-dimensional convolutional neural network based on the airborne hyperspectral data, constructing a data set for training and verifying, completing the training of a deep learning model, and predicting the category of each pixel in the hyperspectral image by using the trained model to obtain the tree species distribution chart of the whole area.

This method is quite labor intensive, requires a lot of manpower and is long in cycle. For complex terrain boundaries, the manual plotting is rough and not high in precision.

In the ground feature classification in the prior art, only a hyperspectral image is used for ground feature classification, the main means is also the traditional genetic algorithm, BP Neural Network (Back Propagation Neural Network) and other optimization algorithms, all information of hyperspectral data cannot be fully utilized, fine ground feature boundary classification usually depends on semantic segmentation of images in different scales, and the algorithm cannot be expressed more generally on characteristics in different scales, so that the classification precision is general.

A three-dimensional maximum pooling layer is added after each convolution layer in the three-dimensional convolutional neural network, and the loss of original image details is caused by introducing a data dimension reduction mode of spectral dimension pooling, so that the capability of model segmentation is reduced. Although the network training parameters can be obviously reduced by adding the spectral dimension pooling, and the model training time is saved, the loss of spectral information can cause the reduction of the precision of the main boundary, the dimensionality of the three-dimensional convolutional neural network is increased, the data calculation amount is larger, the training parameters are more complex, and the efficiency is lower.

The traditional measuring method mainly comprises the steps of manually selecting points and using high-precision instruments such as a height indicator and a theodolite to carry out measurement, wherein the measurement precision is related to the manual operation precision such as instrument performance, placing position and measurement elevation angle setting, and the accurate measurement can not be guaranteed forever.

In the aspect of a ground feature extraction method based on remote sensing, different remote sensing technologies have great differences in extraction accuracy, coverage, data acquisition difficulty, cost and the like.

Disclosure of Invention

The invention provides a method for applying an artificial intelligence recognition technology to three-dimensional power transmission and transformation engineering measurement and calculation, which has the characteristics of high efficiency, low cost, accurate precision and the like. The method comprises the steps of remote sensing image acquisition, data preprocessing, image segmentation, basic training sample labeling, network model selection, model training, prediction and result analysis and evaluation.

Preferably, the data preprocessing step of the method further comprises: the remote sensing image is influenced by factors such as time and the like, so that the image has a white spot problem; at this time, the images need to be preprocessed, including correction, noise reduction and cloud removal, so that the images can be better classified;

the image segmentation step of the method further comprises: because the remote sensing image has wide coverage range and large information amount and is not suitable for directly utilizing the deep learning algorithm to analyze and process, the image needs to be divided into proper sizes, so that the characteristics of the image are better extracted;

the method also comprises the following steps of marking basic training samples: according to the characteristics of targets or categories such as houses and trees in the remote sensing image, vector labeling is carried out on the remote sensing image to form a file in a vector format, so that the analysis and processing of the image data by a deep learning framework are facilitated;

the network model selecting step of the method further comprises: the network model based on the target identification comprises RCNN, R-FCN and SSD, and the network model based on the semantic segmentation comprises DEEPLAB; the models are convolutional neural networks CNN, and the basic structure of the models is composed of an input layer, a convolutional layer, a pooling layer, a full-link layer and an output layer; the convolution has the function of extracting the features of the original image, the pooling can reduce the number of feature parameters, reduce the calculated amount and improve the robustness of the extracted features; an activation function comprising relu and sigmoid is added between the convolutional layer and the pooling layer so as to realize the nonlinear output of the neuron and improve the approximation capability of the model to the complex function;

the model training step of the method also comprises the following steps: training the model by using a deep learning framework; dividing the sample into a training sample and a verification sample by using the parameters of the convolutional neural network model including the training times, the batch and the learning rate according to the size of the sample data, training the model by using the training sample, and performing precision verification on the obtained model by using the verification sample;

the method also comprises the following prediction steps: predicting unknown image data by using the trained model;

the method also comprises the following steps of: and analyzing and evaluating the predicted result so as to help decision analysis.

Preferably, the method further comprises preprocessing data through deep learning, including shadow removal and image denoising; the detailed shadow removing steps are as follows: the shadow has lower brightness than surrounding ground objects, homomorphic filtering is carried out on the image, the uneven brightness distribution of the image is improved, and deep learning is used for training the shadow of the processed image; the detailed image denoising processing step is that in the shooting and transmission processes of the remote sensing image, more random noise can be generated due to factors such as equipment or digitization and the like, the characteristic extraction in the later training process can be influenced to a greater extent, the denoising algorithm is combined by median filtering and Gaussian filtering, and the median filtering is effective for the random noise in the image; the Gaussian filter has good noise reduction effect on Gaussian noise presenting certain normal distribution; in real environments, noise is a complex of noise from different sources; the true noise can be viewed as the sum of random variables of a very large number of different probability distributions, each of which is independent, and then approaches a gaussian distribution as the number of noise sources increases.

Preferably, the method further comprises data enhancement; the data enhancement comprises a space domain and a frequency domain, and the detailed steps are as follows: spatial domain: data enhancement is carried out on an image space domain, and the main means comprises image denoising, image overturning, brightness and contrast enhancement, color intensity improvement and sample quantity increase; frequency domain: fourier transform is carried out on the image to obtain a frequency domain image, then filtering is carried out on the frequency domain image, Gaussian high-pass filtering and low-pass filtering are used, and different filters are used for different ground objects.

Preferably, the method further comprises data training; the detailed steps are as follows:

(1) establishing an initial network model: the method has the advantages that the XCep is selected as the initial model, the network model has the characteristics of high speed and high precision, the characteristics can be continuously analyzed and filtered, and effective characteristics can be extracted; for remote sensing images with large data volume, an Entry Flow module in the model can continuously perform down sampling, the space dimension is reduced, the middle layer can continuously learn the association relation, the characteristics are optimized, and the accuracy of characteristic extraction is improved;

(2) adjusting training parameters: according to the characteristics of different ground features, including the average size and color distribution condition of a single ground feature, adjusting parameters of training data, including the size of a cutting training picture, the void convolution rate and the normalization parameters of data, and finishing the independent training of various ground features;

(3) deep separable training neural network coding and decoding: and (3) coding structure: adopting the void convolution with different void rates and an improved pyramid model to output semantic information containing 256 channels, and adjusting different parameters to obtain output characteristic graphs with different scales;

the decoding structure: performing up-sampling by using 4 times of bilinear interpolation, correspondingly reducing the bottom characteristic diagram of the channel number through 1 × 1 convolution, and then recovering the size of the original image by applying 3 × 3 convolution and 4 times of up-sampling to obtain a training result;

(4) and (3) correcting deviation by back propagation, and determining a loss function: the probability of positive class and negative class of a real sample is evaluated by selecting a cross entropy function according to a loss function, and a sample normalization method is added during backward propagation according to a cross entropy function value, so that the problem of gradient disappearance is solved;

preferably, the method further comprises image data prediction; the detailed steps are as follows:

(1) block prediction strategy: for the remote sensing image with larger data volume, according to the size of a single target, cutting the original prediction image into blocks with the same size, and finally splicing the prediction result, so that the prediction efficiency can be improved, but the defect is that the splicing phenomenon among the blocks is obvious;

(2) multi-scale fusion: and continuously predicting the original image in blocks for multiple times, ensuring that the size and the position of each block are different, and fusing the results of the multiple predictions according to the probability of positive and negative samples of the prediction result of the pixel to finally form the prediction result of the whole image.

Preferably, the method further comprises point cloud extraction; the method comprises the steps of extracting a house from point cloud and extracting forest from the point cloud; point cloud extraction house: firstly, removing ground points, and extracting the house point cloud data according to the elevation difference of the target surface of the building and the reflection attribute of the laser pulse; point cloud extraction of forest trees: vegetation generally has an irregular shape, most of which vertically protrudes on the bare ground and occupies a certain continuous area; the surface coarse sugar degree formed by the vegetation point cloud is higher than that of a building or a bare ground, and the vegetation and other main ground objects are distinguished by utilizing the curved surface change and curvature of the point cloud; however, in case of a particularly rugged bare ground, such as a ground with a stone field, the method for determining the surface roughness may fail, and therefore, the characteristic of local point cloud density should be considered in addition to the rough roughness of the point cloud surface; because the laser beam can penetrate part of the crown to directly strike branches or the ground, whether the vegetation is the vegetation can be judged by utilizing the elevation difference between the first echo and the last echo of the point cloud data; however, the first echo point cloud includes vegetation, a building roof and a part of wall, and the last echo includes ground points and a part of vegetation, so that the elevation difference value of multiple echoes can be used for determining vegetation candidate points, and after building and wall objects are detected by other rules, the forest point cloud can be accurately identified.

Preferably, the method further comprises a house vectorization modeling step;

the method comprises the following steps: separating different plane point clouds by adopting a method combining RANSAC method segmentation and distance method segmentation, solving corresponding plane equation parameters by utilizing foot points in each segmentation plane, and separating each roof plane point cloud;

step two: extracting a group contour node of each partition plane point cloud by using an Alpha Shape algorithm;

step three: extracting initial key points based on the direction angles, and further locking the detection range of the key points;

step four: determining a topological relationship between the roof planes by analyzing distances between the plane boundaries and the remaining plane boundaries;

step five: obtaining three intersection line characteristics including a common line segment characteristic, a common ray characteristic and a common straight line characteristic according to the topological relation between planes, and determining the accurate position of a final key point aiming at different intersection line characteristics;

step six: constructing a roof model by using the detected key points, acquiring the side face of the building by combining corresponding ground information, and finally integrating to obtain a complete three-dimensional model of the building;

preferably, the method further comprises the steps of house removal statistics and forest felling statistics; and (3) house removal statistics: combining a three-dimensional platform, warehousing the vectorized house model, automatically counting the number, area and volume of houses in the line corridor range, and generating a dismantling distribution map; forest felling statistics: and (4) combining a three-dimensional platform, warehousing the divided single-wood models, and automatically counting the number of the trees in the range of the line corridor.

Preferably, the method further comprises a veneer splitting step;

the method comprises the following steps: establishing a characteristic field;

step two: establishing a label field model, and determining the single-wood standing position through CHM;

step three: segmentation was performed using a Potts model.

According to the invention, by automatically classifying the true-color DOM image and the point cloud data, elements such as houses and trees which have important influences on the power corridor of the power transmission line can be efficiently and accurately extracted, the workload of measuring personnel is greatly reduced, and the work of measuring and calculating the engineering quantity in the traditional mode can be completed only by fixed-point manual intervention. The high-precision boundary extraction result of the deep learning method is high, the target prediction result is compared with the label in a test set, the accuracy rate of the ground feature identification can reach more than 85%, the boundary segmentation can be basically accurately carried out on the classified ground features, and reliable basis is provided for the actual measurement and calculation work.

Drawings

In order to more clearly illustrate the detailed description of the invention or the technical solutions in the prior art, the drawings that are needed in the detailed description of the invention or the prior art will be briefly described below. Throughout the drawings, like elements or portions are generally identified by like reference numerals. In the drawings, elements or portions are not necessarily drawn to scale.

FIG. 1 is a flowchart illustrating image depth learning according to the present invention;

FIG. 2 is a flow chart illustrating steps of implementing the present invention;

FIG. 3 is a schematic diagram of the process of deep learning of aerial survey data according to the present invention;

FIG. 4 is a schematic diagram of a point cloud extraction process of the present invention;

FIG. 5 is a flow chart of the point cloud extraction steps of the present invention.

Detailed Description

Embodiments of the present invention will be described in detail below with reference to the accompanying drawings. The following examples are only for illustrating the technical solutions of the present invention more clearly, and therefore are only examples, and the protection scope of the present invention is not limited thereby.

It is to be noted that, unless otherwise specified, technical or scientific terms used herein shall have the ordinary meaning as understood by those skilled in the art to which the invention pertains.

With the appearance of new sensors and more advanced sensor-mounted platforms and the continuous improvement of resolution of remote sensing images, remote sensing images such as multispectral images, high resolution images, high spectral resolution images and the like also appear in succession. However, the remote sensing images have large data volume and multiple information dimensions, so that the remote sensing information processing becomes relatively difficult, and the difficulty of remote sensing image classification also increases relatively. Processing by traditional remote sensing image classification methods has become increasingly impractical. Therefore, it is necessary to research and find a more efficient and accurate image classification method. Deep learning (AI) is derived from artificial neural networks and is a machine learning that determines a distributed feature representation of data by combining low-level features to form more abstract high-level features or classes. In the field of remote sensing, in particular, applications of deep learning in image classification are increasing, and deep learning is favored by remote sensing researchers due to its high efficiency and intelligence.

On the basis of researching the existing remote sensing image classification method, the remote sensing image is interpreted and analyzed by using an open source framework of deep learning, so that the problems of target identification and ground feature classification are solved. The deep neural network model can be conveniently and quickly constructed by utilizing the deep learning framework, and various neural network models can be selected according to different application scenes; meanwhile, the deep learning framework can also use the GPU, computer clusters and even computing resources of a cloud environment, so that a large number of problems of operation and analysis can be effectively solved.

In the surveying industry, the traditional method for measuring and calculating the engineering quantity is to manually draw ground features such as houses, trees, expropriations and the like based on images, generate an ORG file, and import the ORG file into design software for measuring and calculating. There are generally four methods for estimation: the first method is based on full-waveform laser Detection And measurement (LiDAR, Light Detection And Ranging) data, structured information of vegetation is obtained, point cloud attributes are collected in an affected range, comprehensive waveform characteristic parameters are calculated, And measurement accuracy of a target size is obtained And improved through a single data source; the second method is that based on full-waveform LiDAR data, positioning and extraction of monomer trees, houses and land seeking are derived, and a Weber distribution and cubic spline function model are combined to be used as characteristic variables to obtain model parameters; the third method is based on multisource contemporaneous high-resolution remote sensing data, firstly, target classification is carried out based on the high-resolution data and an object-oriented segmentation method, and then, image ground objects are identified based on space details and spectral features extracted from the high-spectral data and by combining a back propagation neural network; and fourthly, recognizing the ground object based on the hyperspectral data by using a three-dimensional convolutional neural network, performing semantic segmentation on the complex ground object by using the three-dimensional convolutional neural network based on the airborne hyperspectral data, constructing a data set for training and verifying, completing the training of a deep learning model, and predicting the category of each pixel in the hyperspectral image by using the trained model to obtain the tree species distribution chart of the whole area.

This method is quite labor intensive, requires a lot of manpower and is long in cycle. For complex terrain boundaries, the manual plotting is rough and not high in precision.

In the ground feature classification in the prior art, only a hyperspectral image is used for ground feature classification, the main means is also the traditional genetic algorithm, BP Neural Network (Back Propagation Neural Network) and other optimization algorithms, all information of hyperspectral data cannot be fully utilized, fine ground feature boundary classification usually depends on semantic segmentation of images in different scales, and the algorithm cannot be expressed more generally on characteristics in different scales, so that the classification precision is general.

A three-dimensional maximum pooling layer is added after each convolution layer in the three-dimensional convolutional neural network, and the loss of original image details is caused by introducing a data dimension reduction mode of spectral dimension pooling, so that the capability of model segmentation is reduced. Although the network training parameters can be obviously reduced by adding the spectral dimension pooling, and the model training time is saved, the loss of spectral information can cause the reduction of the precision of the main boundary, the dimensionality of the three-dimensional convolutional neural network is increased, the data calculation amount is larger, the training parameters are more complex, and the efficiency is lower.

The traditional measuring method mainly comprises the steps of manually selecting points and using high-precision instruments such as a height indicator and a theodolite to carry out measurement, wherein the measurement precision is related to the manual operation precision such as instrument performance, placing position and measurement elevation angle setting, and the accurate measurement can not be guaranteed forever.

In the aspect of a ground feature extraction method based on remote sensing, different remote sensing technologies have great differences in extraction accuracy, coverage, data acquisition difficulty, cost and the like.

As shown in fig. 1 to 5, aiming at the defects in the prior art, the invention provides a method for applying an artificial intelligence recognition technology to three-dimensional power transmission and transformation engineering measurement, which comprises the steps of remote sensing image acquisition, data preprocessing, image segmentation, marking of a basic training sample, network model selection, model training, prediction, and result analysis and evaluation.

Preferably, the data preprocessing step of the method further comprises: the remote sensing image is influenced by factors such as time and the like, so that the image has a white spot problem; at this time, the images need to be preprocessed, including correction, noise reduction and cloud removal, so that the images can be better classified;

the image segmentation step of the method further comprises: because the remote sensing image has wide coverage range and large information amount and is not suitable for directly utilizing the deep learning algorithm to analyze and process, the image needs to be divided into proper sizes, so that the characteristics of the image are better extracted;

the method also comprises the following steps of marking basic training samples: according to the characteristics of targets or categories such as houses and trees in the remote sensing image, vector labeling is carried out on the remote sensing image to form a file in a vector format, so that the analysis and processing of the image data by a deep learning framework are facilitated;

the network model selecting step of the method further comprises: the network model based on the target identification comprises RCNN, R-FCN and SSD, and the network model based on the semantic segmentation comprises DEEPLAB; the models are convolutional neural networks CNN, and the basic structure of the models is composed of an input layer, a convolutional layer, a pooling layer, a full-link layer and an output layer; the convolution has the function of extracting the features of the original image, the pooling can reduce the number of feature parameters, reduce the calculated amount and improve the robustness of the extracted features; an activation function comprising relu and sigmoid is added between the convolutional layer and the pooling layer so as to realize the nonlinear output of the neuron and improve the approximation capability of the model to the complex function;

the model training step of the method also comprises the following steps: training the model by using a deep learning framework; dividing the sample into a training sample and a verification sample by using the parameters of the convolutional neural network model including the training times, the batch and the learning rate according to the size of the sample data, training the model by using the training sample, and performing precision verification on the obtained model by using the verification sample;

the method also comprises the following prediction steps: predicting unknown image data by using the trained model;

the method also comprises the following steps of: and analyzing and evaluating the predicted result so as to help decision analysis.

Preferably, the method further comprises preprocessing data through deep learning, including shadow removal and image denoising; the detailed shadow removing steps are as follows: the shadow has lower brightness than surrounding ground objects, homomorphic filtering is carried out on the image, the uneven brightness distribution of the image is improved, and deep learning is used for training the shadow of the processed image; the detailed image denoising processing step is that in the shooting and transmission processes of the remote sensing image, more random noise can be generated due to factors such as equipment or digitization and the like, the characteristic extraction in the later training process can be influenced to a greater extent, the denoising algorithm is combined by median filtering and Gaussian filtering, and the median filtering is effective for the random noise in the image; the Gaussian filter has good noise reduction effect on Gaussian noise presenting certain normal distribution; in real environments, noise is a complex of noise from different sources; the true noise can be viewed as the sum of random variables of a very large number of different probability distributions, each of which is independent, and then approaches a gaussian distribution as the number of noise sources increases.

Preferably, the method further comprises data enhancement; the data enhancement comprises a space domain and a frequency domain, and the detailed steps are as follows: spatial domain: data enhancement is carried out on an image space domain, and the main means comprises image denoising, image overturning, brightness and contrast enhancement, color intensity improvement and sample quantity increase; frequency domain: fourier transform is carried out on the image to obtain a frequency domain image, then filtering is carried out on the frequency domain image, Gaussian high-pass filtering and low-pass filtering are used, and different filters are used for different ground objects.

Preferably, the method further comprises data training; the detailed steps are as follows:

(1) establishing an initial network model: the method has the advantages that the XCep is selected as the initial model, the network model has the characteristics of high speed and high precision, the characteristics can be continuously analyzed and filtered, and effective characteristics can be extracted; for remote sensing images with large data volume, an Entry Flow module in the model can continuously perform down sampling, the space dimension is reduced, the middle layer can continuously learn the association relation, the characteristics are optimized, and the accuracy of characteristic extraction is improved;

(2) adjusting training parameters: according to the characteristics of different ground features, including the average size and color distribution condition of a single ground feature, adjusting parameters of training data, including the size of a cutting training picture, the void convolution rate and the normalization parameters of data, and finishing the independent training of various ground features;

(3) deep separable training neural network coding and decoding: and (3) coding structure: adopting the void convolution with different void rates and an improved pyramid model to output semantic information containing 256 channels, and adjusting different parameters to obtain output characteristic graphs with different scales;

the decoding structure: performing up-sampling by using 4 times of bilinear interpolation, correspondingly reducing the bottom characteristic diagram of the channel number through 1 × 1 convolution, and then recovering the size of the original image by applying 3 × 3 convolution and 4 times of up-sampling to obtain a training result;

(4) and (3) correcting deviation by back propagation, and determining a loss function: the probability of positive class and negative class of a real sample is evaluated by selecting a cross entropy function according to a loss function, and a sample normalization method is added during backward propagation according to a cross entropy function value, so that the problem of gradient disappearance is solved;

preferably, the method further comprises image data prediction; the detailed steps are as follows:

(1) block prediction strategy: for the remote sensing image with larger data volume, according to the size of a single target, cutting the original prediction image into blocks with the same size, and finally splicing the prediction result, so that the prediction efficiency can be improved, but the defect is that the splicing phenomenon among the blocks is obvious;

(2) multi-scale fusion: and continuously predicting the original image in blocks for multiple times, ensuring that the size and the position of each block are different, and fusing the results of the multiple predictions according to the probability of positive and negative samples of the prediction result of the pixel to finally form the prediction result of the whole image.

Preferably, the method further comprises point cloud extraction; the method comprises the steps of extracting a house from point cloud and extracting forest from the point cloud; point cloud extraction house: firstly, removing ground points, and extracting the house point cloud data according to the elevation difference of the target surface of the building and the reflection attribute of the laser pulse; point cloud extraction of forest trees: vegetation generally has an irregular shape, most of which vertically protrudes on the bare ground and occupies a certain continuous area; the surface coarse sugar degree formed by the vegetation point cloud is higher than that of a building or a bare ground, and the vegetation and other main ground objects are distinguished by utilizing the curved surface change and curvature of the point cloud; however, in case of a particularly rugged bare ground, such as a ground with a stone field, the method for determining the surface roughness may fail, and therefore, the characteristic of local point cloud density should be considered in addition to the rough roughness of the point cloud surface; because the laser beam can penetrate part of the crown to directly strike branches or the ground, whether the vegetation is the vegetation can be judged by utilizing the elevation difference between the first echo and the last echo of the point cloud data; however, the first echo point cloud includes vegetation, a building roof and a part of wall, and the last echo includes ground points and a part of vegetation, so that the elevation difference value of multiple echoes can be used for determining vegetation candidate points, and after building and wall objects are detected by other rules, the forest point cloud can be accurately identified.

Preferably, the method further comprises a house vectorization modeling step;

the method comprises the following steps: separating different plane point clouds by adopting a method combining RANSAC method segmentation and distance method segmentation, solving corresponding plane equation parameters by utilizing foot points in each segmentation plane, and separating each roof plane point cloud;

step two: extracting a group contour node of each partition plane point cloud by using an Alpha Shape algorithm;

step three: extracting initial key points based on the direction angles, and further locking the detection range of the key points;

step four: determining a topological relationship between the roof planes by analyzing distances between the plane boundaries and the remaining plane boundaries;

step five: obtaining three intersection line characteristics including a common line segment characteristic, a common ray characteristic and a common straight line characteristic according to the topological relation between planes, and determining the accurate position of a final key point aiming at different intersection line characteristics;

step six: constructing a roof model by using the detected key points, acquiring the side face of the building by combining corresponding ground information, and finally integrating to obtain a complete three-dimensional model of the building;

preferably, the method further comprises the steps of house removal statistics and forest felling statistics; and (3) house removal statistics: combining a three-dimensional platform, warehousing the vectorized house model, automatically counting the number, area and volume of houses in the line corridor range, and generating a dismantling distribution map; forest felling statistics: and (4) combining a three-dimensional platform, warehousing the divided single-wood models, and automatically counting the number of the trees in the range of the line corridor.

Preferably, the method further comprises a veneer splitting step;

the method comprises the following steps: establishing a characteristic field;

step two: establishing a label field model, and determining the single-wood standing position through CHM;

step three: segmentation was performed using a Potts model.

The first embodiment is as follows:

the specific implementation process of the invention can comprise the following steps:

first, deep learning

Data pre-processing

A. Shadow removal

The shadow has lower brightness than surrounding ground objects, homomorphic filtering is carried out on the image, uneven brightness distribution of the image is improved, and deep learning is used for training the shadow of the processed image.

B. Image denoising process

In the shooting and transmission processes of the remote sensing image, more random noise can be generated due to equipment or digitalization and other factors, the feature extraction in the later training process can be influenced to a greater extent, the denoising algorithm combines median filtering and Gaussian filtering, and the median filtering is effective for the random noise in the image. The Gaussian filter has good noise reduction effect on Gaussian noise which presents a certain normal distribution. In real environments, noise is a complex of noise from different sources. The true noise can be viewed as the sum of random variables of a very large number of different probability distributions, each of which is independent, and then approaches a gaussian distribution as the number of noise sources increases.

Data enhancement

A. Spatial domain: data enhancement is carried out on an image space domain by the main means of image denoising, image overturning, brightness and contrast enhancement, color intensity improvement, sample quantity increase and the like

B. Frequency domain: fourier transform is carried out on the image to obtain a frequency domain image, then filtering is carried out on the frequency domain image, Gaussian high-pass filtering and low-pass filtering are used, and different filters are used for different ground objects.

Data training

A. Initial network model building

The initial model is selected XCEPTION, and the network model has the characteristics of high speed and high precision, can continuously analyze and filter the characteristics and can further extract effective characteristics. For the remote sensing image with large data volume, the Entry Flow module in the model can continuously sample, the space dimensionality is reduced, the middle layer can continuously learn the association relation, the characteristics are optimized, and the accuracy of characteristic extraction is improved.

B. Training parameter adjustment

According to the characteristics of different ground features, such as the average size of a single ground feature and the color distribution condition, parameters of training data, such as the size of a cutting training picture, the void convolution rate, the normalization parameter of data and the like, are adjusted, and the independent training of various ground features is completed.

C. Deep separable training neural network coding and decoding

And (3) coding structure: and (3) outputting semantic information containing 256 channels by adopting the void convolution with different void rates and the improved pyramid model, and adjusting different parameters to obtain output characteristic diagrams with different scales.

The decoding structure: and performing up-sampling by using 4 times of bilinear interpolation, correspondingly reducing the bottom characteristic diagram of the number of channels through 1 × 1 convolution, and then recovering the size of the original image by applying 3 × 3 convolution and 4 times of up-sampling to obtain a training result.

D. Back propagation correction, determining loss function

The probability of positive class and negative class of the real sample is evaluated by selecting a cross entropy function according to the loss function, and a sample normalization method is added during backward propagation according to the cross entropy function value, so that the problem of gradient disappearance is solved.

Image data prediction

A. Block prediction strategy:

for the remote sensing image with larger data volume, the original prediction image is cut into blocks with the same size according to the size of a single target, and finally the prediction results are spliced, so that the prediction efficiency can be improved, but the defect is that the splicing phenomenon among the blocks is obvious.

B. Multi-scale fusion:

and continuously predicting the original image in blocks for multiple times, ensuring that the size and the position of each block are different, and fusing the results of the multiple predictions according to the probability of positive and negative samples of the prediction result of the pixel to finally form the prediction result of the whole image.

Second, point cloud extraction

Point cloud extraction house: ground points are removed, and the extraction of the house point cloud data is realized according to the elevation difference of the target surface of the building and the reflection attribute of the laser pulse.

Point cloud extraction of forest trees: vegetation is generally of irregular shape, mostly protruding vertically on bare ground and occupying a certain continuous area. The vegetation (especially the tall forest) point cloud forms a surface coarse sugar degree higher than that of a building or a bare ground, and the curved surface change and the curvature of the point cloud are utilized to distinguish the vegetation from other main ground objects. However, in case of a particularly rugged bare ground, such as a ground with a stone field, the method for determining the surface roughness may fail, and therefore, the characteristic of local point cloud density should be considered in addition to the rough roughness of the point cloud surface. Since the laser beam can penetrate part of the crown to directly strike branches or the ground, the height difference between the first echo and the last echo of the point cloud data can be used for judging whether the vegetation is. However, the first echo point cloud includes vegetation, a building roof and a part of wall, and the last echo includes ground points and a part of vegetation, so that the elevation difference value of multiple echoes can be used for determining vegetation candidate points, and after objects such as buildings, walls and the like are detected by using other rules, the forest point cloud can be accurately identified.

Vectorization modeling of a house:

the method comprises the following steps: separating different plane point clouds by adopting a method combining RANSAC method segmentation and distance method segmentation, solving corresponding plane equation parameters by utilizing foot points in each segmentation plane, and separating each roof plane point cloud;

step two: extracting a group contour node of each partition plane point cloud by using an Alpha Shape algorithm;

step three: extracting initial key points based on the direction angles, and further locking the detection range of the key points;

step four: determining a topological relationship between the roof planes by analyzing distances between the plane boundaries and the remaining plane boundaries;

step five: obtaining three intersection line characteristics including a common line segment characteristic, a common ray characteristic and a common straight line characteristic according to the topological relation between planes, and determining the accurate position of a final key point aiming at different intersection line characteristics;

step six: constructing a roof model by using the detected key points, acquiring the side face of the building by combining corresponding ground information, and finally integrating to obtain a complete three-dimensional model of the building;

and (3) house removal statistics: and (4) combining a three-dimensional platform, warehousing the vectorized house model, automatically counting the number, area and volume of the houses in the line corridor range, and generating a dismantling distribution map.

Single wood segmentation:

the method comprises the following steps: establishing a characteristic field

Step two: establishing a label field model, and establishing the single-wood standing position through CHM

Step three: segmentation using Potts model

Forest felling statistics: and (4) combining a three-dimensional platform, warehousing the divided single-wood models, and automatically counting the number of the trees in the range of the line corridor.

Example two:

extracting forest trees and houses, and carrying out area statistics:

the extraction of ground objects such as trees only needs to sample areas with different degrees, and all the conditions of the trees in the whole area can be obtained only by sampling a small amount of personnel on site, so that a large amount of manpower and material resources are saved. When the forest sample is marked, the detail of marking is needed, and houses or naked landmarks in the forest are not marked. When the model is trained, the ratio of the training sample to the test sample is 4:1, the training frequency is 100000, the training batch is 4, and the learning rate is 0.001. The trained network model is MOBILENET _ V2. And converting the result into vector data by using a grid-to-vector method for the predicted forest data, and calculating the area of the forest data and the like by calling a function of calculating the vector area by GRASS.

When the house sample is marked, the mark is more detailed as much as possible and is marked along the outline of the house as much as possible. When the model is trained, the ratio of the training sample to the test sample is 4:1, the training frequency is 100000, the training batch is 4, the learning rate is 0.001, and the trained network model is MOBILE ET _ V2. And converting the result into vector data by using a grid-to-vector method for the predicted house data, and calculating the area of the house data by calling a function of calculating the vector area of the GRASS.

Extraction of roads:

by using the artificial intelligence technology, the urban road network of villages and towns can be generated in real time, the traffic flow can be analyzed, the traffic light interval can be adjusted, the waiting time of vehicles can be shortened, and the traffic efficiency of the urban roads of villages and towns can be improved. When the road sample is marked, the marking is detailed as much as possible, the marking is performed along the contour of the road as much as possible, and the marking is performed according to two data when the road sample is branched. When the model is trained, the ratio of the training sample to the test sample is 4:1, the training frequency is 100000, the training batch is 4, the learning rate is 0.001, and the trained network model is MOBILE ET _ V2. And converting the predicted road data into vector data by using a method of converting a grid into a vector, calculating the area of the road data by calling a function of calculating the vector area by GRASS, and extracting the center line of the road and the like.

Example three:

1 overview

The Raster Vision is an open source framework for Python developers to build computer Vision models on satellites, aviation, and other large image sets, including oblique drone images. It can help engineers to quickly and repeatably configure experiments that go through the core components of a machine learning workflow: analyzing training data, creating training chips, training models, creating predictions, evaluating models, and binding model files and configurations for deployment.

The Raster Vision is a framework that can be state configured in a modified and re-used manner and tracks all files at each step of the workflow built by the machine learning model.

2 frame and module

2.1 framework introduction

When there is a set of images and corresponding sample data, the region of interest (AOI) can be selected to describe the marked position of the image, the register Vision Workflow can start working, the RV starts analyzing the data, trains the model, and finally generates a packed model and configuration file, which can easily utilize the model under various deployment conditions. In the master Vision workflow, multiple experiments need to be run to find the best model to deploy.

Images: the original tif image position is set by a Raster Source, and a large scene image can comprise a plurality of sub-images or a file.

Labels: denoted LabelSource in the Rapid Vision, is the content that provides annotations or tags for a scene. The nature of the label produced by the LabelSource is specific to the task that the machine learning model is performing.

AOI: a marked region of interest of the scene image is described.

1. ANALYZE: gathering data set level statistics and metrics

2. CHIP: chips of various image and label sources are created.

3. Train: various "back ends" (e.g., TensorFlow or Keras) were used to train the model.

4. PREDICT: the trained models are used to predict validation and test data.

5. EVAL: evaluation metrics, such as F1 score, accuracy, and recall, are derived from the model's predictions of the validation dataset.

6. BUNDLE: the trained models are bundled into a prediction package that can be deployed in batch processes, active servers, and other workflows.

The Raster Vision framework contains three task sets, respectively, Chip Classification, Object Detection, and Semantic Segmentation. These three task sets all follow a RasterVision unified workflow. The task is configured With the TaskConfig and then set to run using the Width _ Task (task) method.

2.2 Chip Classification

For chip classification, the backend uses the Keras classification, uses the Resnet 50, Keras is an advanced neural network API, and supports the backend as follows: tensorflow by Google, CNTK by Microsoft, and Theano.

2.3 Object Detection

The back-end framework used by Object Detection is the TensorFlow Object Detection API, in which the goal is to predict a bounding box and a class around each Object of interest. This task requires higher target detection accuracy than chip classification, but has the ability to locate and personalize objects. The object detection model requires more time to train and process very close objects.

Four basic steps of target detection (candidate region generation, feature extraction, classification, position refinement)

2.3.1 network model: SSD

SSD is based on a forward-propagating CNN network, yielding a series of fixed-size (fixed-size) bounding boxes, and the possibility of containing an object instance in each box, score. Thereafter, a Non-maximum suppression (Non-maximum suppression) is performed to obtain the final predictions. The SSD network structure is divided into two parts: basic network + pyramid network. On the scale of the original image, a candidate Anchor of the dense hemp is set. Cnn is then used to determine which anchors are targeted inside and which are background without targets. The SSD has a phenomenon of poor detection effect on small targets.

The underlying network of the SSD is the top 4 network of the VGG-16.

A pyramid network is a simple convolutional network with gradually smaller feature maps.

And the SSD evaluates the confidence degrees belonging to the class p in the output default box, and reserves the highest score. As a final test result

Loss function: the total target loss function is weighted by the summation of localization loss (loc) and confidence loss (conf).

2.4 Semantic Segmentation

The backend framework used by Semantic Segmentation is the API of Tensorflow Dellab,

and (3) network model: MobileNet _ V2

Loss function (verification policy): cross entropy

p represents the distribution of the true markers (probability distribution of source samples), q is the distribution of the predicted markers of the trained model (probability distribution of output results), and the cross entropy loss function can measure the similarity between p and q. The cross entropy as the loss function has the advantage that the problem of the learning rate reduction of the mean square error loss function can be avoided when the gradient is reduced by using the sigmoid function, because the learning rate can be controlled by the output error.

3 Tensorflow deep learning

TensorFlow is an open source software library that uses dataflow graphs for numerical calculations, and TensorFlow is not just a powerful neural network library. It may let you build other machine learning algorithms on it, such as decision trees or k-nearest neighbors. TensorFlow defines the model and training of the run by Graph and Session,

the TensorFlow code does not realize least square algorithm and the like, does not control code logic by if-else, and is completely learned by data driving and dynamic adjustment of the Loss value according to a gradient descent algorithm. TensorFlow has encapsulated fully-connected networks, convolutional neural networks, RNN and LSTM, and we can combine various network models.

In the description of the present application, it is to be understood that the terms "upper", "lower", "front", "rear", "left", "right", "horizontal", "inner", "outer", and the like, indicate orientations or positional relationships based on the orientations or positional relationships shown in the drawings, are only for convenience in describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and thus, should not be construed as limiting the present invention.

Furthermore, the terms "first", "second", etc. are used for descriptive purposes only and are not to be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated. In the description of the present invention, "a plurality" means two or more unless specifically defined otherwise.

In the description of the present invention, numerous specific details are set forth. It is understood, however, that embodiments of the invention may be practiced without these specific details. In some instances, well-known methods, structures and techniques have not been shown in detail in order not to obscure an understanding of this description.

In the description herein, particular features, structures, materials, or characteristics may be combined in any suitable manner in any one or more embodiments or examples. Furthermore, various embodiments or examples and features of different embodiments or examples described in this specification can be combined and combined by one skilled in the art without contradiction.

Finally, it should be noted that: the above embodiments are only used to illustrate the technical solution of the present invention, and not to limit the same; while the invention has been described in detail and with reference to the foregoing embodiments, it will be understood by those skilled in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some or all of the technical features may be equivalently replaced; such modifications and substitutions do not depart from the spirit and scope of the present invention, and they should be construed as being included in the following claims and description.

Claims (10)

1. A method for applying artificial intelligence recognition technology to three-dimensional power transmission and transformation engineering measurement and calculation is characterized by comprising the following steps: the method comprises the steps of remote sensing image acquisition, data preprocessing, image segmentation, basic training sample labeling, network model selection, model training, prediction and result analysis and evaluation.

2. The method for applying the artificial intelligence recognition technology to the three-dimensional power transmission and transformation engineering measurement and calculation according to claim 1, wherein the method comprises the following steps:

the data preprocessing step of the method further comprises the following steps: the remote sensing image is influenced by factors such as time and the like, so that the image has a white spot problem; at this time, the images need to be preprocessed, including correction, noise reduction and cloud removal, so that the images can be better classified;

the image segmentation step of the method further comprises: because the remote sensing image has wide coverage range and large information amount and is not suitable for directly utilizing the deep learning algorithm to analyze and process, the image needs to be divided into proper sizes, so that the characteristics of the image are better extracted;

the method also comprises the following steps of marking basic training samples: according to the characteristics of targets or categories such as houses and trees in the remote sensing image, vector labeling is carried out on the remote sensing image to form a file in a vector format, so that the analysis and processing of the image data by a deep learning framework are facilitated;

the network model selecting step of the method further comprises: the network model based on the target identification comprises RCNN, R-FCN and SSD, and the network model based on the semantic segmentation comprises DEEPLAB; the models are convolutional neural networks CNN, and the basic structure of the models is composed of an input layer, a convolutional layer, a pooling layer, a full-link layer and an output layer; the convolution has the function of extracting the features of the original image, the pooling can reduce the number of feature parameters, reduce the calculated amount and improve the robustness of the extracted features; an activation function comprising relu and sigmoid is added between the convolutional layer and the pooling layer so as to realize the nonlinear output of the neuron and improve the approximation capability of the model to the complex function;

the model training step of the method also comprises the following steps: training the model by using a deep learning framework; dividing the sample into a training sample and a verification sample by using the parameters of the convolutional neural network model including the training times, the batch and the learning rate according to the size of the sample data, training the model by using the training sample, and performing precision verification on the obtained model by using the verification sample;

the method also comprises the following prediction steps: predicting unknown image data by using the trained model;

the method also comprises the following steps of: and analyzing and evaluating the predicted result so as to help decision analysis.

3. The method for applying the artificial intelligence recognition technology to the three-dimensional power transmission and transformation engineering measurement and calculation according to claim 1, wherein the method comprises the following steps: the method further comprises the steps of preprocessing data through deep learning, wherein the data comprise shadow removal and image denoising; the detailed shadow removing steps are as follows: the shadow has lower brightness than surrounding ground objects, homomorphic filtering is carried out on the image, the uneven brightness distribution of the image is improved, and deep learning is used for training the shadow of the processed image; the detailed image denoising processing step is that in the shooting and transmission processes of the remote sensing image, more random noise can be generated due to factors such as equipment or digitization and the like, the characteristic extraction in the later training process can be influenced to a greater extent, the denoising algorithm is combined by median filtering and Gaussian filtering, and the median filtering is effective for the random noise in the image; the Gaussian filter has good noise reduction effect on Gaussian noise presenting certain normal distribution; in real environments, noise is a complex of noise from different sources; the true noise can be viewed as the sum of random variables of a very large number of different probability distributions, each of which is independent, and then approaches a gaussian distribution as the number of noise sources increases.

4. The method for applying the artificial intelligence recognition technology to the three-dimensional power transmission and transformation engineering measurement and calculation according to claim 1, wherein the method comprises the following steps: the method further includes data enhancement; the data enhancement comprises a space domain and a frequency domain, and the detailed steps are as follows: spatial domain: data enhancement is carried out on an image space domain, and the main means comprises image denoising, image overturning, brightness and contrast enhancement, color intensity improvement and sample quantity increase; frequency domain: fourier transform is carried out on the image to obtain a frequency domain image, then filtering is carried out on the frequency domain image, Gaussian high-pass filtering and low-pass filtering are used, and different filters are used for different ground objects.

5. The method for applying the artificial intelligence recognition technology to the three-dimensional power transmission and transformation engineering measurement and calculation according to claim 1, wherein the method comprises the following steps: the method further comprises data training; the detailed steps are as follows:

(1) establishing an initial network model: the method has the advantages that the XCep is selected as the initial model, the network model has the characteristics of high speed and high precision, the characteristics can be continuously analyzed and filtered, and effective characteristics can be extracted; for remote sensing images with large data volume, an Entry Flow module in the model can continuously perform down sampling, the space dimension is reduced, the middle layer can continuously learn the association relation, the characteristics are optimized, and the accuracy of characteristic extraction is improved;

(2) adjusting training parameters: according to the characteristics of different ground features, including the average size and color distribution condition of a single ground feature, adjusting parameters of training data, including the size of a cutting training picture, the void convolution rate and the normalization parameters of data, and finishing the independent training of various ground features;

(3) deep separable training neural network coding and decoding: and (3) coding structure: adopting the void convolution with different void rates and an improved pyramid model to output semantic information containing 256 channels, and adjusting different parameters to obtain output characteristic graphs with different scales;

the decoding structure: performing up-sampling by using 4 times of bilinear interpolation, correspondingly reducing the bottom characteristic diagram of the channel number through 1 × 1 convolution, and then recovering the size of the original image by applying 3 × 3 convolution and 4 times of up-sampling to obtain a training result;

(4) and (3) correcting deviation by back propagation, and determining a loss function: the probability of positive class and negative class of the real sample is evaluated by selecting a cross entropy function according to the loss function, and a sample normalization method is added during backward propagation according to the cross entropy function value, so that the problem of gradient disappearance is solved.

6. The method for applying the artificial intelligence recognition technology to the three-dimensional power transmission and transformation engineering measurement and calculation according to claim 1, wherein the method comprises the following steps: the method also includes image data prediction; the detailed steps are as follows:

(1) block prediction strategy: for the remote sensing image with larger data volume, according to the size of a single target, cutting the original prediction image into blocks with the same size, and finally splicing the prediction result, so that the prediction efficiency can be improved, but the defect is that the splicing phenomenon among the blocks is obvious;

(2) multi-scale fusion: and continuously predicting the original image in blocks for multiple times, ensuring that the size and the position of each block are different, and fusing the results of the multiple predictions according to the probability of positive and negative samples of the prediction result of the pixel to finally form the prediction result of the whole image.

7. The method for applying the artificial intelligence recognition technology to the three-dimensional power transmission and transformation engineering measurement and calculation according to claim 1, wherein the method comprises the following steps: the method also includes point cloud extraction; the method comprises the steps of extracting a house from point cloud and extracting forest from the point cloud; point cloud extraction house: firstly, removing ground points, and extracting the house point cloud data according to the elevation difference of the target surface of the building and the reflection attribute of the laser pulse; point cloud extraction of forest trees: vegetation generally has an irregular shape, most of which vertically protrudes on the bare ground and occupies a certain continuous area; the surface coarse sugar degree formed by the vegetation point cloud is higher than that of a building or a bare ground, and the vegetation and other main ground objects are distinguished by utilizing the curved surface change and curvature of the point cloud; however, in case of a particularly rugged bare ground, such as a ground with a stone field, the method for determining the surface roughness may fail, and therefore, the characteristic of local point cloud density should be considered in addition to the rough roughness of the point cloud surface; because the laser beam can penetrate part of the crown to directly strike branches or the ground, whether the vegetation is the vegetation can be judged by utilizing the elevation difference between the first echo and the last echo of the point cloud data; however, the first echo point cloud includes vegetation, a building roof and a part of wall, and the last echo includes ground points and a part of vegetation, so that the elevation difference value of multiple echoes can be used for determining vegetation candidate points, and after building and wall objects are detected by other rules, the forest point cloud can be accurately identified.

8. The method for applying the artificial intelligence recognition technology to the three-dimensional power transmission and transformation engineering measurement and calculation according to claim 1, wherein the method comprises the following steps: the method also comprises a house vectorization modeling step;

the method comprises the following steps: separating different plane point clouds by adopting a method combining RANSAC method segmentation and distance method segmentation, solving corresponding plane equation parameters by utilizing foot points in each segmentation plane, and separating each roof plane point cloud;

step two: extracting a group contour node of each partition plane point cloud by using an Alpha Shape algorithm;

step three: extracting initial key points based on the direction angles, and further locking the detection range of the key points;

step four: determining a topological relationship between the roof planes by analyzing distances between the plane boundaries and the remaining plane boundaries;

step five: obtaining three intersection line characteristics including a common line segment characteristic, a common ray characteristic and a common straight line characteristic according to the topological relation between planes, and determining the accurate position of a final key point aiming at different intersection line characteristics;

step six: and constructing a roof model by using the detected key points, acquiring the side face of the building by combining corresponding ground information, and finally integrating to obtain a complete three-dimensional model of the building.

9. The method for applying the artificial intelligence recognition technology to the three-dimensional power transmission and transformation engineering measurement and calculation according to claim 1, wherein the method comprises the following steps: the method also comprises the following steps of house removal statistics and forest cutting statistics; and (3) house removal statistics: combining a three-dimensional platform, warehousing the vectorized house model, automatically counting the number, area and volume of houses in the line corridor range, and generating a dismantling distribution map; forest felling statistics: and (4) combining a three-dimensional platform, warehousing the divided single-wood models, and automatically counting the number of the trees in the range of the line corridor.

10. The method for applying the artificial intelligence recognition technology to the three-dimensional power transmission and transformation engineering measurement and calculation according to claim 1, wherein the method comprises the following steps: the method also comprises a single wood dividing step;

the method comprises the following steps: establishing a characteristic field;

step two: establishing a label field model, and determining the single-wood standing position through CHM;

step three: segmentation was performed using a Potts model.

Images