CN110569871A - saddle point identification method based on deep convolutional neural network - Google Patents
saddle point identification method based on deep convolutional neural network Download PDFInfo
- Publication number
- CN110569871A CN110569871A CN201910696431.9A CN201910696431A CN110569871A CN 110569871 A CN110569871 A CN 110569871A CN 201910696431 A CN201910696431 A CN 201910696431A CN 110569871 A CN110569871 A CN 110569871A
- Authority
- CN
- China
- Prior art keywords
- saddle
- neural network
- layer
- convolutional neural
- network model
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/048—Activation functions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Abstract
The invention discloses a saddle point identification method based on a deep convolutional neural network, which comprises the following steps of: 1) constructing a training sample set; 2) establishing an improved saddle feature deep learning convolutional neural network model; 3) setting various parameters required for training the improved saddle feature deep learning convolutional neural network model, training the improved saddle feature deep learning convolutional neural network model to enable the improved saddle feature deep learning convolutional neural network model to automatically identify potential saddle points in DEM data, correcting coordinates of the saddle points through a position regression MLP network, and determining initially selected saddle points; 4) and selecting a threshold with the recognition probability larger than a set threshold from the initially selected saddle points as a final saddle point, extracting the corrected coordinates of the final saddle point, and mapping the corrected coordinates of the final saddle point back to the original DEM to obtain a final saddle point identification map.
Description
Technical Field
The invention belongs to the technical field of element recognition in the field of digital terrain analysis, and relates to a saddle point recognition method based on a deep convolutional neural network.
Background
the digital terrain analysis takes a digital elevation model DEM as a basic data source, calculates terrain attributes, extracts terrain elements, and quantitatively analyzes the change of the earth surface form at multiple angles, thereby providing method support for scientific research and application related to geoscience. Among various types of topographic elements, regional topographic attributes, topographic structure lines, and the like have been analyzed and studied in detail for a long time. However, elements such as a mountain top, a saddle, a back slope, a basin and the like are limited by difficulties such as semantic ambiguity, multi-scale features, uncertain space support domain and the like, and the automatic extraction method becomes an attention point for digital terrain analysis research.
the existing saddle extraction method mainly adopts a local analysis window to analyze the geometrical morphological characteristics of the saddle, and extracts saddle points by calculating various surface factors and the interrelation among the factors in the analysis window. For example, a mountain saddle point hierarchical extraction method based on DEM [ J ] mapping science, 2013,38(2):181-183. "Zhang Wei with document", Tang Guan, pottery 26104, and the like ] is adopted in the document, and a saddle point extraction improvement method based on DEM confluence simulation [ J ] mapping science, 2011,36(1):158-159. "is used for researching and designing a saddle point extraction method and a confluence model simulation characteristic line intersection method of relative elevation low points on a basin boundary line; the document ' Kongyuanping, Yiwei, Zhaopng ' uses topological relation to quickly extract saddle points [ J ]. computer engineering and application, 2013,49(16):165-167 ' designs a saddle point extraction method for analyzing topological relation between subdivision equal height surfaces. The key of the methods is to establish a proper topographic element analysis neighborhood or topological relation description model, but the method can not pay attention to macroscopic and microscopic features such as saddle point space extension structures and element semantics, and realizes automatic element identification by establishing a deep learning neural network of saddle features.
disclosure of Invention
The invention aims to overcome the defects of the prior art and provides a saddle point identification method based on a deep convolutional neural network, which can automatically identify saddle points in DEM data through a deep learning neural network of saddle features.
in order to achieve the purpose, the saddle point identification method based on the deep convolutional neural network comprises the following steps:
1) constructing a training sample set, wherein the training sample set is formed by m + n DEM mountain data, m is equal to n, and n is a positive sample data set A including a saddle region, and is equal to { A ═ A1，A2，A3，…，AnM is a negative sample data set B not including the saddle region { B ═ B1，B2，B3，…，Bmrectangular cutting is carried out on DEM sample data, coordinates of saddle points in the positive sample are marked, and positive and negative sample data are divided into a training set and a verification set according to a set proportion;
2) analyzing the spatial morphological structure of the saddle, and establishing an improved saddle feature deep learning convolutional neural network model;
3) Setting various parameters required by training an improved saddle feature deep learning convolutional neural network model, wherein the various parameters comprise a target function, an optimization algorithm and a loss function, and pre-training the Lenet-5 network model by means of an ImageNet data set until the output of the Lenet-5 network model is optimal and reaches a stable state; initializing an improved saddle feature deep learning convolutional neural network model by using a pre-trained Lenet-5 network model, and then training the improved saddle feature deep learning convolutional neural network model by using the training set of positive sample data and the training set of negative sample data obtained in the step 1), so that the improved saddle feature deep learning convolutional neural network model can automatically identify potential saddle points in DEM data, and then correcting the coordinates of the saddle points by using a position regression MLP network to determine initially selected saddle points;
4) And selecting a threshold with the recognition probability larger than a set threshold from the initially selected saddle points as a final saddle point, extracting the corrected coordinates of the final saddle point, and mapping the corrected coordinates of the final saddle point back to the original DEM to obtain a final saddle point identification map, thereby completing the recognition of the saddle point based on the deep convolutional neural network.
the specific operation of the step 1) is as follows:
11) Downloading 1:100 million tiff format topographic data from an SRTM open source website, and then analyzing DEM elevation data from the topographic data;
12) observing the distribution condition of terrain elevation values in the DEM, and cutting off DEM mountain land data;
13) reading DEM mountain data, cutting out a complete saddle area according to morphological characteristics of a saddle, marking position coordinates of a saddle point, and storing the position coordinates in an excel format acceptable by a deep neural network to form a training sample set.
The specific operation of the step 2) is as follows:
The improved saddle feature deep learning convolutional neural network model comprises 1 input layer, 2 convolutional layers, 2 sampling layers, 2 full-link layers and 1 Softmax classification layer, wherein the input layer, the convolutional layers, the sampling layers and the full-link layers adopt correction linear units as activation functions, and weight parameters and bias parameters of the convolutional layers, the sampling layers and the full-link layers are initialized through a pre-trained Lenet-5 network model;
Inputting DEM data into the improved saddle feature deep learning convolution neural network model directly, wherein the convolution feature map k of the i-th layerias shown in formula (1):
wherein S isxand SyRepresenting the length and width, k, of the convolution kernel Si-1represents the input of the ith convolutional layer, wjnFor weights learned in the convolution kernel, biFor biasing of the i-th filter, σ () is an activation function, σ () uses a ReLU function, and the expression of the ReLU function is:
the improved pooling layer of the saddle feature deep learning convolutional neural network model adopts max posing, and the output of the pooling layer is as follows:
Wherein R is a sampling scale;
The convolutional layer and the pooling layer form a layered saddle point feature extractor, and excavation and learning of shallow and deep features of a saddle region are realized through multilayer superposition;
the output of the pooling layer is taken as the input of the full-link layer, and the output of the full-link layer is as follows:
wherein ReLU is chosen as the activation function σ (-),is the weight between the ith node on level l-1 and the jth node on level l,Is the current bias;
the softmax layer outputs the recognition probability P (c | P) that the DEM lattice point is a potential saddle point, wherein,
where c is the target class, L is the index of the last layer, NCis the number of targets.
The specific operation of the step 3) is as follows:
in order to enable the network to reach a stable state in the training process, the back propagation process of the network needs to be analyzed, wherein the back propagation process comprises a weight updating mode, a loss function expression mode and optimization algorithm selection for controlling gradient descent;
The backward propagation starts from the fully connected layer, and the operation shown in formula (6) is performed:
wherein E is a loss function, Is the first layerunit to l +1 th layerthe weight of the cell is such that,is the first layerUnit to l +1 th layerinputting the total unit;
in reverse propagation, the convolutional layer weight is adjusted by equation (7),
Wherein the content of the first and second substances,in order to be a non-linear mapping function,is equal to is equal torepeating the forward and backward propagation processes until the deviation between the predicted result and the real result is minimum;
Performing L2 regularization operation on the weight of the full connection layer, namely adding lambda sigma to the original loss function for the full connection relation of the full connection layerww2Punishment item to reduce the influence of noise in the training data on the network fitting process, wherein for the fully-connected layer with weight w, the expression of the loss function E is as follows:
Wherein E is0for the non-regularized loss function, λ is the weighted decay factor, 1- η λ is the weighted decay factor, wiis the learning rate.
the invention has the following beneficial effects:
The saddle point identification method based on the deep convolutional neural network comprises the steps of firstly constructing a training sample set, then analyzing the spatial morphological structure of a saddle, establishing an improved saddle characteristic deep learning convolutional neural network model, then training the improved saddle characteristic deep learning convolutional neural network model by using the training sample set, correcting the coordinates of the initially selected saddle point by using a position regression MLP network, and finally mapping the corrected saddle coordinate to an original DEM to obtain a saddle point identification map.
drawings
FIG. 1 is a flow chart of the present invention;
FIG. 2 is a block diagram of an improved saddle feature deep learning convolutional neural network model in accordance with the present invention;
FIG. 3 is a diagram of a saddle point position regression network MLP according to the present invention;
FIG. 4a is a three-dimensional panoramic view of test data I of the present invention;
FIG. 4b is a point diagram of a manually calibrated saddle for test data I of the present invention;
FIG. 4c is a graph of saddle point identification results on test data I in accordance with the present invention;
FIG. 4d is a diagram showing the recognition result of the test data I in the "method for rapidly extracting saddle points using topological relations" according to the present invention;
FIG. 5a is a three-dimensional panoramic view of test data II;
FIG. 5b is a saddle point diagram manually calibrated for test data II;
FIG. 5c is a graph of saddle point identification results on test data II in accordance with the present invention;
FIG. 5d is a diagram showing the recognition result of the test data II in the "method for rapidly extracting saddle points using topological relations";
FIG. 6a is a three-dimensional panoramic view of test data III;
FIG. 6b is a saddle point diagram manually calibrated for test data III;
FIG. 6c is a diagram showing the automatic saddle point identification result on data III according to the present invention.
FIG. 6d is a diagram showing the recognition result of the test data III in the "method for rapidly extracting saddle points using topological relations";
Detailed Description
the invention is described in further detail below with reference to the accompanying drawings:
referring to fig. 1, the saddle point identification method based on the deep convolutional neural network according to the present invention includes the following steps:
1) constructing a training sample set, wherein the training sample set is formed by m + n DEM mountain data, m is equal to n, and n is a positive sample data set A including a saddle region, and is equal to { A ═ A1，A2，A3，…，AnM is a negative sample data set B not including the saddle region { B ═ B1，B2，B3，…，BmRectangular cutting is carried out on DEM sample data, coordinates of saddle points in the positive sample are marked, and positive and negative sample data are divided into a training set and a verification set according to a set proportion;
The specific operation of the step 1) is as follows:
11) downloading 1:100 million tiff format topographic data from an SRTM open source website, and then analyzing DEM elevation data from the topographic data;
12) observing the distribution condition of terrain elevation values in the DEM, and cutting off DEM mountain land data;
13) reading DEM mountain data, cutting out a complete saddle area according to morphological characteristics of a saddle, marking position coordinates of a saddle point, and storing the position coordinates in an excel format acceptable by a deep neural network to form a training sample set.
specifically, downloading topographic data in a 1:100 million tiff format from an SRTM open source website, analyzing DEM elevation data from the topographic data, and converting the DEM elevation data into a txt text format;
Reading in txt format data, rendering terrain data into a gray map according to the height value, selecting a highlight area in the gray map, and cutting a 64 multiplied by 64 grid to obtain a DEM data block representing a mountain land;
Reading in a DEM mountain land data block, cutting out a complete saddle region of a 6 multiplied by 6 grid according to morphological characteristics of a saddle, marking coordinates of saddle points, and storing the saddle points into an excel format acceptable by a deep neural network to form a sample data set;
writing a training data generator, wherein the training data generator takes a 6 × 6 marked saddle region as a center, selects peripheral 6 × 6 DEM sub-images, namely cutting out 9 sub-images of 6 × 6 from each DEM mountain land data block, selects 1 sub-image containing a saddle point at the center as a positive sample, traverses the other 8 peripheral sub-images not containing the saddle point, discards the sub-images not meeting the size of 6 × 6, and randomly selects one sub-image from the rest sub-images as a negative sample;
2) Analyzing the spatial morphological structure of the saddle, and establishing an improved saddle feature deep learning convolutional neural network model;
the specific operation of the step 2) is as follows:
the improved saddle feature deep learning convolutional neural network model comprises 1 input layer, 2 convolutional layers, 2 sampling layers, 2 full-link layers and 1 Softmax classification layer, wherein the input layer, the convolutional layers, the sampling layers and the full-link layers adopt correction linear units as activation functions, and weight parameters and bias parameters of the convolutional layers, the sampling layers and the full-link layers are initialized through a pre-trained Lenet-5 network model;
Inputting DEM data into the improved saddle feature deep learning convolution neural network model directly, wherein the convolution feature map k of the i-th layeriAs shown in formula (1):
Wherein S isxand Syrepresenting the length and width, k, of the convolution kernel Si-1represents the input of the ith convolutional layer, wjnfor weights learned in the convolution kernel, bifor biasing of the i-th filter, σ () is an activation function, σ () uses a ReLU function, and the expression of the ReLU function is:
the improved pooling layer of the saddle feature deep learning convolutional neural network model adopts max posing, and the output of the pooling layer is as follows:
Wherein R is a sampling scale;
the convolutional layer and the pooling layer form a layered saddle point feature extractor, and excavation and learning of shallow and deep features of a saddle region are realized through multilayer superposition;
the output of the pooling layer is taken as the input of the full-link layer, and the output of the full-link layer is as follows:
Wherein ReLU is chosen as the activation function σ (-),Is the weight between the ith node on level l-1 and the jth node on level l,Is the current bias;
The softmax layer outputs the recognition probability P (c | P) that the DEM lattice point is a potential saddle point, wherein,
Where c is the target class, L is the index of the last layer, NCIs the number of targets.
3) setting various parameters required by training an improved saddle feature deep learning convolutional neural network model, wherein the various parameters comprise a target function, an optimization algorithm and a loss function, and pre-training the Lenet-5 network model by means of an ImageNet data set until the output of the Lenet-5 network model is optimal and reaches a stable state; initializing an improved saddle feature deep learning convolutional neural network model by using a pre-trained Lenet-5 network model, and then training the improved saddle feature deep learning convolutional neural network model by using the training set of positive sample data and the training set of negative sample data obtained in the step 1), so that the improved saddle feature deep learning convolutional neural network model can automatically identify potential saddle points in DEM data, and then correcting the coordinates of the saddle points by using a position regression MLP network to determine initially selected saddle points;
4) And selecting a threshold with the recognition probability larger than a set threshold from the initially selected saddle points as a final saddle point, extracting the corrected coordinates of the final saddle point, and mapping the corrected coordinates of the final saddle point back to the original DEM to obtain a final saddle point identification map, thereby completing the recognition of the saddle point based on the deep convolutional neural network.
The specific operation of the step 3) is as follows:
In order to enable the network to reach a stable state in the training process, the back propagation process of the network needs to be analyzed, wherein the back propagation process comprises a weight updating mode, a loss function expression mode and optimization algorithm selection for controlling gradient descent;
the backward propagation starts from the fully connected layer, and the operation shown in formula (6) is performed:
wherein E is a loss function, Is the first layerUnit to l +1 th layerthe weight of the cell is such that,Is the first layerUnit to l +1 th layerInputting the total unit;
in reverse propagation, the convolutional layer weight is adjusted by equation (7),
wherein the content of the first and second substances,in order to be a non-linear mapping function,Is equal to is equal torepeating the forward and backward propagation processes until the deviation between the predicted result and the real result is minimum;
performing L2 regularization operation on the weight of the full connection layer, namely adding lambda sigma to the original loss function for the full connection relation of the full connection layerww2Punishment item to reduce the influence of noise in the training data on the network fitting process, wherein for the fully-connected layer with weight w, the expression of the loss function E is as follows:
Wherein E is0For the non-regularized loss function, λ is the weighted decay factor, 1- η λ is the weighted decay factor, wiis the learning rate.
pre-training a Lenet-5 network model by means of an ImageNet data set until the output of the Lenet-5 network model is optimal and reaches a stable state, dividing a self-built positive sample set and a self-built negative sample set into a training set and a verification set, traversing the training set according to a specified batch size (batch _ size is 32), training an improved saddle feature deep learning convolutional neural network model, and completing the training of the improved saddle feature deep learning convolutional neural network model by setting training times, configuring a back propagation optimizer and the like, wherein the specific implementation of a training part is as shown in a Python program:
when the test accuracy of a test set in an improved saddle feature deep learning convolutional neural network model is high and the network is stable, training of a saddle coordinate correction part is carried out, firstly, a saddle point coordinate regression network model MLP is defined, the structure is shown in FIG. 3, parameters of each layer of CNN are frozen, only an MLP network part is trained, the MLP network part is the same as that of the CNN network training, initial parameters of the MLP are firstly set, an Adam optimization algorithm and mean square error are also selected as target functions, deviation change between saddle point prediction and a real position in the network training and testing process is used as a reference standard to judge whether the network is stable, and a Python program of the MLP network training is as follows:
In the training process of the complete network, the model parameters are set as in table 1:
TABLE 1
the diversity and complexity of landforms are key factors influencing accurate extraction of topographic elements, for example, China is taken as an example, China is high in the west and low in the terrain, the altitude from the west to the east is in a descending trend, China is divided into three steps according to the altitude change of the terrain and the terrain, the first step is a Qinghai-Tibet plateau area, the average altitude H is larger than or equal to 4500m, the Qinghai-Tibet plateau area is taken as a starting point and is extended to Taihang mountain and Wushan to form a second step with the most abundant landform types in China, the average altitude is more than or equal to 1000-2000 m, the third step is mainly based on hills, low mountains and plains, and the average altitude is more than 500m, and in order to verify the effectiveness of the invention in identifying the topographic points in the three steps of 1:100 million SRDEM-saddle, low mountain, medium and low mountain and high mountain data blocks I, II and III are selected for automatic identification test.
TABLE 2
The three-dimensional panoramas of the three test data are shown in fig. 4a, 5a and 6a, respectively, and the manual saddle point labeling results on the 3 data are shown in fig. 4b, 5b and 6b as reference standards. The automatic labeling result of the present invention is shown in fig. 4c, fig. 5c and fig. 6c, wherein the red dots are saddle points extracted by the network. In contrast, the recognition is performed by using a "saddle point fast extraction method using topological relations", and the results are shown in fig. 4d, fig. 5d and fig. 6d, where the black dots are extracted saddle points.
for clarity, the labeling result chart of the present invention marks that the correctly extracted saddle is the type a point, the incorrectly extracted saddle is the type B point, and the missed saddle is the type C point. According to experimental results, the saddle points marked by the method have fewer missed extraction and false extraction numbers, and the identified saddle points mostly accord with qualitative description of a saddle region and are located in an intersection line region between two mountain tops and between a ridge line and a valley line. Even if the saddle point is mistakenly extracted by the invention, the surrounding situation of the saddle point also conforms to the morphological characteristics of the saddle area, only a little deviation occurs during accurate positioning, and the deep network is learned by taking the morphology and the spatial structure presented by the saddle area as the main characteristics. The deep neural network trained by the invention only uses 6000 saddle sample data; a small amount of saddle points are missed, because the network training data volume is unknown, the cognition of the unknown saddle region situation change is limited, and the phenomena can be compensated by increasing the number of training samples.
Claims (4)
1. a saddle point identification method based on a deep convolutional neural network is characterized by comprising the following steps:
1) constructing a training sample set, wherein the training sample set is formed by m + n DEM mountain data, m is equal to n, and n is a positive sample data set A including a saddle region, and is equal to { A ═ A1，A2，A3，…，Anm is a negative sample data set B not including the saddle region { B ═ B1，B2，B3，…，Bmrectangular cutting is carried out on DEM sample data, coordinates of saddle points in the positive sample are marked, and positive and negative sample data are divided into a training set and a verification set according to a set proportion;
2) analyzing the spatial morphological structure of the saddle, and establishing an improved saddle feature deep learning convolutional neural network model;
3) setting various parameters required by training an improved saddle feature deep learning convolutional neural network model, wherein the various parameters comprise a target function, an optimization algorithm and a loss function, and pre-training the Lenet-5 network model by means of an ImageNet data set until the output of the Lenet-5 network model is optimal and reaches a stable state; initializing an improved saddle feature deep learning convolutional neural network model by using a pre-trained Lenet-5 network model, and then training the improved saddle feature deep learning convolutional neural network model by using the training set of positive sample data and the training set of negative sample data obtained in the step 1), so that the improved saddle feature deep learning convolutional neural network model can automatically identify potential saddle points in DEM data, and then correcting the coordinates of the saddle points by using a position regression MLP network to determine initially selected saddle points;
4) And selecting a threshold with the recognition probability larger than a set threshold from the initially selected saddle points as a final saddle point, extracting the corrected coordinates of the final saddle point, and mapping the corrected coordinates of the final saddle point back to the original DEM to obtain a final saddle point identification map, thereby completing the recognition of the saddle point based on the deep convolutional neural network.
2. The saddle point identification method based on the deep convolutional neural network as claimed in claim 1, wherein the specific operation of step 1) is:
11) downloading 1:100 million tiff format topographic data from an SRTM open source website, and then analyzing DEM elevation data from the topographic data;
12) Observing the distribution condition of terrain elevation values in the DEM, and cutting off DEM mountain land data;
13) reading DEM mountain data, cutting out a complete saddle area according to morphological characteristics of a saddle, marking position coordinates of a saddle point, and storing the position coordinates in an excel format acceptable by a deep neural network to form a training sample set.
3. the saddle point identification method based on the deep convolutional neural network as claimed in claim 1, wherein the specific operation of step 2) is as follows:
the improved saddle feature deep learning convolutional neural network model comprises 1 input layer, 2 convolutional layers, 2 sampling layers, 2 full-link layers and 1 Softmax classification layer, wherein the input layer, the convolutional layers, the sampling layers and the full-link layers adopt correction linear units as activation functions, and weight parameters and bias parameters of the convolutional layers, the sampling layers and the full-link layers are initialized through a pre-trained Lenet-5 network model;
Inputting DEM data into the improved saddle feature deep learning convolution neural network model directly, wherein the convolution feature map k of the i-th layerias shown in formula (1):
Wherein S isxAnd Syrepresenting the length and width, k, of the convolution kernel Si-1represents the input of the ith convolutional layer, wjnfor weights learned in the convolution kernel, bifor biasing of the i-th filter, σ () is an activation function, σ () uses a ReLU function, and the expression of the ReLU function is:
the improved pooling layer of the saddle feature deep learning convolutional neural network model adopts maxporoling, and the output of the pooling layer is as follows:
Wherein R is a sampling scale;
The convolutional layer and the pooling layer form a layered saddle point feature extractor, and excavation and learning of shallow and deep features of a saddle region are realized through multilayer superposition;
The output of the pooling layer is taken as the input of the full-link layer, and the output of the full-link layer is as follows:
wherein ReLU is chosen as the activation function σ (-),is the weight between the ith node on level l-1 and the jth node on level l,is the current bias;
the softmax layer outputs the recognition probability P (cp) that a DEM lattice point is a potential saddle point, where,
Where c is the target class, L is the index of the last layer, NCis the number of targets.
4. The saddle point identification method based on the deep convolutional neural network as claimed in claim 3, wherein the specific operation of step 3) is:
In order to enable the network to reach a stable state in the training process, the back propagation process of the network needs to be analyzed, wherein the back propagation process comprises a weight updating mode, a loss function expression mode and optimization algorithm selection for controlling gradient descent;
The backward propagation starts from the fully connected layer, and the operation shown in formula (6) is performed:
wherein E is a loss function, is the first layerUnit to l +1 th layerThe weight of the cell is such that,is the first layerunit to l +1 th layerInputting the total unit;
In reverse propagation, the convolutional layer weight is adjusted by equation (7),
wherein the content of the first and second substances,in order to be a non-linear mapping function,is equal to is equal toRepeating the forward and backward propagation processes until the deviation between the predicted result and the real result is minimum;
Performing L2 regularization operation on the weight of the full connection layer, namely adding lambda sigma to the original loss function for the full connection relation of the full connection layerww2Punishment item to reduce the influence of noise in the training data on the network fitting process, wherein for the fully-connected layer with weight w, the expression of the loss function E is as follows:
wherein E is0For the non-regularized loss function, λ is the weighted decay factor, 1- η λ is the weighted decay factor, wiIs the learning rate.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910696431.9A CN110569871B (en) | 2019-07-30 | 2019-07-30 | Saddle point identification method based on deep convolutional neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910696431.9A CN110569871B (en) | 2019-07-30 | 2019-07-30 | Saddle point identification method based on deep convolutional neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110569871A true CN110569871A (en) | 2019-12-13 |
| CN110569871B CN110569871B (en) | 2022-09-30 |
Family
ID=68773068
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910696431.9A Active CN110569871B (en) | 2019-07-30 | 2019-07-30 | Saddle point identification method based on deep convolutional neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110569871B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113283429A (en) * | 2021-07-21 | 2021-08-20 | 四川泓宝润业工程技术有限公司 | Liquid level meter reading method based on deep convolutional neural network |
| CN113411106A (en) * | 2021-05-31 | 2021-09-17 | 海南大学 | Power distribution method based on deep learning in safe space modulation system |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102567701A (en) * | 2010-12-08 | 2012-07-11 | 中国科学院地理科学与资源研究所 | Method for automatically extracting circular impact craters from ChangE DEM (Dynamic Effect Model) data by Hough transformation |
| CN104156571A (en) * | 2014-07-25 | 2014-11-19 | 山西潞安环保能源开发股份有限公司 | Automatic grading detection method for primary mountain range and water system-directional derivative wavelet transformation method |
| CN104899865A (en) * | 2015-05-04 | 2015-09-09 | 西安建筑科技大学 | Mountain line extraction method based on Gaussian scale space |
| CN106033611A (en) * | 2016-05-23 | 2016-10-19 | 西安建筑科技大学 | Method for extracting surface peaks in DEM data |
| CN107895370A (en) * | 2017-11-29 | 2018-04-10 | 南京师范大学 | A kind of saddle point extracting method based on contour |
| US20180101766A1 (en) * | 2016-10-11 | 2018-04-12 | Siemens Aktiengesellschaft | Efficient calculations of negative curvature in a hessian free deep learning framework |
| CN108510467A (en) * | 2018-03-28 | 2018-09-07 | 西安电子科技大学 | SAR image target recognition method based on variable depth shape convolutional neural networks |
| CN109145171A (en) * | 2018-07-23 | 2019-01-04 | 广州市城市规划勘测设计研究院 | A kind of multiple dimensioned map data updating method |
| CN109739942A (en) * | 2018-12-14 | 2019-05-10 | 南京泛在地理信息产业研究院有限公司 | A kind of saddle point extracting method based on contour line model |
-
2019
- 2019-07-30 CN CN201910696431.9A patent/CN110569871B/en active Active
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102567701A (en) * | 2010-12-08 | 2012-07-11 | 中国科学院地理科学与资源研究所 | Method for automatically extracting circular impact craters from ChangE DEM (Dynamic Effect Model) data by Hough transformation |
| CN104156571A (en) * | 2014-07-25 | 2014-11-19 | 山西潞安环保能源开发股份有限公司 | Automatic grading detection method for primary mountain range and water system-directional derivative wavelet transformation method |
| CN104899865A (en) * | 2015-05-04 | 2015-09-09 | 西安建筑科技大学 | Mountain line extraction method based on Gaussian scale space |
| CN106033611A (en) * | 2016-05-23 | 2016-10-19 | 西安建筑科技大学 | Method for extracting surface peaks in DEM data |
| US20180101766A1 (en) * | 2016-10-11 | 2018-04-12 | Siemens Aktiengesellschaft | Efficient calculations of negative curvature in a hessian free deep learning framework |
| CN107895370A (en) * | 2017-11-29 | 2018-04-10 | 南京师范大学 | A kind of saddle point extracting method based on contour |
| CN108510467A (en) * | 2018-03-28 | 2018-09-07 | 西安电子科技大学 | SAR image target recognition method based on variable depth shape convolutional neural networks |
| CN109145171A (en) * | 2018-07-23 | 2019-01-04 | 广州市城市规划勘测设计研究院 | A kind of multiple dimensioned map data updating method |
| CN109739942A (en) * | 2018-12-14 | 2019-05-10 | 南京泛在地理信息产业研究院有限公司 | A kind of saddle point extracting method based on contour line model |
Non-Patent Citations (5)
| Title |
|---|
| ROCIO NAHIME TORRES ET AL.: "A Deep Learning model for identifying mountain summits in Digital Elevation Model data", 《2018 IEEE FIRST INTERNATIONAL CONFERENCE ON ARTIFICIAL INTELLIGENCE AND KNOWLEDGE ENGINEERING (AIKE)》 * |
| 吴婕等: "基于DEM的山顶点和鞍部点一体化提取", 《测绘地理信息》 * |
| 易炜: "基于DEM的地形特征提取算法研究及应用", 《中国优秀博硕士学位论文全文数据库(硕士) 信息科技辑》 * |
| 汤国安等: "我国区域地貌数字地形分析研究进展", 《测绘学报》 * |
| 熊礼阳等: "基于DEM的山地鞍部点分级提取方法", 《测绘科学》 * |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113411106A (en) * | 2021-05-31 | 2021-09-17 | 海南大学 | Power distribution method based on deep learning in safe space modulation system |
| CN113283429A (en) * | 2021-07-21 | 2021-08-20 | 四川泓宝润业工程技术有限公司 | Liquid level meter reading method based on deep convolutional neural network |
| CN113283429B (en) * | 2021-07-21 | 2021-09-21 | 四川泓宝润业工程技术有限公司 | Liquid level meter reading method based on deep convolutional neural network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110569871B (en) | 2022-09-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108399428B (en) | Triple loss function design method based on trace ratio criterion | |
| CN112734775B (en) | Image labeling, image semantic segmentation and model training methods and devices | |
| CN111126386B (en) | Sequence domain adaptation method based on countermeasure learning in scene text recognition | |
| CN114092832B (en) | High-resolution remote sensing image classification method based on parallel hybrid convolutional network | |
| CN110033008B (en) | Image description generation method based on modal transformation and text induction | |
| CN106709486A (en) | Automatic license plate identification method based on deep convolutional neural network | |
| CN110866494B (en) | Urban group extraction method and system based on optical remote sensing image | |
| CN110222767B (en) | Three-dimensional point cloud classification method based on nested neural network and grid map | |
| CN108052966A (en) | Remote sensing images scene based on convolutional neural networks automatically extracts and sorting technique | |
| CN108446616B (en) | Road extraction method based on full convolution neural network ensemble learning | |
| CN113033520A (en) | Tree nematode disease wood identification method and system based on deep learning | |
| CN107516128A (en) | A kind of flowers recognition methods of the convolutional neural networks based on ReLU activation primitives | |
| CN114445634A (en) | Sea wave height prediction method and system based on deep learning model | |
| CN112950780B (en) | Intelligent network map generation method and system based on remote sensing image | |
| CN110929746A (en) | Electronic file title positioning, extracting and classifying method based on deep neural network | |
| CN109145971A (en) | Based on the single sample learning method for improving matching network model | |
| CN113313094B (en) | Vehicle-mounted image target detection method and system based on convolutional neural network | |
| CN110569871B (en) | Saddle point identification method based on deep convolutional neural network | |
| CN116699096B (en) | Water quality detection method and system based on deep learning | |
| CN111339932B (en) | Palm print image preprocessing method and system | |
| CN111310820A (en) | Foundation meteorological cloud chart classification method based on cross validation depth CNN feature integration | |
| CN111414951A (en) | Method and device for finely classifying images | |
| CN112818777B (en) | Remote sensing image target detection method based on dense connection and feature enhancement | |
| CN116012709B (en) | High-resolution remote sensing image building extraction method and system | |
| CN115457366A (en) | Chinese herbal medicine multi-label recognition model based on graph convolution neural network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |