A disaster environment feature extraction method based on AT-EnFCM-SCNN
By combining sparse convolutional neural networks and fuzzy algorithms, the problems of recognition accuracy and efficiency in complex scenarios of traditional disaster environment monitoring systems are solved, and high-accuracy and efficient disaster environment feature extraction is achieved.
Patent Information
- Application Number
- CN202111537080.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-12-13
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2041-12-13
AI Technical Summary
Traditional disaster environmental monitoring systems have low recognition accuracy in complex disaster scenarios and are easily affected by image background and noise interference, leading to recognition errors and making it difficult to quickly and accurately identify disaster types.
An enhanced blur algorithm based on sparse convolutional neural networks (SCNN) is adopted. Through gradient image multi-scale reconstruction, watershed segmentation and cluster center calculation, combined with the loss function of Group Lasso regularization, the accurate extraction of disaster environmental features is achieved.
It improves the accuracy and efficiency of disaster environment identification, reduces noise interference, avoids over-segmentation, and ensures the integrity and accuracy of target extraction.
Smart Images

Figure CN114202536B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the fields of image processing and machine learning, and is a method for extracting features of a disaster environment. Background Art
[0002] At present, disaster environment monitoring relies on the analysis of geographical environment image changes, and the retrieval of various disaster-related images from the image library to identify the disaster type and issue early warnings. Traditional disaster environment monitoring systems have modular feature matching and clustering algorithms. For clustering algorithms, the image background will seriously affect the retrieval accuracy, and the change in the direction trend of the disaster environment will also reduce the image retrieval accuracy, resulting in image retrieval failure. The modular feature matching method combined with fuzzy information fusion and group monitoring methods is prone to feature value weight recognition errors, and the noise interference of key features affects the accurate recognition of disaster environment types. The diversity and complexity of disaster environments lead to a large probability of misidentification in the actual application of the above technologies.
[0003] In the problem of disaster environment identification, the most critical thing is the accuracy and efficiency of complex disaster scene identification. Accuracy is related to whether the type of disaster can be correctly identified and the corresponding judgment can be made. Identification efficiency affects the time to respond quickly after the disaster environment occurs, so higher requirements are put forward for the identification technology of environmental disasters. Therefore, in order to realize the identification and monitoring of disaster environment, an enhanced fuzzy algorithm based on sparse convolutional neural network is proposed. Through the effective segmentation of the image, the image features of the disaster environment are accurately extracted. Compared with the shortcomings of fixed image size and single shape in the deep learning segmentation method, it provides better local spatial information, effectively improves the pre-segmentation, reduces noise interference, avoids the incomplete extraction of targets caused by the transition segmentation phenomenon, and provides guarantee for the accurate and efficient identification of the disaster environment. Summary of the invention
[0004] In view of the problems existing in the above-mentioned prior art, the technical problem to be solved by the present invention is to provide a disaster environment feature extraction method based on AT-EnFCM-SCNN, and its specific process is as follows: Figure 1 shown.
[0005] The technical solution implementation steps are as follows:
[0006] (1) Determine the multi-scale reconstruction R of the gradient image rec (g,r1,r2):
[0007] First, the gradient image is reconstructed to determine R o (g)
[0008] R o (g) = R d (Re (g))
[0009] In the formula, g is the initial gradient image, R o For the reconstruction operation, R d is the expansion operation, R e is a corrosion operation. Multi-scale operations are performed on the open reconstruction results, and different scale radii are selected to reconstruct the gradient image. At the same time, these gradient images are fused to obtain the multi-scale open reconstruction R of the gradient image. rec (g,r1,r2):
[0010]
[0011] In the formula, r1 is the minimum scale radius, r2 is the maximum scale radius, R rec It is a scale reconstruction operation. Watershed is used for segmentation to determine the super-pixel image of the original disaster information.
[0012] (2) Determine the cluster center v k :
[0013] According to the superpixel image obtained in the previous step, determine the method of updating the cluster center:
[0014]
[0015] Where Q is the number of superpixel regions, l represents the number of regions, k is the number of categories of regions, and N l is the R of the lth region l The number of pixels in x p is a pixel with d-dimensional features, p is a point with features, u lk is the membership degree of the lth region to the kth category, and m is the fuzzy weight index. Then the distance between the cluster center and each pixel is calculated.
[0016] (3) Calculate the membership degree u lk :
[0017]
[0018] In the formula, C is the total number of cluster centers, j is the calculation unit, and v j The position corresponding to the calculation unit for j, ||*|| 2 is an arbitrary distance metric.
[0019] (4) Determine whether the iteration condition is met:
[0020]
[0021] Where n is the number of iteration steps, δ is the error threshold, and the iteration condition is met. Otherwise, the cluster center and membership degree are determined, and the iteration condition is judged. The class center label of each image area is obtained, and the label is clustered again to obtain two areas. The average value of the two areas represents the content of the area, and the segmentation result is output.
[0022] (5) Determine the loss function F(W) of the convolutional neural network:
[0023]
[0024] Where W is the set of all weights in the convolutional layer, and F D (W) is the empirical risk, which represents the error between the true value and the predicted value. R(·) is the structural risk, which is the L1 norm regularization term applied to each weight. It reduces or even clears the weight value by penalizing the sum of the absolute values of the weights, thereby reducing the amount of model calculation. λ is the constant corresponding to the structural risk, λ g is the constant corresponding to the structured sparse regularization term, and L is the total number of regions. g (·) is the structural sparsity regularization term. Group Lasso regularization is used to effectively make a large number of parameters in the convolutional layer zero, achieving high sparsity of each convolutional layer. The expansion is:
[0025]
[0026] In the formula, G is the number of groups divided by the overall weight, w (g ) is a group of g weights. Convolutional layers, pooling layers, and fully connected layers are constructed, and the above loss function is used to train the model to extract the features of the disaster environment information.
[0027] The present invention has the following advantages over the prior art:
[0028] (1) The present invention effectively solves the problem that the traditional disaster environment image segmentation has a single form and requires a large number of training samples, provides more local spatial information, and improves the pre-segmentation performance.
[0029] (2) The present invention effectively improves the problems of noise interference and over-segmentation in traditional disaster identification image processing, overcomes the incomplete target extraction at a single scale, and effectively ensures the correct extraction of larger global targets. BRIEF DESCRIPTION OF THE DRAWINGS
[0030] In order to better understand the present invention, further description is given below in conjunction with the accompanying drawings.
[0031] Figure 1 It is a flowchart of the steps to establish the disaster environment feature extraction method based on AT-EnFCM-SCNN;
[0032] Figure 2 It is to establish a flow chart of the disaster environment feature extraction method based on AT-EnFCM-SCNN;
[0033] Figure 3 It is a schematic diagram of the structure of the CNN neural network;
[0034] Figure 4 It is the accuracy result of extracting disaster environment characteristics using the present invention; Specific implementation plan
[0035] The present invention is further described in detail below through implementation cases.
[0036] The data used in this implementation case comes from a typical demonstration area of underground sheltered space, with a total of 1,000 sets of samples, including 5 scenes: underground tunnels, railway tunnels, underground shopping malls, underground parking lots, and subway stations. Each scene has 200 sets of data. A random sampling method is used to extract 140 sets of samples from each of the 5 scenes as training sets, and the rest as test sets. In the end, the total number of samples used as training sets is 700, and the total number of samples used as test sets is 300.
[0037] The overall flow chart of the disaster environment feature extraction method provided by the present invention is as follows: Figure 1 As shown, the specific steps are as follows:
[0038] (1) Determine the multi-scale reconstruction R of the gradient image rec (g,r1,r2):
[0039] First, the gradient image is reconstructed to determine R o (g)
[0040] R o (g) = R d (R e (g))
[0041] In the formula, g is the initial gradient image, R o For the reconstruction operation, R d is the expansion operation, R e is a corrosion operation. Multi-scale operations are performed on the open reconstruction results, and different scale radii are selected to reconstruct the gradient image. At the same time, these gradient images are fused to obtain the multi-scale open reconstruction R of the gradient image. rec (g,1,50):
[0042]
[0043] In the formula, r1 is the minimum scale radius, which is selected as 1 here, r2 is the maximum scale radius, which is selected as 50 here, R recIt is a scale reconstruction operation. Watershed is used for segmentation to determine the super-pixel image of the original disaster information.
[0044] (2) Determine the cluster center v k :
[0045] According to the superpixel image obtained in the previous step, determine the method of updating the cluster center:
[0046]
[0047] Where Q is the number of superpixel regions, l represents the number of regions, k is the number of categories of regions, and N l is the R of the lth region l The number of pixels in x p is a pixel with d-dimensional features, p is a point with features, u lk is the membership degree of the lth region to the kth category, and m is the fuzzy weight index, which is 2 here. Then the distance between the cluster center and each pixel is calculated.
[0048] (3) Calculate the membership degree u lk :
[0049]
[0050] In the formula, C is the total number of cluster centers, the size is 3, j is the calculation unit, and v j The position corresponding to the calculation unit for j, ||*|| 2 is an arbitrary distance metric.
[0051] (4) Determine whether the iteration condition is met:
[0052]
[0053] Where n is the number of iteration steps, δ is the error threshold, and the iteration condition is met. Otherwise, the cluster center and membership degree are determined, and the iteration condition is judged. The class center label of each image area is obtained, and the label is clustered again to obtain two areas. The average value of the two areas represents the content of the area, and the segmentation result is output.
[0054] (5) Determine the loss function F(W) of the convolutional neural network:
[0055]
[0056] Where W is the set of all weights in the convolutional layer, and F D(W) is the empirical risk, which represents the error between the true value and the predicted value. R(·) is the structural risk, which is the L1 norm regularization term applied to each weight. It reduces or even clears the weight value by penalizing the sum of the absolute values of the weights, thereby reducing the amount of model calculation. λ is the constant corresponding to the structural risk, λ g is the constant corresponding to the structured sparse regularization term, and L is the total number of regions. g (·) is the structural sparsity regularization term. Group Lasso regularization is used to effectively make a large number of parameters in the convolutional layer zero, achieving high sparsity of each convolutional layer. The expansion is:
[0057]
[0058] In the formula, G is the number of groups divided by the overall weight, w (g) There are g weight groups. Convolutional layers, pooling layers, and fully connected layers are constructed, and the above loss function is used to train the model to extract the features of disaster environment information.
[0059] The trained sparse convolutional neural network is tested using test set samples to realize feature extraction of disaster environment information, obtain the accuracy results of feature extraction, and complete the disaster environment feature extraction method based on AT-EnFCM-SCNN.
[0060] In order to verify the accuracy of the present invention in extracting disaster environment information features, four sets of disaster environment information experiments were conducted on the present invention. The experimental results are as follows: Figure 4 As shown. Figure 4 It can be seen that the disaster environment information feature extraction experimental method established by the present invention has an accuracy rate of more than 95% for disaster environment information feature extraction, which can achieve a high accuracy rate on the basis of ensuring stability, and the feature extraction effect is good. This shows that the disaster environment information feature extraction method established by the present invention is effective, and provides a better method for establishing an accurate disaster environment information feature extraction model, and has certain practicality.
Claims
1. A disaster environment feature extraction method based on AT-EnFCM-SCNN is characterized by: (1) determining the multi-scale open reconstruction of the gradient image, (2) determining the cluster center, (3) calculating the membership degree, (4) determining whether the iteration condition is met, and (5) determining the loss function of the convolutional neural network. Specifically, it includes the following five steps: Step 1: Determine the multi-scale reconstruction R of the gradient image rec (g,r1,r2); First, the gradient image is reconstructed to determine R o (g) R o (g) = R d (R e (g)); In the formula, g is the initial gradient image, R o For the reconstruction operation, R d is the expansion operation, R e is the corrosion operation; multi-scale operation is performed on the open reconstruction result, and different scale radii are selected to reconstruct the gradient image. At the same time, these gradient images are fused to obtain the multi-scale open reconstruction R of the gradient image. rec (g,r1,r2): In the formula, r1 is the minimum scale radius, r2 is the maximum scale radius, R rec It is a scale reconstruction operation; watershed segmentation is used to determine the super-pixel image of the original disaster information; Step 2: Determine the cluster center v k ; According to the superpixel image obtained in the previous step, determine the method of updating the cluster center: Where Q is the number of superpixel regions, l represents the number of regions, k is the number of categories of regions, and N l is the R of the lth region l The number of pixels in x p is a pixel with d-dimensional features, p is a point with features, u lk is the membership degree of the lth region to the kth category, and m is the fuzzy weight index; then the distance between the cluster center and each pixel is calculated; Step 3: Calculate the membership u lk ; In the formula, C is the total number of cluster centers, j is the calculation unit, and v j The position corresponding to the calculation unit for j, ||*|| 2 is any measure of distance; Step 4: Determine whether the iteration condition is met: Where n is the number of iteration steps, δ is the error threshold, until the iteration condition is met, otherwise continue to determine the cluster center and membership, continue to judge the iteration condition, obtain the type label of each image area, cluster the labels again, get two areas, use the average value of the two areas to represent the content of the area, and output the segmentation result; Step 5: Determine the loss function F(W) of the convolutional neural network; Where W is the set of all weights in the convolutional layer, and F D (W) is the empirical risk, which represents the error between the true value and the predicted value. R(·) is the structural risk, which is the L1 norm regularization term applied to each weight. It reduces or even clears the weight value by penalizing the sum of the absolute values of the weights, thereby reducing the amount of model calculation. λ is the constant corresponding to the structural risk, λ g is the constant corresponding to the structured sparse regularization term, L is the total number of regions; R g (·) is the structural sparsity regularization term. Group Lasso regularization is used to effectively make a large number of parameters in the convolutional layer zero, achieving high sparsity of each convolutional layer. The expansion is: In the formula, G is the number of groups divided by the overall weight, w (g) There are g weight groups; build convolutional layers, pooling layers, and fully connected layers, use the above loss function to train the model, and realize the feature extraction of disaster environment information.
