CN109871875B

CN109871875B - Building change detection method based on deep learning

Info

Publication number: CN109871875B
Application number: CN201910054336.9A
Authority: CN
Inventors: 林秋华; 孙斌
Original assignee: Dalian University of Technology
Current assignee: Dalian University of Technology
Priority date: 2019-01-21
Filing date: 2019-01-21
Publication date: 2021-01-19
Anticipated expiration: 2039-01-21
Also published as: CN109871875A

Abstract

The invention discloses a building change detection method based on deep learning, and belongs to the technical field of computer vision. Firstly, a building is segmented by applying a deep learning image segmentation algorithm U-net to obtain a binary segmentation image; then, combining the binary segmentation images to generate a building area combined binary image; then, taking the merged image as a masking image, and removing noise ground objects in the input image to obtain a noiseless image; and finally, carrying out change detection on the noiseless image by adopting an unsupervised deep learning network PCAnet, and selecting and outputting an optimal change image. Compared with the existing building change detection method, the method utilizes deep learning to carry out change detection without a large amount of labeled training data; when noise and ground objects such as trees, vehicles, pedestrians and the like interfere in the input image, change detection is performed only for the building area. The test result shows that the accuracy of the method is improved by 7% and the false alarm rate is reduced by 59.8% compared with the GDBM model method.

Description

Building change detection method based on deep learning

Technical Field

The invention relates to the technical field of computer vision, in particular to a building change detection method based on deep learning.

Background

The building is a ground object closely related to human life, densely distributed and wide. With the continuous development of remote sensing technology and aerial photography technology, it becomes easier to acquire high-resolution satellite images or aerial images of buildings. The building change detection technology is to analyze the same building image obtained at different time periods, so as to obtain the change information of the building. Along with the lapse of time, three major changes such as new construction, demolition, reconstruction etc. mainly take place for the building. Some buildings change into legal update, reflecting the development of cities and society. Meanwhile, changes of violation also exist, which affect the urban appearance and restrict social progress. Therefore, the high-precision building change detection is carried out, so that the illegal building area extraction is facilitated, the dynamic monitoring of the building is realized, and the building distribution is improved in time; the method can also realize the rapid and accurate statistics of the buildings, and plays an important role in urban planning, geographic information updating, land resource management and the like.

Building change detection methods can be roughly classified into two types, including methods based on classical image processing and methods based on deep learning, according to the difference of detection algorithms. The method based on classical image processing mainly adopts pixel level, feature level or object level image processing, so that more manual participation is realized, and the probability of detecting a pseudo-change region is higher. For example, the pixel-level based detection method only considers the characteristics of a single pixel, lacks spatial information of neighboring pixels, and the like, so that the detection result is sensitive to noise and has poor robustness. The effect of the feature level detection method depends on the result of building feature extraction, and when the change detection is performed by using a single feature, more missed detections and false alarms may be caused. The object-level-based detection method is based on image segmentation, pixels with the same properties are combined to form an object, and then change detection is carried out, so that the defects of the pixel-level detection method are improved to a certain extent. However, the traditional image segmentation method is difficult to realize the optimal extraction of the similar objects, and further influences the detection result of the building change. In addition, some methods based on classical image processing require digital surface models to be constructed, geographic information systems to be utilized, and the like, and are high in cost and complex in calculation.

In recent years, deep learning has become one of the research hotspots in the field of computer vision. With continuous optimization and improvement of the deep neural network, particularly with the proposal of the convolutional neural network, the feature extraction capability of the network is further improved, and the fields of image classification, recognition, segmentation and the like are rapidly developed. The deep learning network can automatically learn deep variation characteristics under various conditions, and the defects of more manual participation, poor robustness, limited characteristic extraction capability and the like in a classical variation detection method are overcome. Therefore, methods based on deep learning have become an important direction for current building change detection research. The existing typical method is that the change detection is realized by taking a change area in an image as a target to be detected and an unchanged area as a background by utilizing the excellent target identification capability of an area fast convolution neural network; and extracting features by using a twin convolution network, and judging a change area by measuring the distance between feature vectors. One common feature of these methods is that a large amount of labeled image data is required as a training set to train the neural network to learn image features. When the labeled training sample is insufficient, it is difficult to achieve the ideal detection effect. And a method for detecting the change of the building by using a Gaussian-Bernoulli depth Boltzmann Machine (GDBM) model, extracting most possible change and unchanged area samples from a change intensity map obtained by preprocessing by setting a self-adaptive sampling interval, training the model to extract change characteristics, and finally generating a change detection map. The method can reduce the requirement on the amount of the labeled training data, but the detection accuracy is not high enough.

The analysis of the current change detection method based on deep learning shows that except that most methods require a large amount of labeled training data, the methods directly judge the change of the whole image to be detected. However, in an image generally used for building change detection, not only buildings but also a large number of noise features exist. The noise feature refers to features of other forms besides buildings, such as trees, vehicles, pedestrians, etc., which inevitably interfere with the detection result to different degrees. In other words, detecting the full map increases the probability of detecting the false change region, i.e., the false alarm rate increases, and at the same time, the accuracy decreases.

Disclosure of Invention

The invention aims to adopt an unsupervised deep learning network to carry out change detection, such as a change detection method based on the unsupervised deep learning network PCANet (Feng Gao, JunyuDong, BoLi, and Qizhi Xu, "Automatic change detection in synthetic aperture images based on PCANet," IEEE Geoscience and Remote Sensing Letters vol, 13, 12, pp.1792-1796,2016), and reduce the requirement on the data volume of label images; by image segmentation, noise ground objects which do not belong to buildings are removed to the maximum extent, so that the accuracy is improved, and the false alarm rate is reduced.

The technical scheme of the invention is that two input images to be detected are given

(i 1.. m; j 1.. n, mxn is the image size); first, a deep learning Image segmentation algorithm such as U-Net (Olif Ronneberger, Philipp Fische, andTimes Brox, "U-Net: volumetric network for biological Image segmentation," Medical Image Computing and Computer-Assisted interaction, vol.9351, pp.234-241,2015) is applied to the I₁(i,j)、I₂(I, j) carrying out building segmentation to obtain a binary segmentation image I_s1(i,j)、I_s2(i, j); then, the binary divided image I_s1(i,j)、I_s2(I, j) merging to generate a building area merged binary image I_M(i, j); then, to merge the image I_M(I, j) removing the input image I for masking image₁(i,j)、I₂(I, j) to obtain a noise-free image I_p1(i,j)、I_p2(i, j); finally, change detection is carried out on the noiseless image by adopting an unsupervised deep learning network such as PCANet, and the optimal change image is selected and output

The specific implementation steps (see fig. 1) are as follows:

the first step is as follows: inputting two images I to be detected₁(i,j)、I₂(i, j) the image single-channel bit depth is B.

The second step is that: and (5) dividing the building area. Using the trained deep learning image segmentation network U-Net (the network structure is shown in FIG. 2), respectively for I₁(i,j)、I₂(i, j) performing building region segmentation to obtain a binary segmentation image

The specific steps of training U-Net (see FIG. 3) are as follows:

step 1: input training set

Wherein N is_RRepresenting the total number of samples in the training set; a set of hyper-parameters required for network training is set, including the learning rate alpha (generally 10)^-3～10^-6) And the number of training rounds epochs (generally 10-20).

Step 2: and starting to train the U-Net network.

And step 3: adjustment of learning rate α (× 10)^-1) And the training round number epochs (+1), and selecting the network parameters with the best segmentation effect (the highest accuracy) from the results obtained by training of a plurality of groups of different hyper-parameters.

And 4, step 4: and outputting the trained U-Net network.

The third step: and merging the building segmentation maps. Merging the binary segmentation images I by adopting the following formula_s1(i,j)、I_s2(i, j) generating a building region-merged binary image

The fourth step: and removing noise and ground objects. Merging binary images I using building regions_M(I, j) for two images I to be detected₁(i,j)、I₂(i, j) removing noise and feature to generate a noise-free image

The fifth step: building change detection. Using unsupervised network PCANet pair I_p1(i,j)、I_p2(i, j) detecting the change. The change detection algorithm based on PCANet is implemented by the following steps (see FIG. 4):

step 1: to I_p1(i,j)、I_p2(i, j) carrying out logarithmic ratio processing to generate a rough estimation change image

Wherein "| · |" represents a modulo operation.

Step 2: pixel pre-classification and labeling. For image I_DExtracting Gabor characteristics from each pixel in (I, j), and matching with Fuzzy C-Means clustering (FCM) algorithm to carry out I pair_DClustering pixels in (i, j) to obtain three types of labeled classes, wherein the change w is (1) in sequence_c: the subscript c indicates a change, and the intra-class pixel (i, j) is labeled as l _ij1 ═ 1; (2) unchanged w_uc: the subscript uc indicates no change, and the pixel (i, j) within the class is labeled as l_ij-1; (3) to be determined w_ud: the subscript ud indicates that the pending, intra-class pixel (i, j) is labeled as l_ij0; the number of pixels included in the three categories is N_c、N_uc、N_udAnd N is_c+N_uc+N_ud＝m×n。

And step 3: and calculating the number of training sample pixels. From step 2: randomly selecting training sample pixels from all m multiplied by n pixels obtained by pixel pre-classification and marking, screening positive sample pixels (changed pixels) and negative sample pixels (unchanged pixels) according to a certain proportion, wherein the label of the training sample pixels (i, j) is l_ijE { -1, +1 }. Total number of training sample pixels N_sumCalculated using the following formula:

N_sum＝m×n×r (7)

where r represents the proportion of randomly chosen sample pixels. Number of positive sample pixels N_posCalculated using the following formula:

negative sample pixel number N_negThen it is:

N_neg＝N_sum-N_pos (9)

and 4, step 4: and generating a training set. In the training set, the generation process of the training sample image is shown in fig. 5; in two input images I respectively_p1(i,j)、I_p2(i, j) selecting image block with size of k × k (k is typically odd number in 3-11) with sample pixel (i, j) as center

Then, two image blocks are put together

Longitudinally connected to form a sample image P_yThe size is 2k × k, i.e.:

sample image P_yLabel l of_yIs the label l of the sample pixel (i, j)_ijI.e. l_yE { -1, +1 }. Co-generation of N_sumA sample image, whereinN_posPositive sample, N_negA negative sample to form a training set

And 5: extracting PCANet characteristics and generating a characteristic vector set. Extracting a training set through an unsupervised deep learning network PCANet

Feature vector F of each sample image_yAnd a label l_yTogether forming a feature vector set

Firstly, a group of parameters to be set for training PCANet is set, wherein the parameters include the number of stages (generally 2) for performing PCA filter convolution (PCA filters convolution), and the number L of filters in each stage_t(generally 5 to 8) t 1.

Step 6: and (5) training a classifier. Using feature vector sets

And training a classifier. The classifier uses a linear kernel Support Vector Machine (Liner SVM). The parameter needed to be set for training the Liner SVM is only one coefficient in the objective function of the classifier, called as an error penalty factor, and is generally 1-10.

And 7: the pixels are reclassified. For step 2: undetermined classification w obtained by pixel pre-classification and marking_udGenerating an image by the method shown in fig. 5 for each pixel; then extracting features of the image by using PCANet; inputting the obtained feature vectors into a Liner SVM which finishes training for classification; dividing pixels into variation classes w according to output results of Liner SVM_{ud_c}Or unchanged class w_{ud_uc}(ii) a Finally, synthesize w obtained in step 2_c、w_ucPixels in class, then full image pixels are classified as changes

And is not changed

In the formula, "u" represents "union".

And a sixth step: generating a change image I_o(i,j)∈^m×n. Set of pixels derived from classification

Generating a binary change image by the following formula:

the seventh step: and selecting the optimal variation image. Fixing the number of PCA filtering convolution stages to be 2, and adjusting the number L of filters in the two PCA filtering convolution stages₁、L₂(5-8 are respectively taken, step 1 is taken), the image block size parameter k (odd number in 3-11) and the error penalty factor C (1-10 are taken, step 1) of the Liner SVM are taken, and the accuracy ACC and the omission factor P of each change image obtained in the sixth step are calculated_MDFalse alarm rate P_FA(ii) a According to ACC, P_MD、P_FASelecting relatively optimal change image from high priority to low priority

And its corresponding parameters. The calculation formula of each evaluation index is as follows:

wherein N is_TC、N_FCRespectively representing the number of pixels that are correctly detected and the number of pixels that are detected as being changed but not actually changed, N_TUC、N_FUCRespectively representing the number of unchanged pixels correctly detected and the number of pixels detected as unchanged but actually changed.

Eighth step: outputting an optimal change image

Compared with the existing building change detection method, the method has the advantages that the change detection is carried out by utilizing deep learning, and a large amount of labeled training data is not needed; when noise ground objects such as trees, vehicles, pedestrians and the like interfere in the input image, change detection is only carried out on the area of the building, accuracy is improved, and false alarm rate is reduced. For example, building change detection is carried out on aerial images of 2008 and 2013 in a certain area, the accuracy of the existing building change detection method by using a GDBM model is 83.64%, and the false alarm rate is 25.36% (Zhang Xinlong, Chen Xiuwan, Lifei, etc.. deep learning change detection method of high-resolution remote sensing images [ J ] survey and drawing report, 2017,46(8): 999-. In comparison, the method can achieve 89.50% of correct rate, and the false alarm rate is only 10.20%; the accuracy is improved by 7%, and the false alarm rate is reduced by 59.8%.

Drawings

Fig. 1 is a general block diagram of the present invention.

Fig. 2 is a structural diagram of the deep learning image segmentation network U-Net.

FIG. 3 is a training block diagram of U-Net.

Fig. 4 is a block diagram of a change detection method based on PCANet.

Fig. 5 is a diagram of a sample image generation process in the change detection method based on PCANet.

Fig. 6 is an aerial image of two years. (a)2008, aerial images; (b) images were taken by plane in 2013.

Fig. 7 is a diagram of the building segmentation result. (a)2008, aerial images; (b) images were taken by plane in 2013.

FIG. 8 is a merged view of two aerial image building segmentation maps.

Fig. 9 is a diagram showing the aerial image noise feature removal result. (a)2008, aerial images; (b) images were taken by plane in 2013.

Fig. 10 is a rough estimation change image.

Fig. 11 is a diagram showing a result of detecting a change in a building and a diagram showing a true value of the change. (a) The result chart of the patent; (b) a building change truth map.

Detailed Description

The following describes a specific embodiment of the present invention in detail with reference to the technical solution and the attached drawings.

Two existing aerial images are aerial images of a certain area in 2008 and 2013, the size of each aerial image is 512 x 512, and the single-channel bit depth B of each aerial image is 8. The building change detection by adopting the invention is as follows:

the first step is as follows: inputting aerial images of 2008 and 2013, and recording as I₁(i,j)、I₂(i, j) as shown in FIGS. 6(a) and (b).

The second step is that: and (5) dividing the building area. Using the trained deep learning image segmentation network U-Net (the network structure is shown in FIG. 2), respectively for I₁(i,j)、I₂(i, j) dividing the building region to obtain a binary divided image according to the formulas (1) and (2) and the image bit depth

As shown in fig. 7(a) and (b). The specific steps of training U-Net (see FIG. 3) are as follows:

step 1: input training set

Using a data set published by CSU-DP laboratory collation, the spatial resolution of the data set was 0.22m, and a total of 1233 aerial images of 512X 512 size, N_R1233; the hyper-parameter required to set a set of training networks is the learning rate α 10^-3The training round number epochs is 10.

Step 2: and starting to train the U-Net network.

And step 3: adjustment of learning rate α (× 10)^-1Get 10 immediately^-4,10^-4,10^-5,10^-6) And training round number epochs (+1, namely 11,12,13,14,15,16,17,18,19 and 20) are taken, and among the results obtained by the sets of hyper-parameters, the network parameter with the best segmentation effect (highest accuracy) is selected.

And 4, step 4: and outputting the trained U-Net network.

The third step: and merging the building segmentation maps. Merging the binary segmentation images I by adopting a formula (3)_s1(i,j)、I_s2(i, j) generating a building region-merged binary image

As shown in fig. 8.

The fourth step: and removing noise and ground objects. Combining the binary images I by using the building region by adopting the formulas (4) and (5)_M(I, j) for two images I to be detected₁(i,j)、I₂(i, j) removing noise and feature to generate a noise-free image

As shown in fig. 9(a) and (b).

step 1: using formula (6) to I_p1(i,j)、I_p2(i, j) carrying out logarithmic ratio processing to generate a rough estimation change image

As shown in fig. 10.

Step 2: pixel pre-classification and labeling. For image I_DExtracting Gabor characteristics of each pixel in (I, j), and matching with an FCM algorithm to carry out I_DClustering pixels in (i, j) to obtain three types of labeled classes, wherein the change w is (1) in sequence_cAnd the intra-class pixel (i, j) is labeled as l _ij1 ═ 1; (2) unchanged, intra-class pixel (i, j) is labeled as l_ij-1; (3) to be determined w_udAnd the intra-class pixel (i, j) is labeled as l_ij0; the number of pixels included in the three categories is N_c＝204018、N_uc＝40563、N_ud＝17563，N_c+N_uc+N_ud＝512×512。

And step 3: and calculating the number of training sample pixels. From step 2: randomly selecting training sample pixels from all pixels obtained by pixel pre-classification and marking, screening positive sample pixels (changed pixels) and negative sample pixels (unchanged pixels) according to the proportion that r is 30%, wherein the label of the training sample pixel (i, j) is l_ijE { -1, +1 }. Calculating the total number of pixels of the training sample to be N by adopting the formula (7)_sum78643; respectively calculating the total number of the positive and negative sample pixels to be N by adopting formulas (8) and (9)_pos＝65600、N_neg＝13043。

And 4, step 4: and generating a training set. In the training set, the generation process of the training sample image is shown in fig. 5; in two input images I respectively_p1(i,j)、I_p2In (i, j), an image block of size 7 × 7 (k: 7) is selected with the sample pixel (i, j) as the center

Then, two image blocks are mapped by equation (10)

Longitudinally connected to form a sample image

Sample image P_yLabel l of_yIs the label l of the sample pixel (i, j)_ijI.e. l_yE { -1, +1 }. 78643 sample images can be generated in total, wherein 65600 positive samples and 13043 negative samples form a training set

Firstly, a group of parameters required for training PCANet is set, wherein the parameters comprise a stage number of stages of 2, and the number of filters in each stage is L₁＝7、L₂＝7。

Step 6: and (5) training a classifier. Using feature vector sets

And training a classifier. The classifier selects a Liner SVM. And setting the hyperparametric error penalty factor as C-1.

And 7: the pixels are reclassified. For step 2: undetermined classification w obtained in pixel pre-classification and labeling_udGenerating an image by the method shown in fig. 5 for each pixel; then extracting features of the image by using PCANet; inputting the obtained feature vectors into a Liner SVM which finishes training for classification; dividing pixels into variation classes w according to output results of Liner SVM_{ud_c}Or unchanged class w_{ud_uc}(ii) a Finally, synthesizing w obtained in step 2 by adopting formulas (11) and (12)_c、w_ucPixels in classes, classifying full image pixels as changes

And is not changed

And a sixth step: generating a change image I_o(i,j)∈^512×512. Set of pixels derived from classification

Generating a binary change image I using equation (13)_o(i,j)。

The seventh step: and selecting the optimal variation image. Adjusting the number L of filters in two PCA filtering convolution stages₁、L₂(5, 6 and 8 are respectively taken), an image block size parameter k (3, 5,9 and 11 is taken), an error penalty factor C (2, 3,4,5,6,7,8,9 and 10 is taken) of the Liner SVM, and the accuracy ACC and the omission factor P of each change image obtained in the sixth step are calculated by adopting the formulas (14), (15) and (16)_MDFalse alarm rate P_FA(ii) a According to ACC, P_MD、P_FASelecting relatively optimal change image from high priority to low priority

And the corresponding parameters: number L of filters in two PCA filtering convolution stages₁、L₂Respectively 8,9 for image block size parameter k and 1 for error penalty factor C of the Liner SVM.

Eighth step: outputting an optimal change image

As shown in fig. 11(a), fig. 11(b) is a change true value map.

Claims

1. A building change detection method based on deep learning is characterized by comprising the following steps:

the first step is as follows: inputting two images I to be detected₁(i,j)、I₂(i, j), the single-channel bit depth of the image is B;

the second step is that: dividing a building area; respectively carrying out I pair on I pair by utilizing a trained deep learning image segmentation network U-Net₁(i,j)、I₂(i, j) is carried outDividing the building region to obtain a binary divided image

The U-Net training method comprises the following specific steps:

step 1: input training set

Wherein N is_RRepresenting the total number of samples in the training set; setting a group of hyper-parameters required by network training, including learning rate alpha, taking 10^-3～10^-6(ii) a Training the number of epochs, and taking 10-20;

step 2: starting to train the U-Net network;

and step 3: adjusting the learning rate alpha, each time multiplying by 10^-1(ii) a Adjusting the number of training rounds epochs, adding 1 each time, and selecting the network parameters with the best segmentation effect, namely the highest accuracy rate, from the results obtained by training of a plurality of groups of different hyper-parameters;

and 4, step 4: outputting the trained U-Net network;

the third step: merging the building segmentation maps; merging the binary segmentation images I by adopting the following formula_s1(i,j)、I_s2(i, j) generating a building region-merged binary image

The fourth step: removing noise and ground objects; merging binary images I using building regions_M(I, j) for two images I to be detected₁(i,j)、I₂(i, j) removing noise and feature to generate a noise-free image

The fifth step: building change detection; using unsupervised network PCANet pair I_p1(i,j)、I_p2(i, j) performing change detection; the change detection algorithm based on the PCANet comprises the following specific implementation steps:

Wherein "| · |" represents a modulo operation;

step 2: pixel pre-classification and marking; for image I_DExtracting Gabor characteristics from each pixel in (I, j), and matching with a fuzzy C mean value clustering algorithm to carry out I_D(i, j) clustering the pixels to obtain threeThe category with labels is sequentially (1) change w_c: the subscript c indicates a change, and the intra-class pixel (i, j) is labeled as l_ij1 ═ 1; (2) unchanged w_uc: the subscript uc indicates no change, and the pixel (i, j) within the class is labeled as l_ij-1; (3) to be determined w_ud: the subscript ud indicates that the pending, intra-class pixel (i, j) is labeled as l_ij0; the number of pixels included in the three categories is N_c、N_uc、N_udAnd N is_c+N_uc+N_ud＝m×n；

And step 3: calculating the number of training sample pixels; from step 2: randomly selecting training sample pixels from all m multiplied by n pixels obtained by pixel pre-classification and marking, screening positive sample pixels and negative sample pixels according to a certain proportion, wherein the label of the training sample pixels (i, j) is l_ijE { -1, +1 }; total number of training sample pixels N_sumCalculated using the following formula:

N_sum＝m×n×r (7)

where r represents the proportion of randomly chosen sample pixels; number of positive sample pixels N_posCalculated using the following formula:

negative sample pixel number N_negThen it is:

N_neg＝N_sum-N_pos (9)

and 4, step 4: generating a training set; in two input images I respectively_p1(i,j)、I_p2(i, j) selecting an image block of size k × k centered on the sample pixel (i, j)

k is an odd number in 3-11; then, two image blocks are put together

Longitudinally connected to form a sample image P_yThe size is 2k × k, i.e.:

sample image P_yLabel l of_yIs the label l of the sample pixel (i, j)_ijI.e. l_yE { -1, +1 }; co-generation of N_sumA sample image of N_posPositive sample, N_negA negative sample to form a training set

And 5: extracting PCANet characteristics and generating a characteristic vector set; extracting a training set through an unsupervised deep learning network PCANet

Firstly, setting a group of parameters to be set for training PCANet, including the number of stages for carrying out PCA filtering convolution, and taking 2; number of filters L in each stage_tTaking 5-8, wherein t is 1.

Step 6: training a classifier; using feature vector sets

Training a classifier; the classifier selects a linear kernel support vector machine; the parameter required to be set for training the Liner SVM is only one coefficient in a classifier target function, namely an error punishment factor, which is marked as C, and is taken as 1-10;

and 7: the pixels are classified again; for step 2: undetermined classification w obtained by pixel pre-classification and marking_udGenerating an image by the method in the step 4 for each pixel; then extracting features of the image by using PCANet; will obtainInputting the feature vectors into a Liner SVM which finishes training for classification; dividing pixels into variation classes w according to output results of Liner SVM_{ud_c}Or unchanged class w_{ud_uc}(ii) a Finally, synthesize w obtained in step 2_c、w_ucPixels in class, then full image pixels are classified as changes

And is not changed

In the formula, "U" represents "union set";

and a sixth step: generating a change image I_o(i,j)∈^m×n(ii) a Set of pixels derived from classification

Generating a binary change image by the following formula:

the seventh step: selecting an optimal change image; fixing the number of PCA filtering convolution stages to be 2, and adjusting the number L of filters in the two PCA filtering convolution stages₁、L₂Respectively taking 5-8, and stepping by 1; taking odd numbers in 3-11 as an image block size parameter k; taking 1-10 error penalty factors C of a Liner SVM, and stepping by 1; calculating the accuracy ACC and the omission factor P of each change image obtained in the sixth step_MDFalse alarm rate P_FA(ii) a According to ACC, P_MD、P_FASelecting relatively optimal change image from high priority to low priority

And its corresponding parameters; the calculation formula of each evaluation index is as follows:

wherein N is_TC、N_FCRespectively representing the number of pixels that are correctly detected and the number of pixels that are detected as being changed but not actually changed, N_TUC、N_FUCRespectively representing the number of unchanged pixels which are correctly detected and the number of pixels which are detected as unchanged but actually changed;

eighth step: outputting an optimal change image