CN109242889B - SAR image change detection method based on context significance detection and SAE - Google Patents

Abstract

Based on context significance detection and SAE SAR image change detection method, inputting two time phase SAR images in the same area; calculating a logarithmic ratio difference map; extracting a saliency region of the difference map based on a context saliency detection algorithm; masking the backgrounds of the SAR images in the two time phases according to the saliency areas of the difference map to obtain a masked SAR image; extracting a change area of the SAR image after the mask based on the fuzzy C-means clustering to obtain a pseudo label training sample; training a multilayer sparse automatic encoder by using the masked SAR image and the pseudo label training sample; and extracting the final SAR image change area by adopting the trained network. The SAR image change detection method based on the multi-temporal SAR image change detection can effectively overcome the influence of speckle noise on SAR image change detection, meanwhile, a multi-layer sparse automatic encoder is designed, effective change characteristics of the SAR image are extracted, the change detection accuracy can be improved, and the SAR image change detection method based on the multi-temporal SAR image change detection can be used in the technical field of multi-temporal SAR image change detection and.

Description

A SAR Image Change Detection Method Based on Context Saliency Detection and SAE

technical field

The invention belongs to the technical field of remote sensing image processing, in particular to SAR image change detection, in particular to a SAR image change detection method based on context saliency detection and SAE (Sparse Auto-Encoders, sparse automatic encoder), which can be used for multi-temporal SAR Image change detection and other technical fields.

Background technique

Synthetic aperture radar (SAR) is an active microwave imaging system, which has the advantages of all-weather, all-day and high imaging resolution, and is an important way for humans to obtain geographic information. In harsh and extreme weather conditions, compared with optical remote sensing images, SAR can obtain stable data, which has obvious advantages. SAR image change detection is the use of data in different phases to mine the change information of ground objects, and it has been applied in the fields of urban coverage research, land development and utilization.

The main steps of SAR image change detection include image preprocessing, change information extraction and change information processing. Image preprocessing includes geometric correction, image registration, radiometric correction, and image filtering. The variation information extraction is mainly generated by the difference method and the ratio method. The difference method is to perform the difference operation between the two time-phase SAR images, which is easy to implement, but is seriously affected by noise; the ratio method is to combine the two time-phase SAR images. Doing ratio operation can suppress the interference of multiplicative speckle noise. Change information processing is to divide the difference map into changing areas and non-changing areas, usually using threshold method and clustering method. In order to improve the detection effect, the classical classification algorithm combined with manual annotation is used for supervised classification, which can better overcome the influence of speckle noise on the detection results.

In recent years, the application prospect of SAR image change detection has become more and more extensive, and many scholars have achieved good research results. For example, Celik et al. published "Unsupervised change detection in satellite images using principal component analysis and k-means clustering" in IEEE Geoscience and Remote Sensing Letters, Vol. 6, Issue 4 in 2009, and proposed a method based on principal component analysis and k-means clustering. The clustering change detection method can effectively reduce data redundancy, highlight change information, and improve detection accuracy. Gong Maoguo et al. published "Neighborhood-based ratio approach for change detection in SAR images" in IEEE Geoscience and Remote Sensing Letters, Vol. 9, No. 2, in 2012, and proposed a SAR image change detection method based on the neighborhood ratio method. , which can reduce speckle noise while maintaining fine details. Gong Maoguo et al. published "Fuzzy clustering with a modified MRF energy function for change detection insynthetic aperture radar images" in IEEE Transactions on Fuzzy Systems, Vol. 22, No. 1 in 2014. Through fuzzy C-means clustering and Markov The random field energy function classifies the changed and unchanged regions, improving the accuracy of the change detection results. Zheng Yaoguo, Jiao Licheng and others published "Unsupervised saliency-guided SAR image change detection" on Pattern Recognition in 2017, and proposed a SAR image change detection based on saliency map guidance, using context saliency detection to extract salient regions of images in different phases , extract the principal components from the logarithmic ratio difference map, and obtain the final change detection result, which effectively reduces the influence of speckle noise. The above methods mainly deal with the difference map, and do not deeply mine the change characteristics of different time-phase SAR images, and there is still room for improvement in the change detection effect.

SUMMARY OF THE INVENTION

The present invention aims to overcome the above-mentioned problems of the prior art, and proposes a SAR image change detection method based on context saliency detection and SAE, so as to achieve the goal of improving the accuracy of SAR image change detection.

Technical scheme of the present invention:

SAR image change detection method based on context saliency detection and SAE, the steps are as follows:

(1) Input two time-phase SAR images of the same area:

Input two temporal SAR images of the same area for image registration, and obtain two temporal SAR images A and B after preprocessing;

(2) Calculate the log ratio difference plot:

Calculate the log ratio difference map D of the two temporal SAR images A and B;

(3) Extract the saliency region of the difference map based on the context saliency detection algorithm:

(3a) The log ratio difference map D is divided into blocks and divided into multiple scales;

(3b) In multiple scales, the similarity between pixel blocks is calculated according to the gray value and the spatial distance, and the salient value of the central pixel of the pixel block is obtained;

(3c) adding context pixels for saliency correction to obtain the final saliency map;

(4) Mask the background of the two temporal SAR images:

(4a) Select a threshold to binarize the saliency map, that is, obtain a 0-1 binary map, where 0 represents an insignificant area, and 1 represents a significant area;

(4b) masking the two time-phase SAR images A and B using the binary image, removing the unchanged part of the background, and obtaining the masked SAR images A' and B';

(5) The change area of the SAR image after the mask is extracted based on the fuzzy C-means clustering:

(5a) Calculate the neighborhood ratio difference map D' of the SAR images A' and B' after the mask;

(5b) Using fuzzy C-means clustering to divide the neighborhood ratio difference map D' into three categories, representing no change, intermediate uncertainty, and change respectively;

(5c) Select p% of samples from the unchanged class and the changed class to form a pseudo-label training sample set;

(6) Train a multi-layer SAE with post-mask SAR images and pseudo-label samples:

(6a) Construct a three-layer SAE network and randomly initialize the weight and bias parameters of the network;

(6b) Pre-training each layer of the network layer by layer using the pseudo-label training sample set to obtain the pre-trained network parameters and the implicit output of the third-layer SAE, and the implicit output is used as the optimized sample feature;

(6c) Input the optimized sample features and pseudo-labels to the softmax classifier, and train to obtain the weight and bias parameters of the softmax classifier;

(6d) Use the pseudo-label training sample set to perform reverse fine-tuning of the parameters of the softmax classifier and the three-layer SAE;

(7) Use the trained network to extract the final SAR image change detection result:

(7a) Input the sample set of the entire masked image into the three-layer SAE, and the implicit output of the third-layer SAE is used as the optimized sample feature;

(7b) Input the optimized sample features into the softmax classifier to obtain the predicted label;

(7c) Draw the final SAR image change detection result map according to the predicted label vector and spatial position.

Compared with the prior art, the present invention mainly has the following advantages:

First, the present invention uses a context-based saliency detection algorithm to obtain saliency regions, and uses fuzzy C-means clustering to obtain training samples with high confidence, which can effectively overcome the influence of speckle noise on SAR image change detection and help improve Change detection accuracy;

Second, the present invention adopts the network structure of SAE, makes use of the excellent feature expression ability of the deep network, can obtain the difference features with better discrimination between the two time-phase SAR images, and extracts the effective change features of the SAR images, thereby improving the change detection. Effect.

Description of drawings

Fig. 1 is the realization flow chart of the SAR image change detection method based on context saliency detection and SAE;

Fig. 2 is the SAR data of the San Francisco area to which the present invention is applicable, wherein Fig. 2(a) is the SAR image taken in August 2003, Fig. 2(b) is the SAR image taken in May 2004, and Fig. 2(c) is the SAR image taken in May 2004. Ground truth map of changing area;

Fig. 3 is the change detection experiment result figure of the applicable San Francisco data of the present invention, wherein Fig. 3 (a) is the change detection result figure of contrast method SG-FCM, Fig. 3 (b) is the change detection result figure of contrast method SAE, Fig. 3(c) is a graph of the change detection result of the comparison method SGK, and Fig. 3(d) is a graph of the change detection result of the method of the present invention.

Detailed ways

The present invention will be described in detail below with reference to specific examples and accompanying drawings.

According to Figure 1, the SAR image change detection method based on context saliency detection and SAE includes the following steps:

(1) Input two time-phase SAR images of the same area:

Input two temporal SAR images of the same area for image registration, and obtain two temporal SAR images A and B after preprocessing;

(2) Calculate the log ratio difference plot:

Calculate the log ratio difference map D of the two time-phase SAR images A and B, and the calculation formula is as follows:

(3) Extract the saliency region of the difference map based on the context saliency detection algorithm:

(3a) The log ratio difference map D is divided into 4 scales according to the window size of 7 × 7 and the overlap of 50%;

(3b) Calculate the similarity between pixel blocks p _t and p _j according to gray value and spatial distance at 4 scales

Among them, μ represents the position distance weight factor, where μ=3; then calculate the significant value of the central pixel block under the scale r

Among them, M represents the number of similar image blocks of the central pixel block, where M=64, and the scale r∈{100%, 80%, 50%, 30%};

(3c) Adding context pixels to correct the saliency of the center pixel block, the calculation is as follows

表示在尺度r下像素t和最近的像素t_nearest的归一化位置距离；得到每个中心像素的显著值后，生成最终的显著性图；Among them, [ ] represents the normalization operator, which can change the current scale to the original image size,

Represents the normalized position distance between pixel t and the nearest pixel t _nearest at scale r; after obtaining the saliency value of each central pixel, the final saliency map is generated;

(4) Mask the background of the two temporal SAR images:

(4a) Select a threshold to binarize the saliency map, that is, obtain a 0-1 binary map, where 0 represents an insignificant area, and 1 represents a significant area;

(4b) masking the two time-phase SAR images A and B using the binary image, removing the unchanged part of the background, and obtaining the masked SAR images A' and B';

(5) The change area of the SAR image after the mask is extracted based on the fuzzy C-means clustering:

(5a) Calculate the neighborhood ratio difference map D' of the SAR images A' and B' after the mask, the calculation formula is

Among them, the first item is the ratio difference between the pixel points, the second item is the ratio difference between the pixel point neighborhood, and Ω represents the neighborhood of the pixel point;

(5b) Using fuzzy C-means clustering to divide the neighborhood ratio difference map D' into three categories, representing no change, intermediate uncertainty, and change respectively;

(5c) Select 30% of the samples from the unchanged class and the changed class to form a pseudo-label training sample set;

(6) Train a multi-layer SAE using the post-mask SAR image and the pseudo-label sample set:

(6a) Construct a three-layer SAE network and randomly initialize the weights and bias parameters of the network

(6b) Use the pseudo-label training sample set to pre-train each layer of the network layer by layer. Let the features of the pseudo-label training samples be x _i and the labels be y _i , and the encoding process for the k-th layer of SAE is:

The SAE decoding process is:

The pre-training loss function of SAE is:

表示编码的权重和偏置，

表示解码的权重和偏置，

表示该层网络的输入，

表示该层网络的隐含输出，

表示该层网络的解码输出。f(·)和g(·)分别表示编码和解码的激活函数，KL(·)表示Kullback–Leibler散度，ρ为稀疏度参数，

表示第z个隐含单元的平均激活值，M_k表示隐含单元数目，N表示训练样本数目，λ和α表示平衡因子，||·||_F表示Forbenius范数。利用梯度下降法进行优化，得到预训练的网络参数和第三层SAE的隐含输出，隐含输出作为优化后的样本特征

Among them, the first term of the above formula is the reconstruction error of the encoding and decoding, which aims to make the decoding output of the network as close as possible to the input before encoding, the second term is the weight constraint term, which aims to prevent the weight from being too large, and the third term is the sparse restriction term, Aims to guarantee the sparsity of hidden units.

represents the encoded weights and biases,

represents the decoding weights and biases,

represents the input of this layer of network,

represents the implicit output of this layer of network,

represents the decoded output of the network at this layer. f( ) and g( ) represent the activation functions for encoding and decoding, respectively, KL( ) represents the Kullback–Leibler divergence, ρ is the sparsity parameter,

represents the average activation value of the zth hidden unit, M _k represents the number of hidden units, N represents the number of training samples, λ and α represent the balance factor, and ||·|| _F represents the Forbenius norm. The gradient descent method is used for optimization, and the pre-trained network parameters and the implicit output of the third-layer SAE are obtained, and the implicit output is used as the optimized sample feature.

(6c) Input the optimized sample features and pseudo-labels to the softmax classifier, and train to obtain the weight and bias parameters of the softmax classifier. The loss function is

表示分类器预测结果，W⁴表示softmax分类器的权重，λ表示平衡因子。利用反向传播算法求解该目标方程，获得softmax的参数；Among them, the first term of the above formula represents the prediction error, and the second term represents the weight penalty term.

represents the classifier prediction result, W4 represents the weight of the ^softmax classifier, and λ represents the balance factor. Use the back propagation algorithm to solve the objective equation and obtain the parameters of softmax;

(6d) Use the pseudo-label training sample set to perform reverse fine-tuning of the parameters of the softmax classifier and the three-layer SAE, and the realized loss function is:

表示第k层网络的隐含输出，

表示第k层网络的输入。利用反向传播算法进行优化。Among them, the first term of the above equation is the average reconstruction error of the entire network, and the second term is the weight constraint term.

represents the implicit output of the k-th layer network,

represents the input of the k-th layer network. Optimized using the backpropagation algorithm.

(7) Use the trained network to extract the final SAR image change detection result:

(7a) Input the sample set of the entire masked image into the three-layer SAE, and the implicit output of the third-layer SAE is used as the optimized sample feature

输入到softmax分类器，得到预测概率如下：(7b) The optimized sample features

Input to the softmax classifier, the predicted probability is obtained as follows:

和

对应第j类的部分权重和偏置，

和

对应第c类的部分权重和偏置。最后，样本x_i的预测标签如下：in,

and

The partial weights and biases corresponding to the jth class,

and

Partial weights and biases corresponding to class c. Finally, the predicted labels for sample _xi are as follows:

(7c) Draw the final SAR image change detection result map according to the predicted label vector and spatial position.

The technical effects of the present invention are described below through simulation experiments:

1. Simulation conditions and content

The experimental data of the present invention are San Francisco data, respectively as shown in Figure 2, the pixel size is 256 × 256, wherein the white part of Figure 2 (c) is the change area, and the black part is the non-change area; Figure 3 is the application of the present invention The experimental results of SanFrancisco data, in which Figure 3(a) is the change detection result of the comparison method SG-FCM, Figure 3(b) is the change detection result of the comparison method SAE, and Figure 3(c) is the comparison method SGK. Figure 3(d) is the result of change detection using the method of the present invention. Table 1 shows the comparison of the change detection accuracy of these techniques. In the simulation experiment, both the present invention and the comparison method are implemented by programming in Matlab R2017a.

Among the three comparison methods used in this paper, SG-FCM is based on context saliency detection method and fuzzy C-means clustering method, SGK is based on context saliency detection method, PCA and k-means method, SAE is based on neighborhood ratio and autoencoder methods.

2. Analysis of simulation results

Table 1 Accuracy comparison of data change detection in San Francisco

method number of errors Number of missed inspections total errors Detection accuracy Kappa coefficient SG-FCM 549 534 1083 0.9835 0.8757 SAE 669 300 969 0.9852 0.8925 SGK 702 383 1085 0.9834 0.8791 this method 343 439 782 0.9881 0.9093

It can be seen from the results in Table 1 that the method of the present invention obtains higher change detection accuracy than the existing SG-FCM, SAE and SGK methods, which proves that the method of the present invention effectively improves the accuracy of SAR image transformation detection. From the comparison of each method in Fig. 3, it can be seen that the results of the method of the present invention are the closest to the real change results, indicating that the context saliency detection algorithm can effectively overcome the speckle noise of SAR images, and the deep network based on SAE can obtain effective change characteristics, so as to achieve The effect of improving the accuracy of change detection.

Claims (1)

1. A SAR image change detection method based on context significance detection and SAE is characterized by comprising the following steps:

(1) two time-phase SAR images of the same area are input:

inputting two time phase SAR images of the same area for image registration to obtain two preprocessed time phase SAR images A and B;

(2) calculating a log ratio difference map:

calculating a logarithmic ratio difference graph D of the two time phase SAR images A and B, wherein the calculation formula is as follows:

(3) extracting a saliency region of the difference map based on a context saliency detection algorithm:

(3a) partitioning the logarithm ratio difference graph D and dividing the logarithm ratio difference graph D into a plurality of scales;

(3b) calculating pixel blocks p from gray values and spatial distances at multiple scales_tAnd p_jSimilarity between them

Wherein μ represents a location distance weighting factor; then calculating the significant value of the central pixel block under the scale r

Wherein M represents the number of similarity image blocks of the central pixel block;

(3c) adding context pixel to carry out significance correction of central pixel block, and calculating as follows

Wherein [ ·]Representing a normalized operator, changing the current scale into the size of the original picture,

representing the pixel t and the nearest pixel t at the scale r_nearestNormalized position distance of (a); after the significant value of each central pixel is obtained, a final significant graph is generated;

(4) the background of the two phase SAR images is masked:

(4a) selecting a threshold value to carry out binarization on the significance map, namely obtaining a 0-1 binary map, wherein 0 represents an insignificant area, and 1 represents a significant area;

(4b) masking the two time phase SAR images A and B by using a binary image, removing a background unchanged part, and obtaining masked SAR images A 'and B';

(5) extracting a change area of the SAR image after the mask based on fuzzy C-means clustering:

(5a) calculating a neighborhood ratio difference map D ' of the SAR images A ' and B ' after the mask, and calculating a formula of

The first term is the ratio difference between pixel points, the second term is the ratio difference of pixel point neighborhoods, and omega represents the neighborhoods of the pixel points;

(5b) dividing the neighborhood ratio difference graph D' into three classes by adopting fuzzy C-means clustering, and respectively representing a non-change class, a middle uncertain class and a change class;

(5c) selecting p% of samples from the unchanged class and the changed class to form a pseudo label training sample set;

(6) training the multi-layer SAE with the masked SAR image and the pseudo label sample set:

(6a) constructing three layers of SAE network, randomly initializing weight and bias parameters of network

(6b) Pre-training each layer of network layer by utilizing a pseudo label training sample set, and setting the characteristic of the pseudo label training sample as x_iThe label is y_iThe encoding procedure for the k-th layer SAE is:

the SAE decoding process is as follows:

the pre-training loss function for SAE is:

the first term of the above formula is a reconstruction error of encoding and decoding, and aims to enable the decoding output of the network to approach the input before encoding as much as possible, the second term is a weight constraint term and aims to prevent the weight from being overlarge, and the third term is a sparse limit term and aims to ensure the sparsity of an implicit unit;

the weights and biases representing the encoding are,

representing the weight and the offset of the decoding,

indicating the input to the layer network(s),

an implicit output of the layer network is indicated,

a decoded output representing the layer network; f (-) and g (-) denote activation functions of encoding and decoding, respectively, KL (-) denotes a Kullback-Leibler divergence, ρ is a sparsity parameter,

to represent the z-th hiddenContaining the mean activation value of the cell, M_kRepresenting the number of hidden elements, N representing the number of training samples, λ and α representing the balance factors, | · | | survival_FRepresents the Forbenius norm; optimizing by using a gradient descent method to obtain pre-trained network parameters and implicit output of the third layer SAE, wherein the implicit output is used as optimized sample characteristics

(6c) Inputting the optimized sample characteristics and the pseudo labels into a softmax classifier, training to obtain the weight and the bias parameters of the softmax classifier, and obtaining a loss function of

Wherein, the first term of the above formula represents the prediction error, and the second term represents the weight penalty term;

represents the classifier prediction, W⁴Representing the weight of the softmax classifier, and λ representing a balance factor; solving the target equation by using a back propagation algorithm to obtain a parameter of softmax;

(6d) and performing reverse fine adjustment on parameters of the softmax classifier and the three layers of SAEs by using a pseudo label training sample set, wherein the realized loss function is as follows:

wherein, the first term of the above equation is the average reconstruction error of the whole network, and the second term is the weight constraint term;

indicating an implicit output of the k-th network,

representing networks in the k-th networkInputting; optimizing by using a back propagation algorithm;

(7) extracting a final SAR image change detection result by adopting a trained network:

(7a) the sample set of the entire masked image is input into the three-layer SAE, the implicit output of the third layer SAE being the optimized sample feature

(7b) Sample characteristics after optimization

Inputting the prediction probability into a softmax classifier, and obtaining the following prediction probability:

wherein,

and

the partial weights and offsets corresponding to class j,

and

partial weights and offsets corresponding to class c; finally, sample x_iThe predictive signatures of (c) are as follows:

(7c) and drawing a final SAR image change detection result graph according to the predicted label vector and the spatial position.

Images