CN110096948B - Remote sensing image identification method based on characteristic aggregation convolutional network - Google Patents

Abstract

The invention relates to a remote sensing image recognition method based on a feature aggregation convolutional network, comprising the following steps: 1) extracting features by using a VGG-16 convolutional neural network; 2) coding convolution features; 3) remote sensing scene expression; 4) training Feature aggregation convolutional network; 5) Predict the scene category of remote sensing images. The invention establishes a remote sensing scene recognition method based on a feature aggregation convolution network, directly learns the remote sensing scene expression from the remote sensing scene database, improves the remote sensing scene recognition accuracy, and can be used in forest fire monitoring, urban planning and other fields.

Description

Remote sensing image recognition method based on feature aggregation convolutional network

technical field

The invention relates to the technical field of remote sensing image processing, in particular to a remote sensing image recognition method.

Background technique

Remote sensing images are a type of earth observation image data captured by remote sensing imaging systems, which contain detailed information on the composition and distribution of ground objects. However, due to the complex composition of remote sensing scenes, the small objects of objects, and the cluttered spatial distribution of objects, the differences between the images of the same type of remote sensing scenes are large, and the differences between the images of different types of remote sensing scenes are small. Therefore, it is still a very challenging task to accurately infer the remote sensing scene category according to the remote sensing scene ground object content.

In recent years, scholars have proposed many remote sensing scene recognition methods. According to the feature operators used in the methods, the existing methods can be divided into two categories: methods based on handcrafted features and methods based on convolutional neural networks. Handcrafted feature-based methods usually employ handcrafted feature descriptors to construct remote sensing scene representations. Such as Wang et al. in the document "Y.Wang et al., "Learning a Discriminative Distance Metric With Label Consistency for Scene Classification," IEEE Transactions on Geoscience and Remote Sensing, vol.55, no.8, pp.4427-4440, 2017. ” uses classical scale-invariant features and sparse coding to extract remote sensing scene representations, and then uses a label-constrained distance metric function to classify remote sensing scenes. However, artificially designed feature operators are often strongly dependent on image prior knowledge and cannot extract high-level semantic information of remote sensing scenes. Methods based on convolutional neural networks often use deep convolutional neural networks to extract high-level semantic information of remote sensing scenes. For example, He et al. in the literature "N. He, L. Fang, S. Li, A. Plaza and J. Plaza, "Remote Sensing Scene Classification Using Multilayer Stacked Covariance Pooling," IEEE Transactions on Geoscience and Remote Sensing, pp.1–12 , 2018." in using convolutional neural network and multi-layer stacked covariance downsampling to construct remote sensing scene representation containing high-level semantic information, and then using support vector machine for remote sensing scene recognition. However, the current methods based on convolutional neural networks still use traditional unsupervised feature aggregation methods to construct remote sensing scene representations after using convolutional neural networks to extract high-level semantic information of remote sensing scenes. process.

To sum up, the existing remote sensing scene recognition methods rely on traditional manual features or traditional feature aggregation methods, and it is difficult to obtain satisfactory recognition accuracy.

SUMMARY OF THE INVENTION

The purpose of the invention is to solve the problem that the original remote sensing scene database training set has too little data and it is difficult to train the convolutional neural network, and proposes a remote sensing image recognition method based on the feature aggregation convolutional network to improve the expression of the convolutional neural network for remote sensing scenes. ability.

The technical scheme of the present invention is as follows:

A remote sensing image recognition method based on a feature aggregation convolutional network, comprising the following steps:

First, the training stage:

1) Extract features using VGG-16 convolutional neural network;

The VGG-16 convolutional neural network framework is built by using convolutional layer, downsampling layer, fully connected layer and activation function layer, and the convolutional features of remote sensing images are extracted by different convolutional layers of VGG-16 respectively. The convolution features of the shallow convolution layer can better extract the spatial information of remote sensing images, while the deeper convolution features are better at extracting the semantic information of remote sensing images;

2) Convolutional feature encoding;

Construct a deep convolutional feature encoding module, which can be embedded in the VGG-16 convolutional neural network, fuse the spatial information and semantic information of different convolutional features, and encode the convolutional features of different convolutional layers as convolutional features. Express;

3) Remote sensing scene expression;

The convolutional representation is fused with the fully connected layer features of the VGG-16 convolutional neural network to obtain a discriminative remote sensing scene representation;

4) Training feature aggregation convolutional network;

The VGG-16 convolutional neural network, convolutional feature encoding module and remote sensing scene expression module are integrated into the same convolutional neural network structure to form an end-to-end feature aggregation convolutional network. End-to-end joint training of feature aggregation convolutional networks using stochastic gradient descent.

2. Types of Measured Scenarios

Input the remote sensing image to be tested into the trained feature aggregation convolution network, and the feature aggregation convolution network operation outputs the category of the remote sensing image.

The above step 1) can be specifically:

The VGG-16 convolutional neural network is built using the convolutional layer, the downsampling layer, the activation function layer, etc., and the 256×256-sized remote sensing scene image is input into the VGG-16 convolutional neural network, through the following series of nonlinear volumes Product operation to extract multi-layer convolution features of remote sensing scene images:

x ⁱ =σ( ^wi *x ^i-1 +b ⁱ )

Among them, x ⁱ represents the convolution feature of the remote sensing scene image in the i convolution layer, i=1, 2, 3,..., x ⁰ represents the initial 256×256 size remote sensing scene image, and w ⁱ represents the i-th convolution layer. Weight, b ⁱ represents the bias of the i-th convolutional layer, "*" represents the convolution operation, and σ(x)=max(0,x) represents the nonlinear activation function.

The above step 2) can be specifically:

2.1) Construct a convolution feature scale unified module, and use the 3rd, 4th, and 5th downsampling layers of VGG-16 to extract the convolution features of remote sensing scene images and denote them as x ¹ , x ² , and x ³ respectively. On the one hand, since the downsampling layer will reduce the resolution of convolutional features, and on the other hand, because the number of convolution kernels in different convolutional layers is inconsistent, the dimensions of x ¹ , x ² and x ³ are 28×28×256, 14 respectively. ×14×512 and 7×7×512. Then, 4×4 downsampling with stride 4 and 2×2 downsampling with stride 2 are applied to x ¹ and x ² respectively, and the length and width of all convolutional features are unified to 7×7.

2.2) Construct an in-channel convolution feature normalization module, and use L2-normalization to normalize the unified scale x ¹ , x ² and x ³ in the channel dimension to [0,1] respectively. The calculation formula is as follows:

为卷积特征xⁱ在(h,w,c)处的值，

和

的尺寸相同，ε＝e^-8用于避免除数为0。in,

is the value of the convolution feature x ⁱ at (h, w, c),

and

of the same size, ε=e- ⁸ is used to avoid division by 0.

和

编码为卷积表达。首先，为避免引入需要训练的参数，采用特征串联的方式直接在通道维将

和

聚合到一起，聚合后的特征尺度为7×7×1280。接着，利用卷积核尺寸为1×1卷积层和ReLU非线性激活函数，在语义标签信息监督下对聚合后的特征进行非线性编码，得到融合了不同卷积层空间信息和语义信息的卷积表达。2.3) Construct a convolutional feature encoding module to normalize the convolutional features

and

Encoded as a convolutional representation. First of all, in order to avoid introducing parameters that need to be trained, the feature concatenation method is used to directly combine the channel dimension.

and

Aggregated together, the aggregated feature scale is 7×7×1280. Then, using the convolution kernel size of 1×1 convolution layer and the ReLU nonlinear activation function, the aggregated features are nonlinearly encoded under the supervision of semantic label information, and a fusion of spatial information and semantic information of different convolution layers is obtained. Convolutional expression.

The above step 3) is specifically:

The remote sensing scene expression module is constructed. The hierarchical structure of the convolutional neural network determines that its different layer features contain different spatial information and semantic information. The convolutional expression obtained by the convolutional feature encoding module integrates the spatial information and semantic information of different convolutional features. . However, convolutional features are local features of remote sensing scenes, and the convolutional representation obtained through the convolutional feature encoding module still cannot capture the global information of remote sensing scenes. The features of the fully connected layer of the VGG-16 convolutional neural network contain the global information of the remote sensing scene. The present invention proposes a progressive feature aggregation strategy, which gradually aggregates features according to the spatial structure information of the features. First, the convolutional representation is obtained by aggregating the convolutional features through the convolutional feature encoding module. Then, the convolutional representation is fused with the fully connected layer features of VGG-16, and finally a discriminative remote sensing scene representation is obtained.

The above step 4) is specifically:

An end-to-end feature aggregation convolutional network is constructed, and the VGG-16 convolutional neural network, convolutional feature encoding module and remote sensing scene expression module are integrated into one network to obtain a feature aggregation convolutional network. The proposed feature aggregation convolutional network integrates feature learning, feature aggregation and classifier into one network, which can be trained end-to-end guided by semantic label information. Next, on the training dataset, the feature aggregation convolutional network is jointly trained end-to-end using stochastic gradient descent. The objective function of the feature aggregation convolutional network is as follows:

表示M张遥感场景图像，K表示场景类别数目，1{·}表示示性函数。由于现有的遥感场景识别数据库中训练数据较少，不足以从头开始训练卷积神经网络，本发明采用微调的方法缓解训练数据不足的问题。首先使用在目标识别数据库ImageNet中预训练的权重来初始化特征聚合卷积网络，然后利用随机梯度下降法更新特征聚合卷积网络的各层权重参数，直至最大迭代次数。where x and y represent the scene representation and semantic label of the remote sensing scene image I, respectively,

represents M remote sensing scene images, K represents the number of scene categories, and 1{·} represents the indicative function. Since there are few training data in the existing remote sensing scene recognition database, it is not enough to train the convolutional neural network from scratch, the present invention adopts the fine-tuning method to alleviate the problem of insufficient training data. First, the weights pre-trained in the target recognition database ImageNet are used to initialize the feature aggregation convolutional network, and then the stochastic gradient descent method is used to update the weight parameters of each layer of the feature aggregation convolutional network until the maximum number of iterations.

When predicting the category of each remote sensing scene to be tested, the remote sensing scene image of size 256×256 is directly input into the trained feature aggregation convolutional network, and the category of the remote sensing scene can be obtained.

Compared with the prior art, the present invention has the following technical effects:

1. The present invention proposes a supervised feature encoding module that can encode features from different convolutional layers of a convolutional neural network under the supervision of label information. In addition, in order to explore the complementarity of different features, the present invention proposes a progressive aggregation strategy, which gradually aggregates features according to their spatial structure information.

2. The present invention proposes an end-to-end feature aggregation convolutional network that directly learns remote sensing scene representations from a remote sensing scene database. The proposed method integrates feature learning, feature aggregation and classifiers into a unified framework for joint training. .

3. The present invention directly learns the remote sensing scene expression from the remote sensing scene database, improves the recognition accuracy of the remote sensing scene, and can be widely used in forest fire monitoring, urban planning and other fields.

Description of drawings

FIG. 1 is a flowchart of a remote sensing scene recognition method based on a feature aggregation convolutional network according to the present invention.

Detailed ways

As shown in Figure 1, the remote sensing scene recognition method based on the feature aggregation convolution network provided by the present invention mainly includes the following steps:

1) Extract features using VGG-16 convolutional neural network;

The VGG-16 convolutional neural network framework is constructed by using convolutional layer, downsampling layer, fully connected layer, and activation function layer, and the convolutional features of remote sensing images are extracted by different convolutional layers of VGG-16;

2) Convolutional feature encoding;

Construct a deep convolutional feature encoding module, which can be embedded in the VGG-16 convolutional neural network, fuse the spatial information and semantic information of different convolutional features, and encode the convolutional features of different convolutional layers as convolutional features. Express;

3) Remote sensing scene expression;

The convolutional representation is fused with the fully connected layer features of the VGG-16 convolutional neural network to obtain a discriminative remote sensing scene representation;

4) Training feature aggregation convolutional network;

The VGG-16 convolutional neural network, convolutional feature encoding module and remote sensing scene expression module are integrated into the same convolutional neural network structure to form an end-to-end feature aggregation convolutional network. End-to-end joint training of feature aggregation convolutional networks using stochastic gradient descent.

5) Predict the scene category of remote sensing images;

The remote sensing image to be tested is input into the feature aggregation convolutional network, and the category of the remote sensing scene is directly obtained.

The steps implemented by the present invention will be further described in detail below with reference to FIG. 1 .

Step 1, using VGG-16 convolutional neural network to extract features.

The VGG-16 convolutional neural network is built using the convolutional layer, the downsampling layer, the activation function layer, etc., and the 256×256-sized remote sensing scene image is input into the VGG-16 convolutional neural network, through the following series of nonlinear volumes Product operation to extract multi-layer convolution features of remote sensing scene images:

x ⁱ =σ( ^wi *x ^i-1 +b ⁱ )

Among them, x ⁱ represents the convolution feature of the remote sensing scene image in the i convolution layer, i=1, 2, 3,..., x ⁰ represents the initial 256×256 size remote sensing scene image, and w ⁱ represents the i-th convolution layer. Weight, b ⁱ represents the bias of the i-th convolutional layer, "*" represents the convolution operation, and σ(x)=max(0,x) represents the nonlinear activation function.

Step 2, convolutional feature encoding.

2.1) Construct a convolution feature scale unified module, and use the 3rd, 4th, and 5th downsampling layers of VGG-16 to extract the convolution features of remote sensing scene images and denote them as x ¹ , x ² , and x ³ respectively. On the one hand, because the downsampling layer will reduce the resolution of convolutional features, and on the other hand, because the number of convolution kernels in different convolutional layers is inconsistent, the dimensions of x ¹ , x ² and x ³ are 28×28×256, 14 respectively. ×14×512 and 7×7×512. Then, 4×4 downsampling with stride 4 and 2×2 downsampling with stride 2 are applied to x ¹ and x ² respectively, and the length and width of all convolutional features are unified to 7×7.

2.2) Construct an in-channel convolution feature normalization module, and use L2-normalization to normalize the unified scale x ¹ , x ² and x ³ in the channel dimension to [0,1] respectively. The calculation formula is as follows:

为卷积特征xⁱ在(h,w,c)处的值，

和

的尺寸相同，ε＝e^-8用于避免除数为0。in,

is the value of the convolution feature x ⁱ at (h, w, c),

and

of the same size, ε=e- ⁸ is used to avoid division by 0.

和

编码为卷积表达。首先，为避免引入需要训练的参数，采用特征串联的方式直接在通道维将

和

聚合到一起，聚合后的特征尺度为7×7×1280。接着，利用卷积核尺寸为1×1卷积层和ReLU非线性激活函数，在语义标签信息监督下对聚合后的特征进行非线性编码，得到融合了不同卷积层空间信息和语义信息的卷积表达。2.3) Construct a convolutional feature encoding module to normalize the convolutional features

and

Encoded as a convolutional representation. First of all, in order to avoid introducing parameters that need to be trained, the feature concatenation method is used to directly combine the channel dimension.

and

Aggregated together, the aggregated feature scale is 7×7×1280. Then, using the convolution kernel size of 1×1 convolution layer and the ReLU nonlinear activation function, the aggregated features are nonlinearly encoded under the supervision of semantic label information, and a fusion of spatial information and semantic information of different convolution layers is obtained. Convolutional expression.

Step 3, remote sensing scene expression.

The remote sensing scene expression module is constructed. The hierarchical structure of the convolutional neural network determines that its different layer features contain different spatial information and semantic information. The convolutional expression obtained by the convolutional feature encoding module integrates the spatial information and semantic information of different convolutional features. . However, convolutional features are local features of remote sensing scenes, and the convolutional representation obtained through the convolutional feature encoding module still cannot capture the global information of remote sensing scenes. The features of the fully connected layer of the VGG-16 convolutional neural network contain the global information of the remote sensing scene. The present invention proposes a progressive feature aggregation strategy, which gradually aggregates features according to the spatial structure information of the features. First, the convolutional representation is obtained by aggregating the convolutional features through the convolutional feature encoding module. Then, the convolutional representation is fused with the fully connected layer features of VGG-16, and finally a discriminative remote sensing scene representation is obtained.

Step 4, train the feature aggregation convolutional network.

An end-to-end feature aggregation convolutional network is constructed, and the VGG-16 convolutional neural network, convolutional feature encoding module and remote sensing scene expression module are integrated into one network to obtain a feature aggregation convolutional network. The proposed feature aggregation convolutional network integrates feature learning, feature aggregation and classifier into one network, which can be trained end-to-end guided by semantic label information. Next, on the training dataset, the feature aggregation convolutional network is jointly trained end-to-end using stochastic gradient descent. The objective function of the feature aggregation convolutional network is as follows:

表示M张遥感场景图像，K表示场景类别数目，1{·}表示示性函数。由于现有的遥感场景识别数据库中训练数据较少，不足以从头开始训练卷积神经网络，本发明采用微调的方法缓解训练数据不足的问题。首先使用在目标识别数据库ImageNet中预训练的权重来初始化特征聚合卷积网络，然后利用随机梯度下降法更新特征聚合卷积网络的各层权重参数，直至最大迭代次数。where x and y represent the scene representation and semantic label of the remote sensing scene image I, respectively,

represents M remote sensing scene images, K represents the number of scene categories, and 1{·} represents the indicative function. Since there are few training data in the existing remote sensing scene recognition database, it is not enough to train the convolutional neural network from scratch, the present invention adopts the fine-tuning method to alleviate the problem of insufficient training data. First, the weights pre-trained in the target recognition database ImageNet are used to initialize the feature aggregation convolutional network, and then the stochastic gradient descent method is used to update the weight parameters of each layer of the feature aggregation convolutional network until the maximum number of iterations.

Step 5, predict the scene category of the remote sensing image.

Predict the category of each remote sensing scene to be tested, and directly input the 256×256 size remote sensing scene image into the trained feature aggregation convolution network to obtain the category of the remote sensing scene. Statistical remote sensing scene recognition accuracy.

The effect of the present invention can be further illustrated by the following experiments.

1. Simulation conditions

The present invention uses MATLAB software for simulation on the central processing unit of Intel(R) Core i7-5930K 3.50GHZ, GeForce GTX TitanX GPU, memory 64G and Linux operating system.

2. Simulation content

To verify the performance of the proposed feature aggregation convolutional network, experiments are conducted on two public remote sensing scene recognition databases: AID database and UC-Merced database. The AID database consists of 30 remote sensing scene categories, each with 200 to 400 images. The image size on the AID database is 600×600. 50% images of each scene category are randomly selected as the training set, and the rest are used as the test set. The UC-Merced database contains 21 remote sensing scene categories. There are 100 images per scene category with dimensions of 256×256. 80 images of each scene category are randomly selected as the training set, and the remaining images are used as the test set.

Taking into account the novelty and popularity of the algorithm, we selected 4 contrasting methods in the two databases for comparison to verify the recognition performance of the proposed method. The comparison algorithms used are from the following documents: 1. "G.-S.Xia, J.Hu, F.Hu, B.Shi, X.Bai, Y.Zhong, L.Zhang, and X.Lu, " Aid: A Benchmark DataSet for Performance Evaluation of Aerial Scene Classification,"IEEETransactions on Geoscience and Remote Sensing,vol.55,no.7,pp.3965–3981,2017.", 2."S.Chaib,H.Liu, Y.Gu, and H.Yao, “Deep Feature Fusion for Vhr RemoteSensing Scene Classification,” IEEE Transactions on Geoscience and RemoteSensing, vol.55, no.8, pp.4775–4784, 2017.” 3. “N.He , L. Fang, S. Li, A. Plaza, and J. Plaza, "Remote Sensing Scene Classification Using Multilayer StackedCovariance Pooling," IEEE Transactions on Geoscience and Remote Sensing, pp.1–12, 2018. "4." Y. Yu and F. Liu, "Aerial Scene Classification via Multilevel Fusionbased on Deep Convolutional Neural Networks," IEEE Geoscience and RemoteSensing Letters, vol.15, no.2, pp.287–291, 2018." and 5. "K.Qi , C. Yang, Q. Guan, H. Wu, and J. Gong, "A Multiscale Deeply Described Correlatons-based Model for Land-useScene Classification," Remote Sensing, vol.9, no.9, pp.917–933, Sep. 2017.”

The accuracy of remote sensing scene recognition is used to quantify the performance of remote sensing scene recognition methods. The experiments were repeated 5 times on the AID database and the UC-Merced database, respectively, and the average recognition accuracy category was calculated, and the results are shown in Table 1 and Table 2.

Table 1 Accuracy of different remote sensing scene recognition methods in AID database

Table 2 Accuracy of different remote sensing scene recognition methods in UC-Merced database

It can be seen from Table 1 and Table 2 that the remote sensing scene recognition accuracy of the present invention is the highest. This is because the invention makes full use of the label information to supervise the feature learning and feature aggregation process, and the remote sensing scene representation proposed by the feature aggregation convolutional network can better extract the feature distribution features of the remote sensing scene. Therefore, the method is more effective than other methods, and the experimental results further verify the effectiveness of the present invention.

Claims (5)

1. a remote sensing image recognition method based on feature aggregation convolutional network, is characterized in that, comprises: 1) Training phase 1.1) Extract features using VGG-16 convolutional neural network: The VGG-16 convolutional neural network framework is constructed by using the convolutional layer, the downsampling layer, the fully connected layer, and the activation function layer, and the convolutional features of the remote sensing images in the training set are extracted by different convolutional layers of the VGG-16; Among them, the convolutional features of the shallower convolutional layers in VGG-16 are used to extract the spatial information of remote sensing images, and the convolutional features of the deeper convolutional layers are used to extract the semantic label information of remote sensing images; 1.2) Convolutional feature encoding: Construct a convolutional feature encoding module so that the module can be embedded in the VGG-16 convolutional neural network formed in step 1.1), fuse the spatial information and semantic label information of different convolutional features, and combine the volumes of different convolutional layers. Product features are encoded as convolutional representations; 1.3) Remote sensing scene expression; A remote sensing scene expression module is constructed to fuse the convolutional expression obtained in step 1.2) with the fully connected layer features of the VGG-16 convolutional neural network formed in step 1.1) to obtain a discriminative remote sensing scene expression; 1.4) Training feature aggregation convolutional network; The VGG-16 convolutional neural network, convolutional feature encoding module and remote sensing scene expression module are integrated into the same convolutional neural network structure to obtain a feature aggregation convolutional network; the feature aggregation convolutional network integrates feature learning, feature aggregation and A classifier, which can be trained end-to-end under the guidance of semantic label information; in the training dataset, the feature aggregation convolutional network is jointly trained end-to-end using stochastic gradient descent; 2) Measured scene category Input the remote sensing image to be tested into the trained feature aggregation convolution network, and the feature aggregation convolution network calculates and outputs the category of the remote sensing image; specifically refer to the training process to perform the following steps: 2.1) Extract the convolution features of the remote sensing images to be tested, and obtain the spatial information and semantic label information of different convolution features; 2.2) Integrate the spatial information and semantic label information of different convolutional features, and encode the convolutional features of different convolutional layers into convolutional expressions; 2.3) The convolutional representation is fused with the features of the fully connected layer to obtain a discriminative remote sensing scene representation; and then the classification of the remote sensing image to be tested is obtained through the classifier. 2. the remote sensing image recognition method based on feature aggregation convolutional network according to claim 1, is characterized in that, step 1.1) specifically carries out non-linear convolution operation according to following formula, extracts the multi-layer convolution feature of remote sensing scene image: x ⁱ =σ( ^wi *x ^i-1 +b ⁱ ) Among them, x ⁱ represents the convolution feature of the remote sensing scene image in the i convolution layer, i=1, 2, 3,..., x ⁰ represents the initial 256×256 size remote sensing scene image, and w ⁱ represents the i-th convolution layer. Weight, b ⁱ represents the bias of the i-th convolutional layer, "*" represents the convolution operation, and σ(x)=max(0,x) represents the nonlinear activation function. 3. the remote sensing image recognition method based on feature aggregation convolutional network according to claim 2, is characterized in that, step 1.2) is specially: 1.2.1) Construct a unified module of convolution feature scale: use the 3rd, 4th, and 5th downsampling layers of VGG-16 to extract the convolution features of remote sensing scene images and denote them as x ¹ , x ² , and x ³ ; where x ¹ , x ² , and x ³ have dimensions 28×28×256, 14×14×512, and 7×7×512, respectively; then, downsampling 4×4 with stride 4 and 2 with stride 2, respectively ×2 downsampling acts on x ¹ and x ² , and the length and width of all convolution features are unified to 7 × 7; 1.2.2) Construct an intra-channel convolution feature normalization module: use L2-normalization to normalize the unified scale x ¹ , x ² and x ³ in the channel dimension to [0,1], the calculation formula is as follows Show:

in,

is the value of the convolution feature x ⁱ at (h, w, c),

and

are the same size, ε=e ^-8 is used to avoid the divisor being 0; 1.2.3) Constructing the convolutional feature encoding module: First, the normalized convolutional features are directly in the channel dimension by feature concatenation

and

Aggregated together, the aggregated feature scale is 7×7×1280; then, using the convolution kernel size of 1×1 convolutional layer and ReLU nonlinear activation function, the aggregated features are processed under the guidance of semantic label information. Linear encoding, resulting in a convolutional representation that fuses spatial and semantic information from different convolutional layers. 4. the remote sensing image recognition method based on feature aggregation convolutional network according to claim 3, is characterized in that, the objective function of feature aggregation convolutional network described in step 1.4) is as follows:

where x and y represent the scene representation and semantic label of the remote sensing scene image I, respectively,

represents M remote sensing scene images, K represents the number of scene categories, and 1{·} represents the indicative function. 5. the remote sensing image recognition method based on feature aggregation convolutional network according to claim 4, is characterized in that: in training stage, at first use the weight pretrained in target recognition database ImageNet to initialize feature aggregation convolutional network, then utilize The stochastic gradient descent method updates the weight parameters of each layer of the feature aggregation convolutional network until the maximum number of iterations.

Images