CN111046756A - Convolutional neural network detection method for high-resolution remote sensing image target scale features - Google Patents

Convolutional neural network detection method for high-resolution remote sensing image target scale features Download PDF

Info

Publication number
CN111046756A
CN111046756A CN201911179838.0A CN201911179838A CN111046756A CN 111046756 A CN111046756 A CN 111046756A CN 201911179838 A CN201911179838 A CN 201911179838A CN 111046756 A CN111046756 A CN 111046756A
Authority
CN
China
Prior art keywords
target
remote sensing
resolution remote
scale
sensing image
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN201911179838.0A
Other languages
Chinese (zh)
Inventor
王密
董志鹏
杨芳
刘思远
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Wuhan University WHU
Original Assignee
Wuhan University WHU
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Wuhan University WHU filed Critical Wuhan University WHU
Priority to CN201911179838.0A priority Critical patent/CN111046756A/en
Publication of CN111046756A publication Critical patent/CN111046756A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00Scenes; Scene-specific elements
    • G06V20/10Terrestrial scenes
    • G06V20/13Satellite images
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/045Combinations of networks
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods
    • G06N3/084Backpropagation, e.g. using gradient descent

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • General Engineering & Computer Science (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Astronomy & Astrophysics (AREA)
  • Remote Sensing (AREA)
  • Multimedia (AREA)
  • Image Analysis (AREA)

Abstract

本发明提供了一种高分辨率遥感影像目标尺度特征的卷积神经网络检测方法,包括通过建立包含不同辐射亮度、不同尺度大小的大规模高分辨率遥感影像多目标检测数据集,对数据集中各目标尺度范围分别统计分析,及所有目标一起统计分析,获得高分辨率遥感影像中各目标尺度的最优分布范围;根据高分辨率遥感影像中各目标尺度分布范围覆盖的区域面积,获得高分辨率遥感影像目标检测建议框的最优尺度大小;根据目标建议框的最优尺度大小,设置适合高分辨率遥感影像目标检测的卷积神经网络架构,实现影像中目标检测。本发明方法可以获得高精度的高分辨率遥感影像目标检测结果,该方法具有简单、可靠、精度高、易于实现的特点。

Figure 201911179838

The present invention provides a convolutional neural network detection method for target scale features of high-resolution remote sensing images, comprising: establishing a large-scale high-resolution remote sensing image multi-target detection data set including different radiances and different scales; Statistical analysis of each target scale range separately, and statistical analysis of all targets together, to obtain the optimal distribution range of each target scale in the high-resolution remote sensing image; The optimal scale size of the high-resolution remote sensing image target detection proposal frame; according to the optimal scale size of the target proposal frame, a convolutional neural network architecture suitable for high-resolution remote sensing image target detection is set to achieve target detection in images. The method of the invention can obtain high-precision high-resolution remote sensing image target detection results, and the method has the characteristics of simplicity, reliability, high precision and easy implementation.

Figure 201911179838

Description

Convolutional neural network detection method for high-resolution remote sensing image target scale features
Technical Field
The invention belongs to the field of remote sensing image processing and information extraction, and particularly relates to a target detection method for realizing an optimal scale suggestion box for a high-resolution remote sensing image target.
Background
With the development of ground-to-ground observation technology, the data acquisition amount of high-resolution remote sensing images is increasing, and the high-resolution remote sensing images are widely used in the aspects of urban planning, disaster monitoring, agricultural management, military reconnaissance and the like. Under the condition of big data, how to automatically and intelligently realize the target detection of the high-resolution remote sensing image has important influence on the exertion of the application value of the high-resolution remote sensing image. For this reason, a lot of research has been carried out by scholars at home and abroad, many of the research methods mainly use artificially designed image target features for target detection, such as features of histogram of gradient (HOG), Local Binary Patterns (LBP), scale-invariant feature transform (SIFT), Gabor, and the like, and then input the features into a conventional classifier, such as a Support Vector Machine (SVM), AdaBoost, a decision tree, and the like, in the form of feature quantities, so as to perform classification, thereby obtaining a better effect in a specific target detection task. However, due to the complex and changeable shooting conditions of the remote sensing satellite, the traditional target detection algorithm is difficult to adapt to remote sensing images under different conditions, and the robustness and universality of the algorithm are poor.
In recent years, a Convolutional Neural Network (CNN) is used as the hottest deep learning model algorithm, and since the target features do not need to be artificially designed, the effective feature extraction and learning can be automatically performed according to mass data and labels; in addition, under the condition that the training data is sufficient, the model has good generalization capability, and can still keep good robustness under the complicated and changeable conditions. Therefore, the convolutional neural network model has been widely applied to the field of image target detection. Currently, the conventional convolutional neural network target detection architectures include fast-regional CNN (fast-RCNN), Young Only Look One (YOLO), Single-shot multi-box Detector (SSD), and the like, and the convolutional neural network target detection architectures are all designed for the target scale of a natural image, and all achieve a better target detection result in the target detection of the natural image. The high-resolution remote sensing satellite generally images the earth surface in a near-earth orbit, and is influenced by illumination, meteorological conditions and the like in the imaging process, and the generated remote sensing images have the characteristics of complex image content, small target scale range, large radiation difference of the remote sensing images generated in different time periods and the like. Compared with a natural image, the high-resolution remote sensing image has the characteristics of more complex background, smaller target area range, larger scale change of similar targets and the like. Therefore, the existing target detection frameworks such as fast-RCNN, YOLO and SSD can not effectively couple the target scale characteristics of the high-resolution remote sensing image, and the high-precision high-resolution remote sensing image target detection result is difficult to obtain. A research team of the university of Wuhan submits a paper 'scale feature convolutional neural network identification method of a remote sensing image target' in the future, and a convolutional neural network detection and identification method based on target scale features is provided aiming at the problem that the robustness and universality of artificially designed features in the traditional image target detection and identification are poor. However, in the paper, statistical analysis is only roughly performed on all target scales in the data set, and a reasonable optimal suggested box scale acquisition technical means is lacked, so that the detection effect is influenced.
Aiming at the problems, the invention provides improvement and provides a novel convolutional neural network detection method for the target scale characteristics of the high-resolution remote sensing image.
Disclosure of Invention
The invention provides a novel convolutional neural network detection method based on remote sensing image target scale characteristics, aiming at the problem of how to obtain a high-resolution remote sensing image target detection optimal scale suggestion frame.
The technical scheme provided by the invention is a convolutional neural network detection method for high-resolution remote sensing image target scale characteristics, which comprises the following steps:
step 1, establishing a large-scale high-resolution remote sensing image multi-target detection data set containing different radiances and different scales, respectively carrying out statistical analysis on the scale range of each target in the data set, and carrying out statistical analysis on all targets together to obtain the optimal distribution range of each target scale in the high-resolution remote sensing image;
step 2, obtaining the optimal dimension of the target detection suggestion frame of the high-resolution remote sensing image according to the area covered by the target dimension distribution range in the high-resolution remote sensing image;
and 3, setting a convolutional neural network architecture suitable for high-resolution remote sensing image target detection according to the optimal size of the target suggestion frame, and realizing target detection in the image.
Furthermore, in step 3, the convolutional neural network architecture comprises a target region suggestion network RPN and a target classification and accurate positioning network CNN,
the target area suggestion network RPN generates various target candidate areas at each position of the characteristic diagram, and transmits the information of the target candidate areas to the CNN;
the target classification and accurate positioning network CNN uses five layers of convolution layers to extract image target characteristic graphs, and combines target candidate region information and the last layer of characteristic graph to obtain characteristic vectors of target candidate regions; then, the feature vector of the target candidate region is transmitted to the region of interest pooling layer to obtain the feature vector of the target candidate region with the specified size; and finally, the feature vectors with the specified sizes are transmitted to a full connection layer for training and testing of target recognition classification and regional coordinate regression.
Moreover, the size of the region-of-interest pooling layer is 7 × 7, and both of the two fully-connected layers behind the region-of-interest pooling layer contain 4096 neurons; then a full connection layer corresponding to the target classification layer comprises n neurons, a target coordinate regression layer comprises 4n neurons, n is the classification number of the target in the image, and 4n represents the coordinate coefficients of n types of targets x, y, w and h corresponding to the target area; wherein x is the coordinate of the horizontal axis of the central point, y is the coordinate of the vertical axis of the central point, w is the width of the target, and h is the height of the target.
The invention provides a convolutional neural network detection method for the scale characteristics of a high-resolution remote sensing image target, which can be used for well coupling the scale characteristics of the high-resolution remote sensing image target and obtaining a high-precision remote sensing image target detection result.
Drawings
Fig. 1 is a graph of a statistical result of target dimensions of a high-resolution remote sensing image according to an embodiment of the present invention, where fig. 1(a) is a distribution range of target dimensions of an airplane in the high-resolution remote sensing image, fig. 1(b) is a distribution range of target dimensions of a storage tank in the high-resolution remote sensing image, fig. 1(c) is a distribution range of target dimensions of a ship in the high-resolution remote sensing image, and fig. 1(d) is a distribution range of dimensions of all targets (the airplane, the storage tank, and the ship) in the high-resolution remote.
Fig. 2 is a diagram of an optimal dimension of a high-resolution remote sensing image target detection suggestion box according to an embodiment of the invention.
Fig. 3 is a convolutional neural network architecture for detecting a target in a high-resolution remote sensing image according to an embodiment of the present invention.
Fig. 4 is a result of detecting a target in a high-resolution remote sensing image according to an embodiment of the present invention.
Detailed Description
The following detailed description of the embodiments of the present invention is provided with reference to the accompanying drawings.
The embodiment of the invention provides a convolutional neural network detection method for high-resolution remote sensing image target scale features, which comprises the following steps:
(1) the optimal scale of the remote sensing image target detection suggestion frame is as follows: counting the target scale range of the remote sensing image in the data set by establishing a large-scale high-resolution remote sensing image target detection data set to obtain the scale range of the remote sensing image target; and obtaining the optimal dimension of the high-resolution remote sensing image target detection suggestion frame according to the high-resolution remote sensing image target dimension distribution range.
The high-resolution remote sensing satellite generally images the earth surface in a top-down mode on a near-earth orbit, and the imaging process is easily influenced by illumination, weather and the like, so that the generated high-resolution remote sensing image scene is complex; and the targets in the image show the characteristics of small size and dense distribution in different groups. The setting of the target candidate region extraction scale in the convolutional neural network target detection architecture is of great importance to the influence of the target detection architecture precision. In order to fully count the scale range of a typical target region of interest of an image, a remote sensing image target detection data set WHU-RSONE comprising an airplane, a storage tank and a ship is established, wherein the data set comprises 5977 high-resolution remote sensing images, and the image size is 600 pixels multiplied by 600 pixels to 1372 pixels multiplied by 1024 pixels. 2460 remote sensing images contain 51866 targets, wherein 15703 aircraft (plane) targets, 24692 storage-tank (tank) targets and 11471 ship (ship) targets. The WHU-RSONE data set comprises target image data with different radiances and different scales, and statistics is carried out on the target scales in the data set, wherein the statistical information is shown in figure 1.
The optimal dimension map of the high-resolution remote sensing image target detection suggestion box obtained by the embodiment is shown in fig. 2. In fig. 2, a target region of interest is extracted from the last layer of feature map of the network, and 256-dimensional feature vectors corresponding to the target region of interest are input into the two classification layers and the coordinate regression layer to perform learning and training on whether the target region of interest is a target or not and learning and training on the coordinates of the target region of interest. In fig. 2, the left side is the scale size of the target detection suggestion frame in the existing network, and the right side is the scale size of the target detection suggestion frame in the network of the present invention.
In FIG. 1(a), only 6.58% of the aircraft targets are within the region covered by the region of interest of the nine targets in the left rectangular box of FIG. 2; 96.68% of the aircraft targets are within the region covered by the region of interest of the twelve targets in the right rectangular box of FIG. 2.
In fig. 1(b), only 0.65% of the tank targets are within the region covered by the region of interest of the nine targets in the left rectangular box of fig. 2; there are 97.5% of the tank targets in the area covered by the twelve target regions of interest in the right rectangular box of fig. 2.
In FIG. 1(c), only 24.24% of the ship targets are within the area covered by the region of interest of the nine targets in the left rectangular box of FIG. 2; 90% of the ship targets are within the area covered by the twelve target interest areas in the right rectangular box of fig. 2.
In FIG. 1(d), only 7.62% of the targets in the WHU-RSONE training set are within the region covered by the region of interest of the nine targets in the left rectangular box of FIG. 2; 95.61% of the objects are located in the area covered by the region of interest of the twelve objects in the right rectangular box of FIG. 2.
Statistics show that the target region of interest generated by setting four scales (16, 32, 64 and 128) and three proportions (1:2, 1:1 and 2:1) can effectively couple the scale range of a typical target in a remote sensing image. Then in the design of the convolutional neural network architecture of the embodiment of the present invention, the region suggestion network (RPN) generates the target region of interest size in the convolutional neural network architecture training and testing process using four scales (16, 32, 64, and 128) and three ratios (1:2, 1:1, and 2: 1).
Compared with the scale characteristic convolutional neural network identification method of the remote sensing image target, the statistical analysis is only roughly carried out on all target scales in the data set in the paper. In the patent, a data set with a larger scale than that in a paper is established, and the number of targets in the data set is 2 times that in the paper; carrying out statistical analysis on each target scale in the data set independently, and carrying out statistical analysis on all targets in the data set together to obtain the optimal recommended frame scale of each target in the image; the theory in the patent is more strict and reasonable compared with the thesis, and the optimal recommended frame size of target detection in the image can be obtained.
(2) The remote sensing image target detection convolutional neural network architecture: and setting a convolutional neural network architecture suitable for the target detection of the high-resolution remote sensing image according to the optimal dimension of the target suggestion frame, thereby realizing the target detection in the image.
The invention uses the design of the fast-RCNN architecture for reference, and the convolutional neural network architecture comprises two modules: the target area recommendation network RPN (generating a multi-scale and rotation-invariant target candidate area) and the target classification and accurate positioning network CNN (classifying the target candidate area and reducing the target candidate area positioning error). Compared with the scale characteristic convolutional neural network identification method of the remote sensing image target, the convolutional neural network structure in the patent is clearer and more reasonable than that in the paper, and the parameter details in the structure are optimally designed. The convolutional neural network architecture is divided into two parts in the patent, including a target area suggestion network (RPN) and a target classification and accurate positioning network (CNN), and the overall details of the architecture are clearer and more reasonable. The schematic diagram of the convolutional neural network architecture is shown in fig. 3, in the target detection network architecture of the present invention, five convolutional layers are used to extract an image target feature map, and the detailed parameters of each layer are as follows.
A first layer: inputting a convolution layer with a convolution template size of 7 multiplied by 7, a batch normalization layer, an activation layer with an activation function of Relu and a maximum pooling layer, and outputting 96 characteristic graphs;
a second layer: inputting a convolution layer with a convolution template size of 5 multiplied by 5, a batch normalization layer, an activation layer with an activation function of Relu and a maximum pooling layer, and outputting 256 characteristic graphs;
and a third layer: inputting a convolution layer with a convolution template size of 26 multiplied by 256 and an activation layer with an activation function of Relu, and outputting 384 characteristic graphs;
a fourth layer: inputting 13 × 13 × 384 convolution layers with convolution template size of 3 × 3 and activation layers with activation function of Relu, and outputting 384 characteristic graphs;
and a fifth layer: inputting a convolution layer with convolution template size of 3 × 3 and an activation layer with activation function of Relu, and outputting 256 feature maps.
1)RPN
The convolutional neural network architecture of the invention uses RPN to obtain a target candidate region on the last layer of feature map of the network architecture, and the generated target candidate region is used for training and testing the target detection of the whole architecture. Because a high-resolution remote sensing satellite generally images the ground in a near-earth orbit (400 km-600 km) from top to bottom, and a target in a generated remote sensing image has the characteristics of small scale, large scale change of the same kind of target, direction uncertainty and the like, the RPN generates 12 target candidate regions at each position of a feature map by using four scales (16, 32, 64 and 128) and three scales (1:2, 1:1 and 2:1) and is used for target detection training and testing of the convolutional neural network architecture.
2) Object classification and accurate positioning CNN
The RPN transmits the obtained target candidate region information to a target classification and accurate positioning CNN, and the target classification and accurate positioning CNN combines the candidate region information and a last layer of feature map in the framework to obtain a feature vector of the target candidate region. The feature vectors of the target candidate region are then passed to a region of interest pooling layer (convolution), obtaining target candidate region feature vectors of a specified size. And finally, the feature vectors with the specified sizes are transmitted to a full connection layer for training and testing of target recognition classification and regional coordinate regression. The size of the region-of-interest pooling layer is 7 multiplied by 7, and both the two fully-connected layers behind the region-of-interest pooling layer comprise 4096 neurons; and then respectively accessing a full connection layer and a target coordinate regression layer corresponding to the target classification layer, wherein the full connection layer corresponding to the target classification layer comprises n neurons, the target coordinate regression layer comprises 4n neurons, n is the classification number of the target in the image, and 4n represents the coordinate coefficient of n types of targets (x, y, w, h) corresponding to the target area. Wherein x is the coordinate of the horizontal axis of the central point, y is the coordinate of the vertical axis of the central point, w is the width of the target, and h is the height of the target.
3) Architecture training and testing
In the training phase, parameters in the convolutional neural network architecture of the present invention are initialized using the network parameters trained on imgNet. And training the parameters in the convolutional neural network architecture by adopting an end-to-end training method. And adding the RPN training loss, the target classification loss and the CNN loss for accurate positioning, and performing back propagation on the loss by using a Stochastic Gradient Descent (SGD) method to update parameters in the network. The RPN and target classification and pinpoint CNN training process is described in detail below:
① the positive and negative samples in the min-batch of each training of RPN are from a high resolution remote sensing image, in one image, 256 target candidate areas are randomly sampled to calculate the training loss of RPN, wherein the proportion of the positive and negative samples in the 256 target candidate areas is close to 1:1, if the number of the positive samples is less than 128, the insufficient part supplements the negative samples.
②, training the target classification and accurate positioning CNN by using the target candidate area generated by RPN, wherein, a large amount of overlapping redundancy exists between the target candidate areas generated by RPN, in order to eliminate the target candidate areas with large amount of overlapping redundancy, the target candidate area is restrained by NMS algorithm based on the probability that the target candidate area is the target, the IOU threshold of NMS algorithm is set to 0.7, after restraining the target candidate area by NMS algorithm, the candidate area with the target probability ranking of 2000 is selected as the primary min-batch of training the target classification and accurate positioning CNN, the training loss of the target classification and accurate positioning CNN is calculated by using 2000 target candidate areas.
In the test, if the IOU of one target detection area and the IOU of the real target area in the image are more than or equal to 0.5, the target detection area is considered as a correct target detection result, otherwise, the target detection area is considered as an incorrect target detection result. The number of iterative training of the architecture is set to 75000 in this document, wherein the learning rate of the previous 50000 iterative training of the architecture is 0.001; the learning rate of the last 25000 architectural iterative trainings is 0.0001. The network training momentum is 0.9 and the attenuation factor is 0.0005.
In the framework test stage, a high-resolution remote sensing image is input into the convolutional neural network framework, and the RPN generates 6000 target candidate areas on the high-resolution remote sensing image. And performing non-maximum suppression on the 6000 target candidate regions by using an NMS algorithm based on the probability that the target candidate regions are targets, wherein the IOU threshold value of the non-maximum suppression is set to be 0.7. And after the non-maximum value is inhibited, selecting the target candidate with the target probability confidence degree of 300 before the ranking, transmitting the target candidate into the target classification and accurate positioning CNN for target classification and coordinate accurate positioning, and realizing target detection in the image. The detection result of the invention on the high-resolution remote sensing image target is shown in figure 4.
In specific implementation, the automatic operation of the method can be realized by adopting a computer software technology.
The specific embodiments described herein are merely illustrative of the spirit of the invention. Various modifications or additions may be made to the described embodiments or alternatives may be employed by those skilled in the art without departing from the spirit or ambit of the invention as defined in the appended claims.

Claims (3)

1.一种高分辨率遥感影像目标尺度特征的卷积神经网络检测方法,包括以下步骤:1. A convolutional neural network detection method for target scale features of high-resolution remote sensing images, comprising the following steps: 步骤1,通过建立包含不同辐射亮度、不同尺度大小的大规模高分辨率遥感影像多目标检测数据集,对数据集中各目标尺度范围分别统计分析,及所有目标一起统计分析,获得高分辨率遥感影像中各目标尺度的最优分布范围;Step 1, by establishing a large-scale high-resolution remote sensing image multi-target detection data set containing different radiances and different scales, and statistical analysis of each target scale range in the data set, and statistical analysis of all targets together, to obtain high-resolution remote sensing. The optimal distribution range of each target scale in the image; 步骤2,根据高分辨率遥感影像中各目标尺度分布范围覆盖的区域面积,获得高分辨率遥感影像目标检测建议框的最优尺度大小;Step 2, according to the area covered by the distribution range of each target scale in the high-resolution remote sensing image, obtain the optimal scale size of the target detection proposal frame of the high-resolution remote sensing image; 步骤3,根据目标建议框的最优尺度大小,设置适合高分辨率遥感影像目标检测的卷积神经网络架构,实现影像中目标检测。Step 3: According to the optimal size of the target proposal frame, a convolutional neural network architecture suitable for target detection in high-resolution remote sensing images is set to realize target detection in images. 2.根据权利要求1所述高分辨率遥感影像目标尺度特征的卷积神经网络检测方法,其特征在于:步骤3中,所述卷积神经网络架构包括目标区域建议网络RPN和目标分类与精确定位网络CNN,2. according to the convolutional neural network detection method of the described high-resolution remote sensing image target scale feature of claim 1, it is characterized in that: in step 3, described convolutional neural network architecture comprises target area suggestion network RPN and target classification and accurate. localization network CNN, 所述目标区域建议网络RPN在特征图的每个位置生成多种目标候选区域,将目标候选区域信息传递给CNN;The target region suggestion network RPN generates multiple target candidate regions at each position of the feature map, and transmits the target candidate region information to CNN; 所述目标分类与精确定位网络CNN使用五层卷积层提取影像目标特征图,结合目标候选区域信息和最后一层特征图,获得目标候选区域的特征向量;然后目标候选区域的特征向量被传递给感兴趣区域池化层,获得指定大小的目标候选区域特征向量;最后指定大小的特征向量被传递至全连接层,用于目标识别分类和区域坐标回归的训练和测试。The target classification and precise localization network CNN uses five layers of convolution layers to extract the image target feature map, and combines the target candidate region information and the last layer of feature maps to obtain the feature vector of the target candidate region; then the feature vector of the target candidate region is passed. For the pooling layer of the region of interest, the feature vector of the target candidate region of the specified size is obtained; finally, the feature vector of the specified size is passed to the fully connected layer for training and testing of target recognition classification and region coordinate regression. 3.根据权利要求2所述高分辨率遥感影像目标尺度特征的卷积神经网络检测方法,其特征在于:感兴趣区池化层大小为7×7,感兴趣区池化层后两个全连接层均包含4096个神经元;然后目标分类层对应的全连接层包含n个神经元,目标坐标回归层包含4n个神经元,n为影像中目标的分类数,4n表示目标区域对应的n类目标x,y,w,h坐标系数;其中,x是指中心点横轴坐标,y指中心点纵轴坐标,w为目标宽度,h为目标高度。3. The convolutional neural network detection method of high-resolution remote sensing image target scale features according to claim 2, wherein the size of the pooling layer in the region of interest is 7×7, and the size of the pooling layer in the region of interest is 7×7. The connection layer contains 4096 neurons; then the fully connected layer corresponding to the target classification layer contains n neurons, and the target coordinate regression layer contains 4n neurons, where n is the number of classifications of the target in the image, and 4n represents the n corresponding to the target area. Class target x, y, w, h coordinate coefficients; where x refers to the horizontal axis coordinate of the center point, y refers to the vertical axis coordinate of the center point, w is the target width, and h is the target height.
CN201911179838.0A 2019-11-27 2019-11-27 Convolutional neural network detection method for high-resolution remote sensing image target scale features Pending CN111046756A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201911179838.0A CN111046756A (en) 2019-11-27 2019-11-27 Convolutional neural network detection method for high-resolution remote sensing image target scale features

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201911179838.0A CN111046756A (en) 2019-11-27 2019-11-27 Convolutional neural network detection method for high-resolution remote sensing image target scale features

Publications (1)

Publication Number Publication Date
CN111046756A true CN111046756A (en) 2020-04-21

Family

ID=70233810

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201911179838.0A Pending CN111046756A (en) 2019-11-27 2019-11-27 Convolutional neural network detection method for high-resolution remote sensing image target scale features

Country Status (1)

Country Link
CN (1) CN111046756A (en)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112419333A (en) * 2020-11-17 2021-02-26 武汉大学 Remote sensing image self-adaptive feature selection segmentation method and system
CN114842353A (en) * 2022-05-06 2022-08-02 自然资源部第一海洋研究所 Neural network remote sensing image target detection method based on self-adaptive target direction
CN114882376A (en) * 2022-05-06 2022-08-09 自然资源部第一海洋研究所 Convolutional neural network remote sensing image target detection method based on optimal anchor point scale
CN116563105A (en) * 2023-04-18 2023-08-08 武汉大学 Method for optimizing crowd-sourced satellite remote sensing image data set and computer readable medium

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108597053A (en) * 2018-04-25 2018-09-28 北京御航智能科技有限公司 Shaft tower and channel targets identification based on image data and neural network and defect diagnostic method
CN108734219A (en) * 2018-05-23 2018-11-02 北京航空航天大学 A kind of detection of end-to-end impact crater and recognition methods based on full convolutional neural networks structure
CN109190636A (en) * 2018-07-30 2019-01-11 北京航空航天大学 A kind of remote sensing images Ship Target information extracting method
CN109409252A (en) * 2018-10-09 2019-03-01 杭州电子科技大学 A kind of traffic multi-target detection method based on modified SSD network
CN109961049A (en) * 2019-03-27 2019-07-02 东南大学 A method for cigarette brand recognition in complex scenes
CN110188807A (en) * 2019-05-21 2019-08-30 重庆大学 Tunnel pedestrian target detection method based on cascaded super-resolution network and improved Faster R-CNN

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108597053A (en) * 2018-04-25 2018-09-28 北京御航智能科技有限公司 Shaft tower and channel targets identification based on image data and neural network and defect diagnostic method
CN108734219A (en) * 2018-05-23 2018-11-02 北京航空航天大学 A kind of detection of end-to-end impact crater and recognition methods based on full convolutional neural networks structure
CN109190636A (en) * 2018-07-30 2019-01-11 北京航空航天大学 A kind of remote sensing images Ship Target information extracting method
CN109409252A (en) * 2018-10-09 2019-03-01 杭州电子科技大学 A kind of traffic multi-target detection method based on modified SSD network
CN109961049A (en) * 2019-03-27 2019-07-02 东南大学 A method for cigarette brand recognition in complex scenes
CN110188807A (en) * 2019-05-21 2019-08-30 重庆大学 Tunnel pedestrian target detection method based on cascaded super-resolution network and improved Faster R-CNN

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
MATTHEW D. ZEILER 等: "Visualizing and Understanding Convolutional Networks", 《ARXIV》 *
董志鹏 等: "遥感影像目标的尺度特征卷积神经网络识别法", 《测绘学报》 *

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112419333A (en) * 2020-11-17 2021-02-26 武汉大学 Remote sensing image self-adaptive feature selection segmentation method and system
CN112419333B (en) * 2020-11-17 2022-04-29 武汉大学 A method and system for adaptive feature selection and segmentation of remote sensing images
CN114842353A (en) * 2022-05-06 2022-08-02 自然资源部第一海洋研究所 Neural network remote sensing image target detection method based on self-adaptive target direction
CN114882376A (en) * 2022-05-06 2022-08-09 自然资源部第一海洋研究所 Convolutional neural network remote sensing image target detection method based on optimal anchor point scale
CN114882376B (en) * 2022-05-06 2024-03-22 自然资源部第一海洋研究所 Convolutional neural network remote sensing image target detection method based on optimal anchor point scale
CN114842353B (en) * 2022-05-06 2024-04-02 自然资源部第一海洋研究所 Neural network remote sensing image target detection method based on self-adaptive target direction
CN116563105A (en) * 2023-04-18 2023-08-08 武汉大学 Method for optimizing crowd-sourced satellite remote sensing image data set and computer readable medium
CN116563105B (en) * 2023-04-18 2024-02-02 武汉大学 Method for optimizing crowd-sourced satellite remote sensing image data set and computer readable medium

Similar Documents

Publication Publication Date Title
Dong et al. Object detection in high resolution remote sensing imagery based on convolutional neural networks with suitable object scale features
US20230252761A1 (en) Method for classifying hyperspectral images on basis of adaptive multi-scale feature extraction model
CN112836713B (en) Mesoscale convective system recognition and tracking method based on image anchor-free frame detection
Karami et al. Automatic plant counting and location based on a few-shot learning technique
CN108416378B (en) A large-scene SAR target recognition method based on deep neural network
CN108647655B (en) Low-altitude aerial image power line foreign object detection method based on light convolutional neural network
Wang et al. A deep-learning-based sea search and rescue algorithm by UAV remote sensing
CN111046756A (en) Convolutional neural network detection method for high-resolution remote sensing image target scale features
Hartawan et al. Disaster victims detection system using convolutional neural network (CNN) method
CN112819830A (en) Individual tree crown segmentation method based on deep learning and airborne laser point cloud
CN111310756A (en) Damaged corn particle detection and classification method based on deep learning
CN117409339A (en) Unmanned aerial vehicle crop state visual identification method for air-ground coordination
CN115546656B (en) A method for extracting aquaculture areas from remote sensing images based on deep learning
CN110866472A (en) A UAV ground moving target recognition and image enhancement system and method
CN112883850A (en) Multi-view aerospace remote sensing image matching method based on convolutional neural network
CN116363532A (en) Traffic target detection method for UAV images based on attention mechanism and reparameterization
Li et al. An aerial image segmentation approach based on enhanced multi-scale convolutional neural network
CN117392382A (en) Single tree fruit tree segmentation method and system based on multi-scale dense instance detection
Gai et al. Cherry detection algorithm based on improved YOLOv5s network
Li et al. Remote sensing image scene classification model based on dual knowledge distillation
Alburshaid et al. Palm trees detection using the integration between GIS and deep learning
Wei et al. Novel green-fruit detection algorithm based on D2D framework
Chen et al. Ship detection with optical image based on attention and loss improved YOLO
CN117011724A (en) Unmanned aerial vehicle target detection positioning method
CN112949726B (en) ISCCP cloud classification method, system, medium and terminal based on FY-4A satellite

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
RJ01 Rejection of invention patent application after publication

Application publication date: 20200421

RJ01 Rejection of invention patent application after publication