WO2020143323A1 - 遥感影像分割方法、装置及存储介质、服务器 - Google Patents
遥感影像分割方法、装置及存储介质、服务器 Download PDFInfo
- Publication number
- WO2020143323A1 WO2020143323A1 PCT/CN2019/118530 CN2019118530W WO2020143323A1 WO 2020143323 A1 WO2020143323 A1 WO 2020143323A1 CN 2019118530 W CN2019118530 W CN 2019118530W WO 2020143323 A1 WO2020143323 A1 WO 2020143323A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- features
- target
- target image
- layer
- neural network
- 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.)
- Ceased
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
Definitions
- the present application relates to the technical field of image detection and image classification, and in particular to a remote sensing image segmentation method, device, storage medium, and server.
- Remote sensing image data is also used as the basic data in the three spatial information technologies of geographic information system (Geographic information system, GIS), global positioning system (Global Positioning System, GPS), remote sensing surveying and mapping technology (remote sensing system, RS), which is widely used Various fields such as environmental monitoring, resource investigation, land use, urban planning, natural disaster analysis and military.
- GIS Geographic information system
- GPS Global Positioning System
- RS remote sensing surveying and mapping technology
- a remote sensing image segmentation method includes: acquiring a target image, and extracting the layered features of the target image through convolutional layers of different scales in a deep convolutional neural network; The layered features extracted by the convolutional layers of two different scales determine the target features of the target image; determine the object class of each object in the target image according to the target feature, and compare the target image according to the object class Split.
- embodiments of the present application also provide a remote sensing image segmentation device, including: a target image acquisition module for acquiring a target image, and extracting the target image through convolutional layers of different scales in a deep convolutional neural network Layered features; target feature extraction module, used to determine the target features of the target image based on the layered features extracted by convolutional layers of two different scales in turn; segmentation module, used to determine the target features based on the target features For the object category of each object in the target image, segment the target image according to the object category.
- the embodiments of the present application also provide a computer-readable storage medium.
- the computer-readable storage medium stores a computer program.
- a remote sensing image segmentation method is implemented.
- the remote sensing image segmentation method includes: obtaining a target image, extracting the layered features of the target image through convolutional layers of different scales in a deep convolutional neural network; and sequentially extracting the layered features based on two different scale convolutional layers Determine the target feature of the target image; determine the object category of each object in the target image according to the target feature, and segment the target image according to the object category.
- embodiments of the present application further provide a server, including: one or more processors; a memory; one or more application programs, wherein the one or more application programs are stored in the memory And is configured to be executed by the one or more processors, and the one or more application programs are configured to perform a remote sensing image segmentation method
- the remote sensing image segmentation method includes: acquiring a target image, through the depth volume The convolutional layers of different scales of the product neural network extract the layered features of the target image; sequentially determine the target features of the target image based on the layered features extracted by the convolutional layers of two different scales; based on the target features The object category of each object in the target image is determined, and the target image is segmented according to the object category.
- the embodiments of the present application can achieve pixel-level segmentation of remote sensing images, and can also reduce the degree of image segmentation of remote sensing images by factors such as clouds and strong light.
- FIG. 1 is a schematic flowchart of an implementation manner in a typical example of a remote sensing image segmentation method of this application;
- FIG. 2 is a schematic structural diagram of a middle convolutional neural network of a remote sensing image segmentation method of the present application
- FIG. 3 is a schematic structural diagram of a typical embodiment of a remote sensing image segmentation device of the present application.
- FIG. 4 is a schematic structural diagram of an embodiment of an application server.
- a remote sensing image segmentation method provided by an embodiment of the present application in one of the embodiments, as shown in FIG. 1, includes: S100, S200, and S300.
- S100 Obtain a target image, and extract layered features of the target image through convolutional layers of different scales in a deep convolutional neural network;
- S200 Determine target features of the target image based on the layered features extracted by convolutional layers of two different scales in sequence;
- S300 Determine the object category of each object in the target image according to the target feature, and segment the target image according to the object category.
- the target image in this application is a remote sensing image, which includes a type of special images taken by satellites and aircraft.
- the segmentation of the remote sensing image in this application is mainly achieved by deep convolutional neural networks.
- the remote sensing image is input into a deep convolutional neural network.
- the deep convolutional neural network After acquiring the remote sensing image, the deep convolutional neural network convolves the remote sensing image, and extracts each convolution by convolution Layered features in layers and pooling layers, where layered features include features after passing through convolutional layers and pooling layers at different scales.
- the neural network results, at the same scale, there is a pooling layer and more than one convolutional layer.
- the convolutional layer extracts the features of the input image, and the pooling layer: compresses the input feature map, on the one hand makes the feature map smaller, simplifies the network calculation complexity; on the one hand, features compression to extract the main features.
- the pooling layer of the next scale samples the features extracted by the previous scale convolutional layer, and the pooling layer downsamples the features extracted by the previous scale convolutional layer.
- the convolution layer at the scale of the pooling layer is convolved to refine the features after the pooling layer, and then more accurate and/or more features can be obtained, so
- the extracted features will be more able to express the corresponding target, and thus the deep convolutional neural network can more accurately distinguish the object category in the remote sensing image, such as distinguishing the target from the background .
- different features will be extracted at different scales, and then hierarchical features can be obtained.
- the layered features extracted by the convolutional layer at different scales are fused.
- the target features include one or more of the underlying geometric features, mid-level texture features, and high-level semantic features of a target, making the target features rich
- the deep convolutional neural network can synthesize the target
- the features accurately classify each object in the target image, and then segment the remote sensing image according to the classification. If the target feature and the background are distinguished, the remote sensing image is segmented to cover the same type of object with the same color, different The color of the objects covered by the class is different, which can distinguish the objects in the remote sensing image, that is, the segmentation of the remote sensing image.
- the segmentation of the remote sensing image is achieved by up-sampling through the deep convolutional neural network to achieve spatial output.
- Pixel-level segmentation while implementing the steps of this method through a deep convolutional neural network, because the deep convolutional neural network extracts shallow to deep multi-scale features and performs jumpy fusion at different scales, which can improve remote sensing images in the cloud, Segmentation accuracy under strong light and other interference factors.
- the sequentially determining the target features of the target image based on the layered features extracted by the convolutional layers of two different scales includes:
- the target feature of the target image is determined based on the fusion feature.
- the feature of the target is richer, and then the accurate description and discrimination (classification) of the target are achieved, and the extracted from the last two convolutional layers in two different scales are fused in sequence Hierarchical features, due to the different scale convolution layers, the extracted feature levels are different.
- the target features in the target image can also be determined by the fusion features including two scales, in particular, the same target has the aforementioned fusion features After that, it makes the features of the same target more abundant, which is convenient for the accurate description and classification of the objects in the remote sensing image through the rich target features.
- the sequentially fusing the layered features extracted from the last two convolutional layers in two different scales to obtain the fusion feature includes:
- the deconvolution layer is located in the adjacent scales, and the restored features are located in the deconvolution layer in;
- the determining the target feature of the target image based on the fusion feature includes:
- the fusion feature is determined as the target feature.
- FIG. 2 the structure of the deep convolutional neural network of this application is shown in FIG. 2, which includes the convolution layer A, the deconvolution layer B, and the pooling layer C.
- the layer C and/or the convolution layer A the features of the objects in the remote sensing image are realized.
- the last convolution layer of the last scale is used as the reference, that is, from the convolution layer , Advance forward for deconvolution, where the deconvolution process is implemented in the deconvolution layer.
- the features extracted by the last convolutional layer A are deconvolved at the deconvolution layer B to obtain the convolutional layer at the previous scale of the scale, where
- the deconvolution layer contains the hierarchical features extracted at this scale.
- the deconvolution layer includes reduced features, that is, the features extracted by the convolution at this scale are converted to the previous scale to facilitate the features at the two scales Perform fusion.
- the fusion process is to merge the restored features obtained by deconvolution of the convolutional layer at the last scale L with the layered features of the last convolutional layer at the previous scale M adjacent to the last scale L, that is, to connect the last scale
- the deconvolution layer corresponding to the convolution layer under L and the last convolution layer at the previous scale M adjacent to the scale can obtain the aforementioned convolution union, which includes the aforementioned
- the fusion feature connects the deconvolution layer and the convolution layer. Therefore, the fusion feature should include the restoration feature in the deconvolution layer and the layered feature of the convolution layer.
- the convolution union that is, after at least one convolution layer, that is, extract the features of the fusion feature
- the layered features extracted by the last convolutional layer at scale N and the restored features extracted by the deconvolution at scale N are fused, that is, the last one at scale N is connected
- the convolutional layer and the deconvolution layer at scale N obtain another convolutional union, and then the convolutional union is sequentially processed as described above to obtain the target feature.
- the remote sensing image can be segmented according to the target feature.
- the target features include one or more of semantic features, geometric features, and texture features.
- the determining the object type of each object in the target image according to the target feature includes:
- the object category of each object in the target image is determined based on the mapping relationship.
- the target feature is a pixel-level feature, combined with the type of the target feature, in order to be able to segment the remote sensing image, obtain the target feature and object
- the mapping relationship of categories that is, the voice features, texture features, geometric features, etc. corresponding to the same object, based on the mapping relationship and the target features, the category of each object in the remote sensing image can be determined, so that the subsequent remote sensing image segmentation as described above can be achieved later.
- the remote sensing image is divided into pixels.
- the layered features of the remote sensing image are first obtained through convolution and pooling layers of different scales, and then the last convolutional base layer is deconvoluted to restore the scale to the size of the previous convolution.
- the last convolutional layer of the previous part is connected to the deconvolutional layer (as shown by the dotted line and part C in Figure 2).
- the features at this time include the features with the highest abstraction and the features with the next highest abstraction.
- the subsequent steps can be deduced by analogy. Each time the end convolution is deconvolved, and then connected to the corresponding convolution layer, and then the convolution is performed.
- the underlying convolution reflects the underlying features such as the edges, corners, and points of the target
- the intermediate convolution reflects the intermediate features such as texture
- the high-level convolution reflects the semantic features such as high-level features.
- the feature map is mapped to two features at the back end of the entire network, and the indexes corresponding to the two features acquired at this time are categories. There are two types in this network: target features and background.
- the extracting layered features of the target image through convolutional layers of different scales of the deep convolutional neural network includes:
- the layered features of the target image in each convolutional layer of the deep convolutional neural network are extracted through the image segmentation model.
- the training the deep convolutional neural network according to the target image to obtain an image segmentation model includes:
- the loss function is input to the deep convolutional neural network to obtain the image segmentation model.
- the remote sensing image needs to be input into a deep convolutional neural network for training to determine the image segmentation model of the remote sensing image.
- the participation of the deep convolutional neural network in this application is as follows:
- the image segmentation model is used to perform the process as described above.
- the target image is mainly used for training to determine the loss function of the deep convolutional neural network. After inputting the loss function into the deep convolutional neural network, the image segmentation model can be obtained.
- the determination process of the loss function is as follows:
- the loss function uses the sum of the cross-entropy cost function and the regular term.
- the cross-entropy function is as follows:
- y is the expected output
- ⁇ is the actual output of the neural network
- x is the input value.
- the cross-entropy loss function can measure the similarity between the actual output and the expected output of the neural network. As a loss function, it can avoid the problem of a decrease in the learning rate of the mean square error loss function when the gradient is reduced. At the same time, as the logarithmic function of cross entropy, the gradient value is still high when approaching the upper boundary, so the convergence rate of the model will not slow down.
- the regularization term is used to attenuate the parameters of the convolutional neural network, resulting in sparsity and reduced feature vectors, thereby reducing the complexity of the model and preventing overfitting of deep neural networks during training.
- the loss function is:
- ⁇ is the coefficient of the regular term
- ⁇ is the model parameter
- k is the number of parameters.
- the optimizer used for training is the batch stochastic gradient descent method.
- the momentum term In order to stabilize the update process, we introduced the momentum term.
- the formula is as follows:
- ⁇ is the momentum term, set to 0.99
- t is the number of iterations
- ⁇ is the learning rate, set to 0.00004
- g t is the gradient of x at time t.
- the number of samples in each batch is 4 during training.
- An embodiment of the present application further provides a remote sensing image segmentation device.
- the target image acquisition module 100, target feature extraction module 200, and segmentation module 300 are included:
- the target image acquisition module 100 is used to acquire a target image, and extract layered features of the target image through convolutional layers of different scales in a deep convolutional neural network;
- the target feature extraction module 200 is used to sequentially determine the target features of the target image based on the layered features extracted by the convolutional layers of two different scales;
- the segmentation module 300 is configured to determine the object category of each object in the target image according to the target feature, and segment the target image according to the object category.
- a remote sensing image segmentation method device provided in an embodiment of the present application further includes: a fusion unit 210 for sequentially fusing the layers extracted by the last two convolutional layers in two different scales Feature to obtain a fusion feature; a first target feature determination unit 220 is used to determine the target feature of the target image based on the fusion feature.
- the last convolution layer acquisition unit 211 is used to acquire the last scale and the features extracted by the last convolution layer of the scale; the deconvolution unit 212 is used to deconvolution the features to obtain the Reduction features in adjacent scales of one scale and one greater scale, the deconvolution layer is located in the adjacent scale, the reduction features are located in the deconvolution layer; a connection unit 213 is used to connect the A deconvolution layer and the last convolution layer in the same scale as the adjacent scale for extracting the hierarchical features to obtain a convolution union; a convolution unit 214 is used to The convolution union performs convolution to obtain the fusion feature, the convolution union and the deconvolution are in the same scale; the judgment unit 215 is used to judge whether the adjacent scale is the deep convolutional neural network If not, obtain the last convolutional layer of the adjacent scale, and continue the step of deconvolution of the features extracted by the convolutional layer.
- the second target feature determining unit 221 is configured to determine the fusion feature as the target feature when the adjacent scale is the first layer of the deep convolutional neural network.
- the mapping relationship obtaining unit 310 is used to obtain the mapping relationship between the target feature and the object class; the object class determining unit 320 is used to determine the object class of each object in the target image based on the mapping relationship.
- a segmentation model obtaining unit 110 is used to train the deep convolutional neural network according to the target image to obtain an image segmentation model; a layered feature extraction unit 120 is used to extract the target image through the image segmentation model in The layered features in each convolutional layer of the deep convolutional neural network.
- the loss function determining unit 111 is used to train the deep convolutional neural network according to the target image and determine the loss function of the deep convolutional neural network; the image segmentation model obtaining unit 112 is used to convert the loss function Input the deep convolutional neural network to obtain the image segmentation model.
- a remote sensing image segmentation method device provided by an embodiment of the present application can implement the foregoing remote sensing image segmentation method embodiment.
- a computer-readable storage medium provided by an embodiment of the present application.
- a computer program is stored on the computer-readable storage medium.
- the remote sensing image segmentation method according to any one of the technical solutions is implemented.
- the computer-readable storage medium includes but is not limited to any type of disk (including floppy disk, hard disk, optical disk, CD-ROM, and magneto-optical disk), ROM (Read-Only Memory, read-only memory), RAM (Random Access Memory (random memory), EPROM (Erasable Programmable Read-Only Memory, erasable programmable read-only memory), EEPROM (Electrically Erasable Programmable Read-Only Memory, erasable programmable read-only memory), flash memory, magnetic card or Ray card.
- the storage device includes any medium that stores or transmits information in a readable form by a device (eg, computer, mobile phone), and may be a read-only memory, a magnetic disk, or an optical disk.
- a computer-readable storage medium provided by an embodiment of the present application.
- the storage medium may be non-volatile or volatile.
- the computer-readable storage medium may implement the embodiment of the above method for segmenting remote sensing images.
- the convolutional neural network is used to extract multi-scale features from shallow to deep, and the deep, middle and shallow features are also merged in a jumping manner to obtain high-level semantic information and low-level geometric information, and finally through upsampling Spatial output, to achieve the purpose of pixel-level segmentation, using the convolutional neural network in this application can improve the segmentation accuracy of remote sensing images under interference factors such as clouds and strong light; a remote sensing image segmentation method provided by embodiments of this application , Including: acquiring a target image, extracting the layered features of the target image through convolutional layers of different scales of the deep convolutional neural network; determining the target based on the layered features extracted by convolutional layers of two different scales The target feature of the image; determine the object category of each object in the
- the target image in this application is a remote sensing image, which includes a type of special images taken by satellites and aircraft.
- the segmentation of the remote sensing image in this application is mainly achieved by deep convolutional neural networks.
- the remote sensing image is input into a deep convolutional neural network.
- the deep convolutional neural network After acquiring the remote sensing image, the deep convolutional neural network convolves the remote sensing image, and extracts each convolution by convolution Layered features in layers and pooling layers, where layered features include features after passing through convolutional layers and pooling layers at different scales.
- the neural network results, at the same scale, there is a pooling layer and more than one convolutional layer.
- the convolutional layer extracts the features of the input image, and the pooling layer: compresses the input feature map, on the one hand makes the feature map smaller, simplifies the network calculation complexity; on the one hand, features compression to extract the main features.
- the pooling layer of the next scale samples the features extracted by the previous scale convolutional layer, and the pooling layer downsamples the features extracted by the previous scale convolutional layer.
- the convolution layer at the scale of the pooling layer is convolved to refine the features after passing the pooling layer, and then more accurate and/or more features can be obtained, so
- the extracted features will be more able to express the corresponding target, and thus the deep convolutional neural network can more accurately distinguish the object category in the remote sensing image, such as distinguishing the target from the background .
- different features will be extracted at different scales, and then hierarchical features can be obtained.
- the layered features extracted by the convolutional layer at different scales are fused.
- the target features include one or more of the underlying geometric features, mid-level texture features, and high-level semantic features of a target, making the target features rich
- the deep convolutional neural network can synthesize the target
- the features accurately classify each object in the target image, and then segment the remote sensing image according to the classification. If the target feature and the background are distinguished, the remote sensing image is segmented to cover the same type of object with the same color, different The color of the objects covered by the class is different, which can distinguish the objects in the remote sensing image, that is, the segmentation of the remote sensing image.
- the segmentation of the remote sensing image is achieved by up-sampling through the deep convolutional neural network to achieve spatial output.
- Pixel-level segmentation while implementing the steps of this method through a deep convolutional neural network, because the deep convolutional neural network extracts multi-scale features from shallow to deep, and performs jumping fusion at different scales, which can improve remote sensing images in the cloud, Segmentation accuracy under strong light and other interference factors.
- the present application further provides a server.
- the memory 505 may be used to store application programs 501 and various functional modules.
- the processor 503 runs the application programs 501 stored in the memory 505 to execute various functional applications of the device and data processing.
- the memory 505 may be internal memory or external memory, or include both internal memory and external memory.
- the internal memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory, or random access memory.
- ROM read-only memory
- PROM programmable ROM
- EPROM electrically programmable ROM
- EEPROM electrically erasable programmable ROM
- flash memory or random access memory.
- the external memory may include hard disks, floppy disks, ZIP disks, U disks, magnetic tapes, etc.
- the memories disclosed in this application include but are not limited to these types of memories.
- the memory 505 disclosed in this application is only an example and not a limitation.
- the input unit 507 is used to receive input of signals, as well as personal information and related physical condition information input by the user.
- the input unit 507 may include a touch panel and other input devices.
- the touch panel can collect the touch operations of the customer on or near it (such as the operation of the customer using any finger or stylus on the touch panel or near the touch panel), and according to the preset
- the program drives the corresponding connection device; other input devices may include but are not limited to one or more of a physical keyboard, function keys (such as playback control keys, switch keys, etc.), trackball, mouse, joystick, and so on.
- the display unit 509 can be used to display information input by the customer or information provided to the customer and various menus of the computer device.
- the display unit 509 may take the form of a liquid crystal display, an organic light-emitting diode, or the like.
- the processor 503 is the control center of the computer equipment. It uses various interfaces and lines to connect the various parts of the entire computer. By running or executing the software programs and/or modules stored in the memory 503, and calling the data stored in the memory, the execution Various functions and processing data.
- One or more processors 503 shown in FIG. 4 can execute and realize the function of the target image acquisition module 100 shown in FIG.
- the function of the target feature extraction module 200, the function of the segmentation module 300, the target image acquisition module 100 Function, the function of the target feature extraction module 200, the function of the segmentation module 300, the function of the fusion unit 210, the function of the first target feature determination unit 220, the function of the last convolution layer acquisition unit 211, the function of the deconvolution unit 212 Function, function of connection unit 213, function of convolution unit 214, function of judgment unit 215, function of connection unit 212, function of second target feature determination unit 221, function of mapping relationship acquisition unit 310, object category determination unit 320 ,
- the server includes one or more processors 503, and one or more memories 505, one or more applications 501, wherein the one or more applications 501 are stored in the memory 505 Is configured to be executed by the one or more processors 503, and the one or more application programs 301 are configured to execute the remote sensing image segmentation method described in the above embodiment.
- a server provided by an embodiment of the present application can implement the above-mentioned embodiment of the remote sensing image segmentation method.
- a convolutional neural network is used to extract multi-scale features from shallow to deep, and the deep, middle and Shallow features to obtain high-level semantic information and low-level geometric information, and finally achieve spatial output through upsampling to achieve the purpose of pixel-level segmentation.
- a remote sensing image segmentation method includes: acquiring a target image, and extracting the target image through convolutional layers of different scales in a deep convolutional neural network Hierarchical features; sequentially determining the target features of the target image based on the hierarchical features extracted by convolutional layers of two different scales; determining the object class of each object in the target image based on the target features, and based on the objects The category divides the target image.
- the target image in this application is a remote sensing image, which includes a type of special images taken by satellites and aircraft.
- the segmentation of the remote sensing image in this application is mainly achieved by deep convolutional neural networks.
- the remote sensing image is input into a deep convolutional neural network.
- the deep convolutional neural network After acquiring the remote sensing image, the deep convolutional neural network convolves the remote sensing image, and extracts each convolution by convolution Layered features in layers and pooling layers, where layered features include features after passing through convolutional layers and pooling layers at different scales.
- the neural network results, at the same scale, there is a pooling layer and more than one convolutional layer.
- the convolutional layer extracts the features of the input image, and the pooling layer: compresses the input feature map, on the one hand makes the feature map smaller, simplifies the network calculation complexity; on the one hand, features compression to extract the main features.
- the pooling layer of the next scale samples the features extracted by the previous scale convolutional layer, and the pooling layer downsamples the features extracted by the previous scale convolutional layer.
- the convolution layer at the scale of the pooling layer is convolved to refine the features after passing the pooling layer, and then more accurate and/or more features can be obtained, so
- the extracted features will be more able to express the corresponding target, and thus the deep convolutional neural network can more accurately distinguish the object category in the remote sensing image, such as distinguishing the target from the background .
- different features will be extracted at different scales, and then scored layer features can be obtained.
- the layered features extracted by the convolutional layer at different scales are fused.
- the target features include one or more of the underlying geometric features, mid-level texture features, and high-level semantic features of a target, making the target features rich
- the deep convolutional neural network can synthesize the target
- the features accurately classify each object in the target image, and then segment the remote sensing image according to the classification. If the target feature and the background are distinguished, the remote sensing image is segmented to cover the same type of object with the same color, different The color of the objects covered by the class is different, which can distinguish the objects in the remote sensing image, that is, the segmentation of the remote sensing image.
- the segmentation of the remote sensing image is achieved by up-sampling through the deep convolutional neural network to achieve spatial output.
- Pixel-level segmentation while implementing the steps of this method through a deep convolutional neural network, because the deep convolutional neural network extracts shallow to deep multi-scale features and performs jumpy fusion at different scales, which can improve remote sensing images in the cloud, Segmentation accuracy under strong light and other interference factors.
- the server provided by the embodiment of the present application can implement the embodiment of the remote sensing image segmentation method provided above.
Landscapes
- Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
Abstract
一种遥感影像分割方法,图像检测、图像分类技术领域。该方法包括:获取目标图像,通过深度卷积神经网络的各卷积层提取目标图像的分层特征(S100);依次基于两不同尺度卷积层提取出的分层特征确定目标图像的目标特征(S200);依据目标特征确定目标图像中各物体的物体类别,依据物体类别对目标图像进行分割(S300)。在该方法中采用卷积神经网络提取由浅到深的多尺度特征,还通过跳跃式的方式融合深层、中层和浅层特征,用以获取高级别的语义信息和低级别的几何信息,之后再通过上采样实现空间输出,以达到像素级别分割的目的,提高遥感影像在云、强光照等干扰因素下的分割精度。
Description
本申请要求于2019年1月8日提交中国专利局、申请号为2019100153442,发明名称为“遥感影像分割方法、装置及存储介质、服务器”的中国专利申请的优先权,其全部内容通过引用结合在本申请中。
本申请涉及图像检测、图像分类技术领域,具体涉及一种遥感影像分割方法、装置及存储介质、服务器。
遥感影像数据还作为地理信息系统(Geographic information System,GIS),全球定位系统(Global Positioning System,GPS)、遥感测绘技术(remote sensing system,RS)三大空间信息技术中的基础数据,广泛应用于环境监测、资源调查、土地利用、城市规划、自然灾害分析和军事等各个领域。近年来,随着高分辨率遥感卫星、成像雷达以及无人机驾驶飞机(Unmanned Aerial Vehicle)技术的发展,发明人意识到遥感影像数据进一步呈现海量、复杂和高分辨率的特点,实现精细分割对于促进遥感影像信息的准确提取和数据共享具有重要的研究意义和应用价值。
发明内容
为克服以上技术问题,特别是现有遥感影像中信息提取不够准确,导致不能够精准分割遥感影像的问题,特提出以下技术方案:
为解决以上技术问题,本申请实施例提供的一种遥感影像分割方法,包括:获取目标图像,通过深度卷积神经网络不同尺度的各卷积层提取所述目标图像的分层特征;依次基于两不同尺度卷积层提取出的所述分层特征确定所述目标图像的目标特征;依据所述目标特征确定所述目标图像中各物体的物体类别,依据所述物体类别对所述目标图像进行分割。
为解决上述技术问题,本申请实施例还提供一种遥感影像分割装置, 包括:目标图像获取模块,用于获取目标图像,通过深度卷积神经网络不同尺度的各卷积层提取所述目标图像的分层特征;目标特征提取模块,用于依次基于两不同尺度卷积层提取出的所述分层特征确定所述目标图像的目标特征;分割模块,用于依据所述目标特征确定所述目标图像中各物体的物体类别,依据所述物体类别对所述目标图像进行分割。
为解决上述技术问题,本申请实施例还提供了一种计算机可读存储介质,所述计算机可读存储介质上存储有计算机程序,该程序被处理器执行时实现一种遥感影像分割方法,所述遥感影像分割方法包括:获取目标图像,通过深度卷积神经网络不同尺度的各卷积层提取所述目标图像的分层特征;依次基于两不同尺度卷积层提取出的所述分层特征确定所述目标图像的目标特征;依据所述目标特征确定所述目标图像中各物体的物体类别,依据所述物体类别对所述目标图像进行分割。
为解决上述技术问题,本申请实施例还提供了一种服务器,包括:一个或多个处理器;存储器;一个或多个应用程序,其中所述一个或多个应用程序被存储在所述存储器中并被配置为由所述一个或多个处理器执行,所述一个或多个应用程序配置用于执行一种遥感影像分割方法,所述遥感影像分割方法包括:获取目标图像,通过深度卷积神经网络不同尺度的各卷积层提取所述目标图像的分层特征;依次基于两不同尺度卷积层提取出的所述分层特征确定所述目标图像的目标特征;依据所述目标特征确定所述目标图像中各物体的物体类别,依据所述物体类别对所述目标图像进行分割。
本申请实施例能实现对遥感影像的像素级分割,也能降低云、强光照等因素对遥感影像分割影像程度。
图1为本申请遥感影像分割方法的典型实施例中一种实施方式的流程示意图;
图2为本申请遥感影像分割方法的中卷积神经网络的结构示意图;
图3为本申请遥感影像分割装置的典型实施例的结构示意图;
图4为本申请服务器的一实施例结构示意图。
下面详细描述本申请的实施例,所述实施例的示例在附图中示出,其中自始至终相同或类似的标号表示相同或类似的元件或具有相同或类似功能的元件。下面通过参考附图描述的实施例是示例性的,仅用于解释本申请,而不能解释为对本申请的限制。
本技术领域技术人员可以理解,除非特意声明,这里使用的单数形式“一”、“一个”、“所述”和“该”也可包括复数形式。应该进一步理解的是,本申请的说明书中使用的措辞“包括”是指存在所述特征、整数、步骤、操作,但是并不排除存在或添加一个或多个其他特征、整数、步骤、操作。
本技术领域技术人员可以理解,除非另外定义,这里使用的所有术语(包括技术术语和科学术语),具有与本申请所属领域中的普通技术人员的一般理解相同的意义。还应该理解的是,诸如通用字典中定义的那些术语,应该被理解为具有与现有技术的上下文中的意义一致的意义,并且除非像这里一样被特定定义,否则不会用理想化或过于正式的含义来解释。
本领域技术人员应当理解,本申请所称的“应用”、“应用程序”、“应用软件”以及类似表述的概念,是业内技术人员所公知的相同概念,是指由一系列计算机指令及相关数据资源有机构造的适于电子运行的计算机软件。除非特别指定,这种命名本身不受编程语言种类、级别,也不受其赖以运行的操作系统或平台所限制。理所当然地,此类概念也不受任何形式的终端所限制。
本申请实施例提供的一种遥感影像分割方法,在其中一种实施方式中,如图1所示,包括:S100、S200、S300。
S100:获取目标图像,通过深度卷积神经网络不同尺度的各卷积层提取所述目标图像的分层特征;
S200:依次基于两不同尺度卷积层提取出的所述分层特征确定所述目标图像的目标特征;
S300:依据所述目标特征确定所述目标图像中各物体的物体类别,依 据所述物体类别对所述目标图像进行分割。
本申请中的目标图像为遥感影像,其包括卫星及航空器拍摄的一类特殊图像,本申请中遥感影像的分割主要通过深度卷积神经网络实现。在本申请提供的实施例中,将所述遥感影像输入深度卷积神经网络中,深度卷积神经网络在获取遥感影像后,对该遥感影像进行卷积,通过卷积提取其在各卷积层和池化层中的分层特征,其中分层特征包括通过不同尺度下各卷积层和池化层之后的特征。由于在神经网络结果中,同一个尺度下,具有一个池化层和一个以上的卷积层。卷积层是对输入的图片进行特征提取,池化层:对输入的特征图进行压缩,一方面使特征图变小,简化网络计算复杂度;一方面进行特征压缩,提取主要特征。在一个卷积层之后,下一个尺度的池化层对上一个尺度卷积层提取出的特征进行采样,池化层对上个上一个尺度卷积层提取出的特征做了下采样,可以得到了更小的同样数量的特征;之后通过该池化层所在尺度下的卷积层卷积,细化经过池化层之后的特征,进而可以得到更准确和/或更多的特征,因此,在后续随着卷积的深入,提取的特征会更能表达出对应的目标物,进而使得深度卷积神经网络能够更为准确的区分遥感影像中的物体类别,如区分出目标物和背景。进一步,结合前述可知,不同尺度下则会提取到不同的特征,进而可以获得分层特征。为了能够使得遥感影像中同一目标物对应的特征更多,对不同尺度下卷积层提取出的所述分层特征进行融合,具体的分层特征详见后文,在此不做赘述。在获得目标特征之后,由于目标特征包括一个目标的底层几何特征、中层纹理特征以及高层语义特征中的一种或者多种,使得该目标的特征较为丰富,则深度卷积神经网络可以综合该目标特征对目标图像中各物体进行准确的分类,之后则依据该分类对遥感影像进行分割,如区分出目标地物和背景,遥感影像分割为将同一类的物体进行采用相同的颜色进行覆盖,不同类的物体覆盖的颜色不同,实现对遥感影像中物体的区分,即是对遥感影像的分割,遥感影像的分割是通过深度卷积神经网络上采样实现空间输出,以通过该方法实现对遥感影像像素级别的分割,同时通过深度卷积神经网络实现本方法的步骤,由于深度卷积神经网络提取由浅到深的多尺度特征,并进行不同尺度的跳跃性融合,进 而可以提高遥感影像在云、强光照等干扰因素下的分割精度。
可选地,所述依次基于两不同尺度卷积层提取出的所述分层特征确定所述目标图像的目标特征,包括:
依次融合两不同尺度中最后两卷积层提取出的所述分层特征,获得融合特征;
基于所述融合特征确定所述目标图像的所述目标特征。
结合前述过程,为了能够得到同一目标多层次的特征,使得该目标的特征更为丰富,进而实现对目标的准确描述和判别(分类),依次融合两不同尺度中最后两卷积层提取出的分层特征,由于不同尺度卷积层,其提取出的特征层次不同,进而为了在深度卷积神经网络卷积按照尺度完成卷积之后,依次融合不同尺度中最后两卷积层提取出的分层特征,获得包括两个尺度的特征,即前述融合特征,在后续过程中,也可以通过包括两个尺度的融合特征确定目标图像中的目标特征,特别地,同一个目标具有前述的融合特征之后,则使得同一个目标的特征更为丰富,进而便于通过丰富的目标特征对遥感影像中的物体进行准确的描述和分类。
可选地,所述依次融合两不同尺度中最后两卷积层提取出的所述分层特征,获得融合特征,包括:
获取最后一个尺度,获取所述尺度的最后一个卷积层提取出的特征;
对所述特征进行反卷积,得到该尺度相邻且大一尺度的相邻尺度中的还原特征,所述反卷积层位于所述相邻尺度中,所述还原特征位于反卷积层中;
连接所述反卷积层和与所述相邻尺度位于同一尺度中的用于提取所述分层特征的最后一个所述卷积层,获得卷积联合体;
对所述卷积联合体执行卷积获得所述融合特征,所述卷积联合体和所述反卷积位于同一尺度中;
判断所述相邻尺度是否为所述深度卷积神经网络的第一层,若不是,获取所述相邻尺度的最后一个所述卷积层,继续所述对所述卷积层提取出的特征进行反卷积的步骤。
可选地,所述基于所述融合特征确定所述目标图像的所述目标特征, 包括:
当所述相邻尺度为所述深度卷积神经网络的第一层时,将所述融合特征确定为所述目标特征。
结合前述说明,本申请的深度卷积神经网络结构如图2所示,包括卷积层A、反卷积层B以及池化层C,在具体的特征融合过程中,在遥感影像依次经过池化层C和/或卷积层A之后,则实现了对遥感影像中各物体的特征,在完成特征提取之后,以最后一个尺度的最后一个卷积层为基准,即从该卷积层开始,向前推进进行反卷积,其中,反卷积过程在反卷积层实现。在最后一个尺度下的卷积层完成卷积之后,则对最后一个卷积层A提取出的特征在反卷积层B进行反卷积,获得该尺度前一个尺度下的卷积层,其中反卷积层中包含有该尺度提取出来的分层特征,所述反卷积层包括还原特征,即将该尺度卷积提取出的特征转换为到上一尺度,以便于两种尺度下的特征进行融合。融合过程为将最后一个尺度L下的卷积层反卷积得到的还原特征,与最后一个尺度L相邻前一个尺度M下最后一个卷积层的分层特征进行融合,即将连接最后一个尺度L下的卷积层对应的反卷积层和与该尺度相邻前一个尺度M下的最后一个卷积层,进而可以获得前述的卷积联合体,卷积联合体中则包括了前述的融合特征,由于是将反卷积层和卷积层进行连接,因此,融合特征应包括反卷积层中的还原特征和卷积层的分层特征。之后对该卷积联合体进行卷积,即经过至少一个卷积层,即对融合特征进行特征提取,将尺度M下对卷积联合体执行卷积获得特征进行反卷积,获得尺度M相邻前一个尺度N下的还原特征,然后将尺度N下的最后一个卷积层提取出的分层特征与尺度N下反卷积提取出来的还原特征进行融合,即连接尺度N下的最后一个卷积层和尺度N下的反卷积层,获得又一个卷积联合体,之后依次对该卷积联合体进行如前述的过程,进而可以得到目标特征。获取到目标特征之后,则可以依据该目标特征实现对遥感影像的分割。可选地,所述目标特征包括语义特征、几何特征、纹理特征中的一种或者多种。
在具体的实施过程如:
可选地,所述依据所述目标特征确定所述目标图像中各物体的物体类 别,包括:
获取所述目标特征与物体类别的映射关系;
基于所述映射关系确定所述目标图像中各物体的物体类别。
在获得到遥感影像目标特征,经过前述过程的多次卷积和反卷积,使得该目标特征为像素级特征,结合前述目标特征的种类,为了能够对遥感影像进行分割,获取目标特征和物体类别的映射关系,即同一个物体对应的语音特征、纹理特征、几何特征等,基于该映射关系和目标特征则可以确定遥感影像中各物体的类别,以便于后续实现如前述的遥感影像分割,结合前述可知,该遥感影像为像素级的分割。
综合前述过程可知,首先通过不同尺度的卷积和池化层获取遥感影像的分层特征,然后对最后一个卷基层进行反卷积,使之的尺度还原到上一部分卷积的尺寸。同时将上一部分的最后一个卷积层与反卷积层连接(如图2中虚线部分和C部分)。然后对联合体执行卷积获取特征,此时的特征包含了抽象程度最高的特征以及抽象程度次高的特征。之后的步骤以此类推,每次将末端卷积反卷积,然后与对应卷积层连接,再执行卷积。这样融合的原因是底层的卷积反映的是目标的边、角、点等底层特征,中间层卷积反映的是纹理等中层特征,高层卷积反映的是语义等高层特征,当一个目标的特征越丰富,描述和判别越准确。在整个网络后端将特征图映射至两个特征,此时获取的两个特征所对应的索引即为类别。在本网络中分为两类:目标地物和背景。
可选地,所述通过深度卷积神经网络不同尺度的各卷积层提取所述目标图像的分层特征,包括:
依据所述目标图像对所述深度卷积神经网络进行训练,获得图像分割模型;
通过所述图像分割模型提取所述目标图像在所述深度卷积神经网络各卷积层中的分层特征。
可选地,所述依据所述目标图像对所述深度卷积神经网络进行训练,获得图像分割模型,包括:
依据所述目标图像对所述深度卷积神经网络进行训练,确定所述深度 卷积神经网络的损失函数;
将所述损失函数输入所述深度卷积神经网络,获得所述图像分割模型。
为了实现更精准对遥感影像的分割,在获取到遥感影像之后,需要将该遥感影像输入深度卷积神经网络进行训练,以确定遥感影像的图像分割模型。其中,本申请中深度卷积神经网络的参入如下:
进而便于通过图像分割模型实现对图像的分割,即通过该图像分割模型进行如前述的过程。在图像分割模型确定过程中,主要通过目标图像进 行训练,确定深度卷积神经网络的损失函数,将该损失函数输入深度卷积神经网络之后,即可获得图像分割模型。
其中,损失函数的确定过程如下:
损失函数采用交叉熵代价函数和正则项之和。其中交叉熵函数如下所示:
其中,y为期望输出,α为神经网络实际输出,x为输入值。交叉熵损失函数可以衡量神经网络实际输出与期望输出之间的相似性,其作为损失函数在梯度下降时能避免均方误差损失函数学习速率降低的问题。同时,交叉熵作为对数函数,在接近上边界的时候,梯度值仍然较高,因此模型的收敛速度不会减慢。
正则项用以使卷积神经网络参数衰减,产生了稀疏性,减少了特征向量的,从而降低模型的复杂度,防止深度神经网络在训练过程中过拟合。综合交叉熵代价函数和正则项,损失函数为:
其中,λ为正则项的系数,θ为模型参数,k为参数个数。
训练采用的优化器是批量随机梯度下降法,为了稳定更新过程,我们引入了动量项。公式如下:
x
t+1=x
t+Δx
t
Δx
t=ρΔx
t-1-ηg
t
其中x为要学习的参数,ρ为动量项,设置为0.99,t为迭代的次数,η为学习率,设置为0.00004,g
t为x在t时刻的梯度。训练时每个批次的样本数为4。
本申请实施例还提供了一种遥感影像分割装置,在其中一种实施方式中,如图3所示,包括:目标图像获取模块100、目标特征提取模块200、分割模块300:
目标图像获取模块100,用于获取目标图像,通过深度卷积神经网络不同尺度的各卷积层提取所述目标图像的分层特征;
目标特征提取模块200,用于依次基于两不同尺度卷积层提取出的所述分层特征确定所述目标图像的目标特征;
分割模块300,用于依据所述目标特征确定所述目标图像中各物体的物体类别,依据所述物体类别对所述目标图像进行分割。
进一步地,如图3所示,本申请实施例中提供的一种遥感影像分割方法装置还包括:融合单元210,用于依次融合两不同尺度中最后两卷积层提取出的所述分层特征,获得融合特征;第一目标特征确定单元220,用于基于所述融合特征确定所述目标图像的所述目标特征。最后一个卷积层获取单元211,用于获取最后一个尺度,获取所述尺度的最后一个卷积层提取出的特征;反卷积单元212,用于对所述特征进行反卷积,得到该尺度相邻且大一尺度的相邻尺度中的还原特征,所述反卷积层位于所述相邻尺度中,所述还原特征位于反卷积层中;连接单元213,用于连接所述反卷积层和与所述相邻尺度位于同一尺度中的用于提取所述分层特征的最后一个所述卷积层,获得卷积联合体;卷积单元214,用于对所述卷积联合体执行卷积获得所述融合特征,所述卷积联合体和所述反卷积位于同一尺度中;判断单元215,用于判断所述相邻尺度是否为所述深度卷积神经网络的第一层,若不是,获取所述相邻尺度的最后一个所述卷积层,继续所述对所述卷积层提取出的特征进行反卷积的步骤。第二目标特征确定单元221,用于当所述相邻尺度为所述深度卷积神经网络的第一层时,将所述融合特征确定为所述目标特征。映射关系获取单元310,用于获取所述目标特征与物体类别的映射关系;物体类别确定单元320,用于基于所述映射关系确定所述目标图像中各物体的物体类别。分割模型获得单元110,用于依据所述目标图像对所述深度卷积神经网络进行训练,获得图像分割模型;分层特征提取单元120,用于通过所述图像分割模型提取所述目标图像在所述深度卷积神经网络各卷积层中的分层特征。损失函数确定单元111,用于依据所述目标图像对所述深度卷积神经网络进行训练,确定所述深度卷积神经网络的损失函数;图像分割模型获得单元112,用于将所述损失函数输入所述深度卷积神经网络,获得所述图像分割模型。
本申请实施例提供的一种遥感影像分割方法装置可以实现上述遥感 影像分割方法的实施例,具体功能实现请参见方法实施例中的说明,在此不再赘述。
本申请实施例提供的一种计算机可读存储介质,所述计算机可读存储介质上存储有计算机程序,该程序被处理器执行时实现任一项技术方案所述的遥感影像分割方法。其中,所述计算机可读存储介质包括但不限于任何类型的盘(包括软盘、硬盘、光盘、CD-ROM、和磁光盘)、ROM(Read-Only Memory,只读存储器)、RAM(Random Access Memory,随即存储器)、EPROM(Erasable Programmable Read-Only Memory,可擦写可编程只读存储器)、EEPROM(Electrically Erasable Programmable Read-Only Memory,电可擦可编程只读存储器)、闪存、磁性卡片或光线卡片。也就是,存储设备包括由设备(例如,计算机、手机)以能够读的形式存储或传输信息的任何介质,可以是只读存储器,磁盘或光盘等。
本申请实施例提供的一种计算机可读存储介质,存储介质可以是非易失性的,也可以是易失性的,计算机可读存储介质可实现上述遥感影像分割方法的实施例,在本申请中采用卷积神经网络提取由浅到深的多尺度特征,还通过跳跃式的方式融合深层、中层和浅层特征,用以获取高级别的语义信息和低级别的几何信息,最后通过上采样实现空间输出,以达到像素级别分割的目的,利用本申请中的卷积神经网络,可以提高遥感影像在云、强光照等干扰因素下的分割精度;本申请实施例提供的一种遥感影像分割方法,包括:获取目标图像,通过深度卷积神经网络不同尺度的各卷积层提取所述目标图像的分层特征;依次基于两不同尺度卷积层提取出的所述分层特征确定所述目标图像的目标特征;依据所述目标特征确定所述目标图像中各物体的物体类别,依据所述物体类别对所述目标图像进行分割。本申请中的目标图像为遥感影像,其包括卫星及航空器拍摄的一类特殊图像,本申请中遥感影像的分割主要通过深度卷积神经网络实现。在本申请提供的实施例中,将所述遥感影像输入深度卷积神经网络中,深度卷积神经网络在获取遥感影像后,对该遥感影像进行卷积,通过卷积提取其在各卷积层和池化层中的分层特征,其中分层特征包括通过不同尺度下各卷积层和池化层之后的特征。由于在神经网络结果中,同一个尺度下,具 有一个池化层和一个以上的卷积层。卷积层是对输入的图片进行特征提取,池化层:对输入的特征图进行压缩,一方面使特征图变小,简化网络计算复杂度;一方面进行特征压缩,提取主要特征。在一个卷积层之后,下一个尺度的池化层对上一个尺度卷积层提取出的特征进行采样,池化层对上个上一个尺度卷积层提取出的特征做了下采样,可以得到了更小的同样数量的特征;之后通过该池化层所在尺度下的卷积层卷积,细化经过池化层之后的特征,进而可以得到更准确和/或更多的特征,因此,在后续随着卷积的深入,提取的特征会更能表达出对应的目标物,进而使得深度卷积神经网络能够更为准确的区分遥感影像中的物体类别,如区分出目标物和背景。进一步,结合前述可知,不同尺度下则会提取到不同的特征,进而可以获得分层特征。为了能够使得遥感影像中同一目标物对应的特征更多,对不同尺度下卷积层提取出的所述分层特征进行融合,具体的分层特征详见后文,在此不做赘述。在获得目标特征之后,由于目标特征包括一个目标的底层几何特征、中层纹理特征以及高层语义特征中的一种或者多种,使得该目标的特征较为丰富,则深度卷积神经网络可以综合该目标特征对目标图像中各物体进行准确的分类,之后则依据该分类对遥感影像进行分割,如区分出目标地物和背景,遥感影像分割为将同一类的物体进行采用相同的颜色进行覆盖,不同类的物体覆盖的颜色不同,实现对遥感影像中物体的区分,即是对遥感影像的分割,遥感影像的分割是通过深度卷积神经网络上采样实现空间输出,以通过该方法实现对遥感影像像素级别的分割,同时通过深度卷积神经网络实现本方法的步骤,由于深度卷积神经网络提取由浅到深的多尺度特征,并进行不同尺度的跳跃性融合,进而可以提高遥感影像在云、强光照等干扰因素下的分割精度。
此外,在又一种实施例中,本申请还提供一种服务器,如图4所示,所述服务器处理器503、存储器505、输入单元507以及显示单元509等器件。本领域技术人员可以理解,图4示出的结构器件并不构成对所有服务器的限定,可以包括比图示更多或更少的部件,或者组合某些部件。存储器505可用于存储应用程序501以及各功能模块,处理器503运行存储在存储器505的应用程序501,从而执行设备的各种功能应用以及数据处 理。存储器505可以是内存储器或外存储器,或者包括内存储器和外存储器两者。内存储器可以包括只读存储器(ROM)、可编程ROM(PROM)、电可编程ROM(EPROM)、电可擦写可编程ROM(EEPROM)、快闪存储器、或者随机存储器。外存储器可以包括硬盘、软盘、ZIP盘、U盘、磁带等。本申请所公开的存储器包括但不限于这些类型的存储器。本申请所公开的存储器505只作为例子而非作为限定。
输入单元507用于接收信号的输入,以及用户输入的个人信息和相关的身体状况信息。输入单元507可包括触控面板以及其它输入设备。触控面板可收集客户在其上或附近的触摸操作(比如客户使用手指、触笔等任何适合的物体或附件在触控面板上或在触控面板附近的操作),并根据预先设定的程序驱动相应的连接装置;其它输入设备可以包括但不限于物理键盘、功能键(比如播放控制按键、开关按键等)、轨迹球、鼠标、操作杆等中的一种或多种。显示单元509可用于显示客户输入的信息或提供给客户的信息以及计算机设备的各种菜单。显示单元509可采用液晶显示器、有机发光二极管等形式。处理器503是计算机设备的控制中心,利用各种接口和线路连接整个电脑的各个部分,通过运行或执行存储在存储器503内的软件程序和/或模块,以及调用存储在存储器内的数据,执行各种功能和处理数据。图4中所示的一个或多个处理器503能够执行、实现图3中所示的目标图像获取模块100的功能、目标特征提取模块200的功能、分割模块300的功能、目标图像获取模块100的功能、目标特征提取模块200的功能、分割模块300的功能、融合单元210的功能、第一目标特征确定单元220的功能、最后一个卷积层获取单元211的功能、反卷积单元212的功能、连接单元213的功能、卷积单元214的功能、判断单元215的功能、连接单元212的功能、第二目标特征确定单元221的功能、映射关系获取单元310的功能、物体类别确定单元320的功能、分割模型获得单元110的功能、分层特征提取单元120的功能、损失函数确定单元111的功能、图像分割模型获得单元112的功能。
在一种实施方式中,所述服务器包括一个或多个处理器503,以及一个或多个存储器505,一个或多个应用程序501,其中所述一个或多个应 用程序501被存储在存储器505中并被配置为由所述一个或多个处理器503执行,所述一个或多个应用程序301配置用于执行以上实施例所述的遥感影像分割方法。
本申请实施例提供的一种服务器,可实现上述遥感影像分割方法的实施例,在本申请中采用卷积神经网络提取由浅到深的多尺度特征,还通过跳跃式的方式融合深层、中层和浅层特征,用以获取高级别的语义信息和低级别的几何信息,最后通过上采样实现空间输出,以达到像素级别分割的目的,利用本申请中的卷积神经网络,可以提高遥感影像在云、强光照等干扰因素下的分割精度;本申请实施例提供的一种遥感影像分割方法,包括:获取目标图像,通过深度卷积神经网络不同尺度的各卷积层提取所述目标图像的分层特征;依次基于两不同尺度卷积层提取出的所述分层特征确定所述目标图像的目标特征;依据所述目标特征确定所述目标图像中各物体的物体类别,依据所述物体类别对所述目标图像进行分割。本申请中的目标图像为遥感影像,其包括卫星及航空器拍摄的一类特殊图像,本申请中遥感影像的分割主要通过深度卷积神经网络实现。在本申请提供的实施例中,将所述遥感影像输入深度卷积神经网络中,深度卷积神经网络在获取遥感影像后,对该遥感影像进行卷积,通过卷积提取其在各卷积层和池化层中的分层特征,其中分层特征包括通过不同尺度下各卷积层和池化层之后的特征。由于在神经网络结果中,同一个尺度下,具有一个池化层和一个以上的卷积层。卷积层是对输入的图片进行特征提取,池化层:对输入的特征图进行压缩,一方面使特征图变小,简化网络计算复杂度;一方面进行特征压缩,提取主要特征。在一个卷积层之后,下一个尺度的池化层对上一个尺度卷积层提取出的特征进行采样,池化层对上个上一个尺度卷积层提取出的特征做了下采样,可以得到了更小的同样数量的特征;之后通过该池化层所在尺度下的卷积层卷积,细化经过池化层之后的特征,进而可以得到更准确和/或更多的特征,因此,在后续随着卷积的深入,提取的特征会更能表达出对应的目标物,进而使得深度卷积神经网络能够更为准确的区分遥感影像中的物体类别,如区分出目标物和背景。进一步,结合前述可知,不同尺度下则会提取到不同的特征,进而可以获 得分层特征。为了能够使得遥感影像中同一目标物对应的特征更多,对不同尺度下卷积层提取出的所述分层特征进行融合,具体的分层特征详见后文,在此不做赘述。在获得目标特征之后,由于目标特征包括一个目标的底层几何特征、中层纹理特征以及高层语义特征中的一种或者多种,使得该目标的特征较为丰富,则深度卷积神经网络可以综合该目标特征对目标图像中各物体进行准确的分类,之后则依据该分类对遥感影像进行分割,如区分出目标地物和背景,遥感影像分割为将同一类的物体进行采用相同的颜色进行覆盖,不同类的物体覆盖的颜色不同,实现对遥感影像中物体的区分,即是对遥感影像的分割,遥感影像的分割是通过深度卷积神经网络上采样实现空间输出,以通过该方法实现对遥感影像像素级别的分割,同时通过深度卷积神经网络实现本方法的步骤,由于深度卷积神经网络提取由浅到深的多尺度特征,并进行不同尺度的跳跃性融合,进而可以提高遥感影像在云、强光照等干扰因素下的分割精度。
本申请实施例提供的服务器可以实现上述提供的遥感影像分割方法的实施例,具体功能实现请参见方法实施例中的说明,在此不再赘述。
以上所述仅是本申请的部分实施方式,应当指出,对于本技术领域的普通技术人员来说,在不脱离本申请原理的前提下,还可以做出若干改进和润饰,这些改进和润饰也应视为本申请的保护范围。
Claims (20)
- 一种遥感影像分割方法,包括:获取目标图像,通过深度卷积神经网络不同尺度的各卷积层提取所述目标图像的分层特征;依次基于两不同尺度卷积层提取出的所述分层特征确定所述目标图像的目标特征;依据所述目标特征确定所述目标图像中各物体的物体类别,依据所述物体类别对所述目标图像进行分割。
- 根据权利要求1所述的遥感影像分割方法,所述依次基于两不同尺度卷积层提取出的所述分层特征确定所述目标图像的目标特征,包括:依次融合两不同尺度中最后两卷积层提取出的所述分层特征,获得融合特征;基于所述融合特征确定所述目标图像的所述目标特征。
- 根据权利要求2所述的遥感影像分割方法,所述依次融合两不同尺度中最后两卷积层提取出的所述分层特征,获得融合特征,包括:获取最后一个尺度,获取所述尺度的最后一个卷积层提取出的特征;对所述特征进行反卷积,得到该尺度相邻且大一尺度的相邻尺度中的还原特征,所述反卷积层位于所述相邻尺度中,所述还原特征位于反卷积层中;连接所述反卷积层和与所述相邻尺度位于同一尺度中的用于提取所述分层特征的最后一个所述卷积层,获得卷积联合体;对所述卷积联合体执行卷积获得所述融合特征,所述卷积联合体和所述反卷积位于同一尺度中;判断所述相邻尺度是否为所述深度卷积神经网络的第一层,若不是,获取所述相邻尺度的最后一个所述卷积层,继续所述对所述卷积层提取出的特征进行反卷积的步骤。
- 根据权利要求3所述的遥感影像分割方法,所述基于所述融合特征确定所述目标图像的所述目标特征,包括:当所述相邻尺度为所述深度卷积神经网络的第一层时,将所述融合特征确定为所述目标特征。
- 根据权利要求1至4任一项所述的遥感影像分割方法,所述目标特征包括语义特征、几何特征、纹理特征中的一种或者多种。
- 根据权利要求1至4任一项所述的遥感影像分割方法,所述依据所述目标特征确定所述目标图像中各物体的物体类别,包括:获取所述目标特征与物体类别的映射关系;基于所述映射关系确定所述目标图像中各物体的物体类别。
- 根据权利要求1至4任一项所述的遥感影像分割方法,所述通过深度卷积神经网络不同尺度的各卷积层提取所述目标图像的分层特征,包括:依据所述目标图像对所述深度卷积神经网络进行训练,获得图像分割模型;通过所述图像分割模型提取所述目标图像在所述深度卷积神经网络各卷积层中的分层特征。
- 根据权利要求7所述的遥感影像分割方法,所述依据所述目标图像对所述深度卷积神经网络进行训练,获得图像分割模型,包括:依据所述目标图像对所述深度卷积神经网络进行训练,确定所述深度卷积神经网络的损失函数;将所述损失函数输入所述深度卷积神经网络,获得所述图像分割模型。
- 一种计算机可读存储介质,所述计算机可读存储介质上存储有计算机程序,该程序被处理器执行时实现一种遥感影像分割方法,所述遥感影像分割方法包括:获取目标图像,通过深度卷积神经网络不同尺度的各卷积层提取所述目标图像的分层特征;依次基于两不同尺度卷积层提取出的所述分层特征确定所述目标图像的目标特征;依据所述目标特征确定所述目标图像中各物体的物体类别,依据所述 物体类别对所述目标图像进行分割。
- 根据权利要求9所述的计算机可读存储介质,所述依次基于两不同尺度卷积层提取出的所述分层特征确定所述目标图像的目标特征,包括:依次融合两不同尺度中最后两卷积层提取出的所述分层特征,获得融合特征;基于所述融合特征确定所述目标图像的所述目标特征。
- 根据权利要求10所述的计算机可读存储介质,所述依次融合两不同尺度中最后两卷积层提取出的所述分层特征,获得融合特征,包括:获取最后一个尺度,获取所述尺度的最后一个卷积层提取出的特征;对所述特征进行反卷积,得到该尺度相邻且大一尺度的相邻尺度中的还原特征,所述反卷积层位于所述相邻尺度中,所述还原特征位于反卷积层中;连接所述反卷积层和与所述相邻尺度位于同一尺度中的用于提取所述分层特征的最后一个所述卷积层,获得卷积联合体;对所述卷积联合体执行卷积获得所述融合特征,所述卷积联合体和所述反卷积位于同一尺度中;判断所述相邻尺度是否为所述深度卷积神经网络的第一层,若不是,获取所述相邻尺度的最后一个所述卷积层,继续所述对所述卷积层提取出的特征进行反卷积的步骤。
- 根据权利要求11所述的计算机可读存储介质,所述基于所述融合特征确定所述目标图像的所述目标特征,包括:当所述相邻尺度为所述深度卷积神经网络的第一层时,将所述融合特征确定为所述目标特征。
- 根据权利要求9-12任一项所述的计算机可读存储介质,所述目标特征包括语义特征、几何特征、纹理特征中的一种或者多种。
- 根据权利要求9-12任一项所述的计算机可读存储介质,所述依据所述目标特征确定所述目标图像中各物体的物体类别,包括:获取所述目标特征与物体类别的映射关系;基于所述映射关系确定所述目标图像中各物体的物体类别。
- 根据权利要求9-12任一项所述的计算机可读存储介质,所述通过深度卷积神经网络不同尺度的各卷积层提取所述目标图像的分层特征,包括:依据所述目标图像对所述深度卷积神经网络进行训练,获得图像分割模型;通过所述图像分割模型提取所述目标图像在所述深度卷积神经网络各卷积层中的分层特征。
- 根据权利要求15所述的计算机可读存储介质,所述依据所述目标图像对所述深度卷积神经网络进行训练,获得图像分割模型,包括:依据所述目标图像对所述深度卷积神经网络进行训练,确定所述深度卷积神经网络的损失函数;将所述损失函数输入所述深度卷积神经网络,获得所述图像分割模型。
- 一种服务器,包括:一个或多个处理器;存储器;一个或多个应用程序,其中所述一个或多个应用程序被存储在所述存储器中并被配置为由所述一个或多个处理器执行,所述一个或多个应用程序配置用于执行一种遥感影像分割方法,所述遥感影像分割方法包括:获取目标图像,通过深度卷积神经网络不同尺度的各卷积层提取所述目标图像的分层特征;依次基于两不同尺度卷积层提取出的所述分层特征确定所述目标图像的目标特征;依据所述目标特征确定所述目标图像中各物体的物体类别,依据所述物体类别对所述目标图像进行分割。
- 根据权利要求17所述的服务器,所述依次基于两不同尺度卷积层提取出的所述分层特征确定所述目标图像的目标特征,包括:依次融合两不同尺度中最后两卷积层提取出的所述分层特征,获得融 合特征;基于所述融合特征确定所述目标图像的所述目标特征。
- 根据权利要求18所述的服务器,所述依次融合两不同尺度中最后两卷积层提取出的所述分层特征,获得融合特征,包括:获取最后一个尺度,获取所述尺度的最后一个卷积层提取出的特征;对所述特征进行反卷积,得到该尺度相邻且大一尺度的相邻尺度中的还原特征,所述反卷积层位于所述相邻尺度中,所述还原特征位于反卷积层中;连接所述反卷积层和与所述相邻尺度位于同一尺度中的用于提取所述分层特征的最后一个所述卷积层,获得卷积联合体;对所述卷积联合体执行卷积获得所述融合特征,所述卷积联合体和所述反卷积位于同一尺度中;判断所述相邻尺度是否为所述深度卷积神经网络的第一层,若不是,获取所述相邻尺度的最后一个所述卷积层,继续所述对所述卷积层提取出的特征进行反卷积的步骤。
- 根据权利要求19所述的服务器,所述基于所述融合特征确定所述目标图像的所述目标特征,包括:当所述相邻尺度为所述深度卷积神经网络的第一层时,将所述融合特征确定为所述目标特征。
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910015344.2A CN109801293B (zh) | 2019-01-08 | 2019-01-08 | 遥感影像分割方法、装置及存储介质、服务器 |
| CN201910015344.2 | 2019-01-08 |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| WO2020143323A1 true WO2020143323A1 (zh) | 2020-07-16 |
Family
ID=66556995
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| PCT/CN2019/118530 Ceased WO2020143323A1 (zh) | 2019-01-08 | 2019-11-14 | 遥感影像分割方法、装置及存储介质、服务器 |
Country Status (2)
| Country | Link |
|---|---|
| CN (1) | CN109801293B (zh) |
| WO (1) | WO2020143323A1 (zh) |
Cited By (45)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109859209A (zh) * | 2019-01-08 | 2019-06-07 | 平安科技(深圳)有限公司 | 遥感影像分割方法、装置及存储介质、服务器 |
| CN112017178A (zh) * | 2020-09-08 | 2020-12-01 | 重庆市地理信息和遥感应用中心 | 一种基于双时相差值图像法的遥感图像区域变化检测方法 |
| CN112183672A (zh) * | 2020-11-05 | 2021-01-05 | 北京金山云网络技术有限公司 | 图像分类方法、特征提取网络的训练方法和装置 |
| CN112215243A (zh) * | 2020-10-30 | 2021-01-12 | 百度(中国)有限公司 | 图像特征提取方法、装置、设备及存储介质 |
| CN112257496A (zh) * | 2020-09-14 | 2021-01-22 | 中国电力科学研究院有限公司 | 一种基于深度学习的输电通道周围环境分类方法及系统 |
| CN112364699A (zh) * | 2020-10-14 | 2021-02-12 | 珠海欧比特宇航科技股份有限公司 | 基于加权损失融合网络的遥感图像分割方法、装置及介质 |
| CN112530011A (zh) * | 2020-12-10 | 2021-03-19 | 北京市生态环境监测中心 | 一种消纳场环境级别确定方法、装置及设备 |
| CN112800964A (zh) * | 2021-01-27 | 2021-05-14 | 中国人民解放军战略支援部队信息工程大学 | 基于多模块融合的遥感影像目标检测方法及系统 |
| CN113139552A (zh) * | 2021-04-29 | 2021-07-20 | 航天新气象科技有限公司 | 一种小麦抽穗期识别方法、装置、电子设备及存储介质 |
| CN113344871A (zh) * | 2021-05-27 | 2021-09-03 | 中国农业大学 | 农业遥感图像分析方法及系统 |
| CN113487483A (zh) * | 2021-07-05 | 2021-10-08 | 上海商汤智能科技有限公司 | 影像分割网络的训练方法和装置 |
| CN113554032A (zh) * | 2021-09-22 | 2021-10-26 | 南京信息工程大学 | 基于高度感知的多路并行网络的遥感图像分割方法 |
| CN113673369A (zh) * | 2021-07-30 | 2021-11-19 | 中国科学院自动化研究所 | 遥感图像场景规划方法、装置、电子设备和存储介质 |
| CN113763381A (zh) * | 2021-09-28 | 2021-12-07 | 北京工业大学 | 一种融合图像全局信息的对象检测方法及系统 |
| CN113762196A (zh) * | 2021-09-16 | 2021-12-07 | 哈尔滨理工大学 | 一种基于改进Dense-Net的遥感图像分类方法 |
| CN113920311A (zh) * | 2021-09-17 | 2022-01-11 | 山东锋士信息技术有限公司 | 一种基于边缘辅助信息的遥感图像分割方法及系统 |
| CN114078204A (zh) * | 2020-08-20 | 2022-02-22 | 深圳绿米联创科技有限公司 | 目标检测方法、装置、电子设备及存储介质 |
| CN114494283A (zh) * | 2021-12-24 | 2022-05-13 | 南京信息工程大学 | 一种农田自动分割方法及系统 |
| CN114529833A (zh) * | 2022-02-18 | 2022-05-24 | 中国工商银行股份有限公司 | 一种遥感图像分割方法及装置 |
| CN114639060A (zh) * | 2022-04-01 | 2022-06-17 | 中国科学技术大学先进技术研究院 | 多尺度火焰图像检测方法、装置、设备及存储介质 |
| CN114648704A (zh) * | 2022-03-17 | 2022-06-21 | 山东师范大学 | 一种农田边界高精度提取方法及系统 |
| CN114708513A (zh) * | 2022-03-04 | 2022-07-05 | 深圳市规划和自然资源数据管理中心 | 一种顾及角点特征的边缘建筑物提取方法及系统 |
| CN114764904A (zh) * | 2020-12-31 | 2022-07-19 | 上海智臻智能网络科技股份有限公司 | 交通场景的检测方法及设备 |
| CN114764905A (zh) * | 2020-12-31 | 2022-07-19 | 上海智臻智能网络科技股份有限公司 | 用于交通环境检测的方法及设备 |
| CN114913330A (zh) * | 2022-07-18 | 2022-08-16 | 中科视语(北京)科技有限公司 | 点云部件分割方法、装置、电子设备与存储介质 |
| CN115457379A (zh) * | 2022-07-01 | 2022-12-09 | 北京理工大学 | 一种联合语义分割与角度预测的遥感影像道路提取方法及系统 |
| CN115731461A (zh) * | 2022-11-04 | 2023-03-03 | 北京理工大学 | 一种多层特征解耦的光学遥感图像建筑物提取方法 |
| CN116229272A (zh) * | 2023-03-14 | 2023-06-06 | 中国人民解放军陆军军事交通学院镇江校区 | 一种基于代表点表示的高精度遥感图像检测方法及系统 |
| CN116452973A (zh) * | 2023-03-28 | 2023-07-18 | 西南交通大学 | 地理信息遥感测绘系统及其方法 |
| CN116630819A (zh) * | 2023-05-09 | 2023-08-22 | 浙江大立科技股份有限公司 | 一种像素级的卫星遥感小目标识别方法及系统 |
| CN116703744A (zh) * | 2023-04-18 | 2023-09-05 | 二十一世纪空间技术应用股份有限公司 | 一种基于卷积神经网络的遥感影像匀光匀色方法和装置 |
| CN116863352A (zh) * | 2023-09-05 | 2023-10-10 | 江西啄木蜂科技有限公司 | 一种超大幅面遥感影像异常区域检测方法 |
| CN117095299A (zh) * | 2023-10-18 | 2023-11-21 | 浙江省测绘科学技术研究院 | 破碎化耕作区的粮食作物提取方法、系统、设备及介质 |
| CN117237666A (zh) * | 2023-09-19 | 2023-12-15 | 中移(苏州)软件技术有限公司 | 多尺度边缘检测方法、装置、设备和可读存储介质 |
| CN117253156A (zh) * | 2023-11-17 | 2023-12-19 | 深圳元戎启行科技有限公司 | 基于图像分割的特征描述提取方法、装置、终端及介质 |
| CN117274808A (zh) * | 2023-09-22 | 2023-12-22 | 中国工商银行股份有限公司 | 农作物监测方法、装置、设备及存储介质 |
| CN117409339A (zh) * | 2023-10-13 | 2024-01-16 | 东南大学 | 一种用于空地协同的无人机作物状态视觉识别方法 |
| CN117876538A (zh) * | 2024-03-11 | 2024-04-12 | 成都环境工程建设有限公司 | 一种垃圾发电厂可视化管理方法及系统 |
| CN117933309A (zh) * | 2024-03-13 | 2024-04-26 | 西安理工大学 | 一种用于双时相遥感图像变化发现的三路神经网络及方法 |
| CN119579979A (zh) * | 2024-11-19 | 2025-03-07 | 北京理工大学 | 基于光学影像和合成孔径雷达影像的地物分类方法、设备、介质及产品 |
| CN119694321A (zh) * | 2025-02-27 | 2025-03-25 | 山东浪潮科学研究院有限公司 | 身份识别方法、装置、电子设备及存储介质 |
| CN120279258A (zh) * | 2025-06-09 | 2025-07-08 | 深圳大学 | 水上遇险目标遥感识别与漂移预测方法、装置及设备 |
| CN120765934A (zh) * | 2025-06-26 | 2025-10-10 | 耕宇牧星(北京)空间科技有限公司 | 一种基于多尺度语境增强的遥感图像分割方法及系统 |
| CN120808135A (zh) * | 2025-06-17 | 2025-10-17 | 浙江大学 | 一种基于深度学习的遥感影像用地超限识别方法及系统 |
| CN121170620A (zh) * | 2025-09-22 | 2025-12-19 | 中国水利水电科学研究院 | 一种基于多尺度遥感小微陆表水体精化提取方法 |
Families Citing this family (20)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109801293B (zh) * | 2019-01-08 | 2023-07-14 | 平安科技(深圳)有限公司 | 遥感影像分割方法、装置及存储介质、服务器 |
| CN110443813B (zh) * | 2019-07-29 | 2024-02-27 | 腾讯医疗健康(深圳)有限公司 | 血管、眼底图像的分割方法、装置、设备及可读存储介质 |
| CN110909195A (zh) * | 2019-10-12 | 2020-03-24 | 平安科技(深圳)有限公司 | 基于区块链的图片标注方法、装置及存储介质、服务器 |
| CN110942013A (zh) * | 2019-11-22 | 2020-03-31 | 佛山科学技术学院 | 一种基于深度神经网络的卫星影像特征提取方法及系统 |
| CN111047602A (zh) * | 2019-11-26 | 2020-04-21 | 中国科学院深圳先进技术研究院 | 图像分割方法、装置及终端设备 |
| CN111091091A (zh) * | 2019-12-16 | 2020-05-01 | 北京迈格威科技有限公司 | 目标对象重识别特征的提取方法、装置、设备及存储介质 |
| CN111563439B (zh) * | 2020-04-28 | 2023-08-08 | 京东科技信息技术有限公司 | 一种水产生物病害检测方法、装置及设备 |
| CN111798461B (zh) * | 2020-06-19 | 2022-04-01 | 武汉大学 | 粗粒度标签引导深度学习的像素级遥感影像云区检测方法 |
| CN111916050A (zh) * | 2020-08-03 | 2020-11-10 | 北京字节跳动网络技术有限公司 | 语音合成方法、装置、存储介质和电子设备 |
| CN112084865A (zh) * | 2020-08-06 | 2020-12-15 | 中国科学院空天信息创新研究院 | 目标检测方法、装置、电子设备和存储介质 |
| CN111985487B (zh) * | 2020-08-31 | 2024-03-19 | 香港中文大学(深圳) | 一种遥感影像目标提取方法、电子设备及存储介质 |
| CN112580641B (zh) * | 2020-11-23 | 2024-06-04 | 上海明略人工智能(集团)有限公司 | 图像特征的提取方法及装置、存储介质、电子设备 |
| CN112634277B (zh) * | 2020-12-30 | 2025-05-09 | 上海联影智能医疗科技有限公司 | 图像分割方法、系统、电子设备及存储介质 |
| CN113222039B (zh) * | 2021-05-24 | 2023-09-05 | 中国平安财产保险股份有限公司 | 遥感影像标的查勘模型的训练方法、装置、设备及介质 |
| CN114821048A (zh) * | 2022-04-11 | 2022-07-29 | 北京奕斯伟计算技术有限公司 | 目标物分割方法和相关装置 |
| CN114638767B (zh) * | 2022-05-18 | 2022-08-16 | 山东师范大学 | 基于生成对抗网络的腹腔镜影像除烟方法 |
| CN117333496A (zh) * | 2022-06-27 | 2024-01-02 | 成都圭目机器人有限公司 | 一种地下目标分割方法及系统 |
| CN115393591A (zh) * | 2022-08-30 | 2022-11-25 | 河南讯飞人工智能科技有限公司 | 一种对象分割方法、装置、设备及存储介质 |
| CN116664996A (zh) * | 2023-05-05 | 2023-08-29 | 中汽创智科技有限公司 | 特征融合方法、装置、计算机设备及其存储介质 |
| CN117197878B (zh) * | 2023-11-07 | 2024-03-05 | 中影年年(北京)科技有限公司 | 基于机器学习的人物面部表情捕捉方法及系统 |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105512661A (zh) * | 2015-11-25 | 2016-04-20 | 中国人民解放军信息工程大学 | 一种基于多模态特征融合的遥感影像分类方法 |
| US20180174046A1 (en) * | 2016-12-15 | 2018-06-21 | Beijing Kuangshi Technology Co., Ltd. | Target detection method and device, neural network training method and device |
| CN108460411A (zh) * | 2018-02-09 | 2018-08-28 | 北京市商汤科技开发有限公司 | 实例分割方法和装置、电子设备、程序和介质 |
| CN108830322A (zh) * | 2018-06-15 | 2018-11-16 | 联想(北京)有限公司 | 一种图像处理方法及装置、设备、存储介质 |
| CN109801293A (zh) * | 2019-01-08 | 2019-05-24 | 平安科技(深圳)有限公司 | 遥感影像分割方法、装置及存储介质、服务器 |
Family Cites Families (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107784654B (zh) * | 2016-08-26 | 2020-09-25 | 杭州海康威视数字技术股份有限公司 | 图像分割方法、装置及全卷积网络系统 |
| GB2555136A (en) * | 2016-10-21 | 2018-04-25 | Nokia Technologies Oy | A method for analysing media content |
| CN107180430A (zh) * | 2017-05-16 | 2017-09-19 | 华中科技大学 | 一种适用于语义分割的深度学习网络构建方法及系统 |
| CN107610141B (zh) * | 2017-09-05 | 2020-04-03 | 华南理工大学 | 一种基于深度学习的遥感图像语义分割方法 |
| CN108230329B (zh) * | 2017-12-18 | 2021-09-21 | 孙颖 | 基于多尺度卷积神经网络的语义分割方法 |
-
2019
- 2019-01-08 CN CN201910015344.2A patent/CN109801293B/zh active Active
- 2019-11-14 WO PCT/CN2019/118530 patent/WO2020143323A1/zh not_active Ceased
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105512661A (zh) * | 2015-11-25 | 2016-04-20 | 中国人民解放军信息工程大学 | 一种基于多模态特征融合的遥感影像分类方法 |
| US20180174046A1 (en) * | 2016-12-15 | 2018-06-21 | Beijing Kuangshi Technology Co., Ltd. | Target detection method and device, neural network training method and device |
| CN108460411A (zh) * | 2018-02-09 | 2018-08-28 | 北京市商汤科技开发有限公司 | 实例分割方法和装置、电子设备、程序和介质 |
| CN108830322A (zh) * | 2018-06-15 | 2018-11-16 | 联想(北京)有限公司 | 一种图像处理方法及装置、设备、存储介质 |
| CN109801293A (zh) * | 2019-01-08 | 2019-05-24 | 平安科技(深圳)有限公司 | 遥感影像分割方法、装置及存储介质、服务器 |
Cited By (56)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109859209A (zh) * | 2019-01-08 | 2019-06-07 | 平安科技(深圳)有限公司 | 遥感影像分割方法、装置及存储介质、服务器 |
| CN114078204A (zh) * | 2020-08-20 | 2022-02-22 | 深圳绿米联创科技有限公司 | 目标检测方法、装置、电子设备及存储介质 |
| CN114078204B (zh) * | 2020-08-20 | 2025-01-17 | 深圳绿米联创科技有限公司 | 目标检测方法、装置、电子设备及存储介质 |
| CN112017178A (zh) * | 2020-09-08 | 2020-12-01 | 重庆市地理信息和遥感应用中心 | 一种基于双时相差值图像法的遥感图像区域变化检测方法 |
| CN112257496A (zh) * | 2020-09-14 | 2021-01-22 | 中国电力科学研究院有限公司 | 一种基于深度学习的输电通道周围环境分类方法及系统 |
| CN112364699A (zh) * | 2020-10-14 | 2021-02-12 | 珠海欧比特宇航科技股份有限公司 | 基于加权损失融合网络的遥感图像分割方法、装置及介质 |
| CN112215243A (zh) * | 2020-10-30 | 2021-01-12 | 百度(中国)有限公司 | 图像特征提取方法、装置、设备及存储介质 |
| CN112183672A (zh) * | 2020-11-05 | 2021-01-05 | 北京金山云网络技术有限公司 | 图像分类方法、特征提取网络的训练方法和装置 |
| CN112530011A (zh) * | 2020-12-10 | 2021-03-19 | 北京市生态环境监测中心 | 一种消纳场环境级别确定方法、装置及设备 |
| CN114764904A (zh) * | 2020-12-31 | 2022-07-19 | 上海智臻智能网络科技股份有限公司 | 交通场景的检测方法及设备 |
| CN114764905A (zh) * | 2020-12-31 | 2022-07-19 | 上海智臻智能网络科技股份有限公司 | 用于交通环境检测的方法及设备 |
| CN112800964A (zh) * | 2021-01-27 | 2021-05-14 | 中国人民解放军战略支援部队信息工程大学 | 基于多模块融合的遥感影像目标检测方法及系统 |
| CN113139552A (zh) * | 2021-04-29 | 2021-07-20 | 航天新气象科技有限公司 | 一种小麦抽穗期识别方法、装置、电子设备及存储介质 |
| CN113344871A (zh) * | 2021-05-27 | 2021-09-03 | 中国农业大学 | 农业遥感图像分析方法及系统 |
| CN113487483A (zh) * | 2021-07-05 | 2021-10-08 | 上海商汤智能科技有限公司 | 影像分割网络的训练方法和装置 |
| CN113673369A (zh) * | 2021-07-30 | 2021-11-19 | 中国科学院自动化研究所 | 遥感图像场景规划方法、装置、电子设备和存储介质 |
| CN113762196A (zh) * | 2021-09-16 | 2021-12-07 | 哈尔滨理工大学 | 一种基于改进Dense-Net的遥感图像分类方法 |
| CN113920311A (zh) * | 2021-09-17 | 2022-01-11 | 山东锋士信息技术有限公司 | 一种基于边缘辅助信息的遥感图像分割方法及系统 |
| CN113554032B (zh) * | 2021-09-22 | 2021-12-14 | 南京信息工程大学 | 基于高度感知的多路并行网络的遥感图像分割方法 |
| CN113554032A (zh) * | 2021-09-22 | 2021-10-26 | 南京信息工程大学 | 基于高度感知的多路并行网络的遥感图像分割方法 |
| CN113763381A (zh) * | 2021-09-28 | 2021-12-07 | 北京工业大学 | 一种融合图像全局信息的对象检测方法及系统 |
| CN114494283A (zh) * | 2021-12-24 | 2022-05-13 | 南京信息工程大学 | 一种农田自动分割方法及系统 |
| CN114529833A (zh) * | 2022-02-18 | 2022-05-24 | 中国工商银行股份有限公司 | 一种遥感图像分割方法及装置 |
| CN114708513A (zh) * | 2022-03-04 | 2022-07-05 | 深圳市规划和自然资源数据管理中心 | 一种顾及角点特征的边缘建筑物提取方法及系统 |
| CN114708513B (zh) * | 2022-03-04 | 2023-12-01 | 深圳市规划和自然资源数据管理中心 | 一种顾及角点特征的边缘建筑物提取方法及系统 |
| CN114648704A (zh) * | 2022-03-17 | 2022-06-21 | 山东师范大学 | 一种农田边界高精度提取方法及系统 |
| CN114639060A (zh) * | 2022-04-01 | 2022-06-17 | 中国科学技术大学先进技术研究院 | 多尺度火焰图像检测方法、装置、设备及存储介质 |
| CN115457379A (zh) * | 2022-07-01 | 2022-12-09 | 北京理工大学 | 一种联合语义分割与角度预测的遥感影像道路提取方法及系统 |
| CN114913330A (zh) * | 2022-07-18 | 2022-08-16 | 中科视语(北京)科技有限公司 | 点云部件分割方法、装置、电子设备与存储介质 |
| CN115731461A (zh) * | 2022-11-04 | 2023-03-03 | 北京理工大学 | 一种多层特征解耦的光学遥感图像建筑物提取方法 |
| CN116229272A (zh) * | 2023-03-14 | 2023-06-06 | 中国人民解放军陆军军事交通学院镇江校区 | 一种基于代表点表示的高精度遥感图像检测方法及系统 |
| CN116229272B (zh) * | 2023-03-14 | 2023-10-31 | 中国人民解放军陆军军事交通学院镇江校区 | 一种基于代表点表示的高精度遥感图像检测方法及系统 |
| CN116452973A (zh) * | 2023-03-28 | 2023-07-18 | 西南交通大学 | 地理信息遥感测绘系统及其方法 |
| CN116703744A (zh) * | 2023-04-18 | 2023-09-05 | 二十一世纪空间技术应用股份有限公司 | 一种基于卷积神经网络的遥感影像匀光匀色方法和装置 |
| CN116703744B (zh) * | 2023-04-18 | 2024-05-28 | 二十一世纪空间技术应用股份有限公司 | 一种基于卷积神经网络的遥感影像匀光匀色方法和装置 |
| CN116630819A (zh) * | 2023-05-09 | 2023-08-22 | 浙江大立科技股份有限公司 | 一种像素级的卫星遥感小目标识别方法及系统 |
| CN116863352A (zh) * | 2023-09-05 | 2023-10-10 | 江西啄木蜂科技有限公司 | 一种超大幅面遥感影像异常区域检测方法 |
| CN116863352B (zh) * | 2023-09-05 | 2024-01-02 | 江西啄木蜂科技有限公司 | 一种超大幅面遥感影像异常区域检测方法 |
| CN117237666A (zh) * | 2023-09-19 | 2023-12-15 | 中移(苏州)软件技术有限公司 | 多尺度边缘检测方法、装置、设备和可读存储介质 |
| CN117237666B (zh) * | 2023-09-19 | 2025-11-04 | 中移(苏州)软件技术有限公司 | 多尺度边缘检测方法、装置、设备和可读存储介质 |
| CN117274808A (zh) * | 2023-09-22 | 2023-12-22 | 中国工商银行股份有限公司 | 农作物监测方法、装置、设备及存储介质 |
| CN117409339A (zh) * | 2023-10-13 | 2024-01-16 | 东南大学 | 一种用于空地协同的无人机作物状态视觉识别方法 |
| CN117095299A (zh) * | 2023-10-18 | 2023-11-21 | 浙江省测绘科学技术研究院 | 破碎化耕作区的粮食作物提取方法、系统、设备及介质 |
| CN117095299B (zh) * | 2023-10-18 | 2024-01-26 | 浙江省测绘科学技术研究院 | 破碎化耕作区的粮食作物提取方法、系统、设备及介质 |
| CN117253156B (zh) * | 2023-11-17 | 2024-03-29 | 深圳元戎启行科技有限公司 | 基于图像分割的特征描述提取方法、装置、终端及介质 |
| CN117253156A (zh) * | 2023-11-17 | 2023-12-19 | 深圳元戎启行科技有限公司 | 基于图像分割的特征描述提取方法、装置、终端及介质 |
| CN117876538A (zh) * | 2024-03-11 | 2024-04-12 | 成都环境工程建设有限公司 | 一种垃圾发电厂可视化管理方法及系统 |
| CN117876538B (zh) * | 2024-03-11 | 2024-05-31 | 成都环境工程建设有限公司 | 一种垃圾发电厂可视化管理方法及系统 |
| CN117933309A (zh) * | 2024-03-13 | 2024-04-26 | 西安理工大学 | 一种用于双时相遥感图像变化发现的三路神经网络及方法 |
| CN119579979A (zh) * | 2024-11-19 | 2025-03-07 | 北京理工大学 | 基于光学影像和合成孔径雷达影像的地物分类方法、设备、介质及产品 |
| CN119694321A (zh) * | 2025-02-27 | 2025-03-25 | 山东浪潮科学研究院有限公司 | 身份识别方法、装置、电子设备及存储介质 |
| CN120279258A (zh) * | 2025-06-09 | 2025-07-08 | 深圳大学 | 水上遇险目标遥感识别与漂移预测方法、装置及设备 |
| CN120279258B (zh) * | 2025-06-09 | 2025-08-12 | 深圳大学 | 水上遇险目标遥感识别与漂移预测方法、装置及设备 |
| CN120808135A (zh) * | 2025-06-17 | 2025-10-17 | 浙江大学 | 一种基于深度学习的遥感影像用地超限识别方法及系统 |
| CN120765934A (zh) * | 2025-06-26 | 2025-10-10 | 耕宇牧星(北京)空间科技有限公司 | 一种基于多尺度语境增强的遥感图像分割方法及系统 |
| CN121170620A (zh) * | 2025-09-22 | 2025-12-19 | 中国水利水电科学研究院 | 一种基于多尺度遥感小微陆表水体精化提取方法 |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109801293A (zh) | 2019-05-24 |
| CN109801293B (zh) | 2023-07-14 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| WO2020143323A1 (zh) | 遥感影像分割方法、装置及存储介质、服务器 | |
| Zhou et al. | BOMSC-Net: Boundary optimization and multi-scale context awareness based building extraction from high-resolution remote sensing imagery | |
| US9152870B2 (en) | Computer vision as a service | |
| JP7265034B2 (ja) | 人体検出用の方法及び装置 | |
| CN109859209B (zh) | 遥感影像分割方法、装置及存储介质、服务器 | |
| WO2022193335A1 (zh) | 点云数据处理方法、装置、计算机设备和存储介质 | |
| WO2020098076A1 (zh) | 跟踪目标的定位方法、装置、设备及存储介质 | |
| CN118691815A (zh) | 一种基于sam大模型微调的遥感图像高质量自动实例分割方法 | |
| US20230418861A1 (en) | Generating embeddings for text and image queries within a common embedding space for visual-text image searches | |
| CN111860398A (zh) | 遥感图像目标检测方法、系统及终端设备 | |
| WO2021203865A1 (zh) | 分子结合位点检测方法、装置、电子设备及存储介质 | |
| CN111160288A (zh) | 手势关键点检测方法、装置、计算机设备和存储介质 | |
| EP4443396A1 (en) | Image encoder training method and apparatus, device, and medium | |
| Sun et al. | ShuffleNetv2-YOLOv3: a real-time recognition method of static sign language based on a lightweight network | |
| US20240004924A1 (en) | Retrieving digital images in response to search queries for search-driven image editing | |
| US20230419571A1 (en) | Generating unified embeddings from multi-modal canvas inputs for image retrieval | |
| Wang et al. | Global contextual guided residual attention network for salient object detection | |
| CN117593762A (zh) | 一种融合视觉和压力的人体姿态估计方法、设备及介质 | |
| Gao et al. | Traffic sign detection based on ssd | |
| CN119648913A (zh) | 一种街道场景的重建方法、电子设备及存储介质 | |
| Yang et al. | Progressive depth decoupling and modulating for flexible depth completion | |
| CN113610856A (zh) | 训练图像分割模型和图像分割的方法和装置 | |
| Li et al. | Semantic segmentation of remote sensing image based on bilateral branch network | |
| CN121330274A (zh) | 基于解耦原型的小样本目标检测方法、系统、设备和介质 | |
| CN116416507A (zh) | 多目标图像检测方法、装置、计算机设备及介质 |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| 121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 19908526 Country of ref document: EP Kind code of ref document: A1 |
|
| NENP | Non-entry into the national phase |
Ref country code: DE |
|
| 122 | Ep: pct application non-entry in european phase |
Ref document number: 19908526 Country of ref document: EP Kind code of ref document: A1 |
