WO2019129032A1 - 遥感图像识别方法、装置、存储介质以及电子设备 - Google Patents

遥感图像识别方法、装置、存储介质以及电子设备 Download PDF

Info

Publication number
WO2019129032A1
WO2019129032A1 PCT/CN2018/123807 CN2018123807W WO2019129032A1 WO 2019129032 A1 WO2019129032 A1 WO 2019129032A1 CN 2018123807 W CN2018123807 W CN 2018123807W WO 2019129032 A1 WO2019129032 A1 WO 2019129032A1
Authority
WO
WIPO (PCT)
Prior art keywords
remote sensing
sensing image
image block
pixel
probability information
Prior art date
Application number
PCT/CN2018/123807
Other languages
English (en)
French (fr)
Inventor
李聪
Original Assignee
北京市商汤科技开发有限公司
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 北京市商汤科技开发有限公司 filed Critical 北京市商汤科技开发有限公司
Priority to KR1020207021508A priority Critical patent/KR20200106909A/ko
Priority to SG11202006035RA priority patent/SG11202006035RA/en
Priority to JP2020535252A priority patent/JP7080978B2/ja
Publication of WO2019129032A1 publication Critical patent/WO2019129032A1/zh
Priority to US16/909,291 priority patent/US11074445B2/en

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N7/00—Computing arrangements based on specific mathematical models
    • G06N7/01—Probabilistic graphical models, e.g. probabilistic networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
    • G06V10/443—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components by matching or filtering
    • G06V10/449—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters
    • G06V10/451—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters with interaction between the filter responses, e.g. cortical complex cells
    • G06V10/454—Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/10—Terrestrial scenes
    • G06V20/13—Satellite images

Definitions

  • the present application relates to computer vision technology, and in particular to a remote sensing image recognition method, a remote sensing image recognition device, a computer readable storage medium, and an electronic device.
  • the prediction result of one pixel is usually determined by other pixels in a certain area around the pixel, and this area may be called the receptive field of the convolutional neural network neurons.
  • the receptive field of convolutional neural network neurons is usually limited (such as a few hundred pixels), while the size of remote sensing images is usually large (such as the length and width of remote sensing images can reach thousands or even tens of thousands of pixels), therefore, neurons are limited. Receptive fields may not be able to obtain enough environmental information, which often leads to false recognition of pixels in remote sensing images by convolutional neural networks.
  • the embodiment of the present application provides a technical solution for remote sensing image recognition.
  • a remote sensing image recognition method which mainly includes: performing a resolution reduction process on a remote sensing image to be obtained, and obtaining a remote sensing image after the resolution reduction; Separating at least one remote sensing image block from the processed remote sensing image; determining a remote sensing image block to be processed from the at least one remote sensing image block; obtaining classification probability information of pixels in the to-be-processed remote sensing image block via a neural network; Determining, according to the classification probability information of the pixels in the remote sensing image block to be processed, the recognition result of the remote sensing image to be identified.
  • a remote sensing image recognition apparatus includes: a resolution reduction module, configured to perform a resolution reduction process on the remote sensing image to be obtained, and obtain a remote sensing image after the resolution reduction processing a remote sensing image segmentation module, configured to slice at least one remote sensing image block from the reduced resolution processed remote sensing image; and select an image block module to determine a remote sensing to be processed from the at least one remote sensing image block a categorization processing module, configured to obtain, by using a neural network, classification probability information of pixels in the to-be-processed remote sensing image block; and determining a recognition result module, configured to perform classification probability information according to pixels in the to-be-processed remote sensing image block Determining the recognition result of the remote sensing image to be identified.
  • a resolution reduction module configured to perform a resolution reduction process on the remote sensing image to be obtained, and obtain a remote sensing image after the resolution reduction processing
  • a remote sensing image segmentation module configured to slice at least one remote sensing image block from the
  • an electronic device comprising: a memory for storing a computer program; a processor for executing a computer program stored in the memory, and when the computer program is executed The method of any of the embodiments of the present application is implemented.
  • a computer readable storage medium having stored thereon a computer program, which when executed by a processor, implements the method of any of the embodiments of the present application.
  • a computer program is provided that, when executed by a processor in a device, implements the method of any of the embodiments of the present application.
  • the present application selects a remote sensing image block from the remote sensing image after the resolution reduction processing, and inputs the neural network. It avoids the phenomenon that all the remote sensing image blocks that are segmented are input into the neural network after each resolution reduction, and the calculation of the neural network is large.
  • FIG. 1 is a schematic flow chart of an embodiment of a remote sensing image recognition method according to the present application.
  • FIG. 2 is a schematic flow chart of another embodiment of a remote sensing image recognition method according to the present application.
  • FIG. 3 is a schematic flow chart of an embodiment of selecting a remote sensing image block according to a current confidence of a pixel of a remote sensing image to be identified according to the present application;
  • FIG. 4 is a schematic flowchart of an embodiment of maintaining current classification probability information of at least one pixel of a remote sensing image to be recognized according to the present application;
  • FIG. 5 is a schematic diagram of an embodiment of a remote sensing image recognition apparatus according to the present application.
  • FIG. 6 is a block diagram of an exemplary device that implements an embodiment of the present application.
  • FIG. 7 is a schematic diagram of an application scenario of the present application.
  • Embodiments of the present application can be applied to electronic devices such as terminal devices, computer systems, servers, etc., which can operate with numerous other general purpose or special purpose computing system environments or configurations.
  • Examples of well-known computing systems, environments, and/or configurations suitable for use with electronic devices such as terminal devices, computer systems, servers, and the like include, but are not limited to, personal computer systems, server computer systems, thin clients, thick clients, handhelds Or laptop devices, microprocessor-based systems, set-top boxes, programmable consumer electronics, networked personal computers, small computer systems, mainframe computer systems, and distributed cloud computing technology environments including any of the above, and the like.
  • Electronic devices such as terminal devices, computer systems, servers, etc., can be described in the general context of computer system executable instructions (such as program modules) being executed by a computer system.
  • program modules may include routines, programs, target programs, components, logic, and data structures, etc., which perform particular tasks or implement particular abstract data types.
  • the computer system/server can be implemented in a distributed cloud computing environment where tasks are performed by remote processing devices that are linked through a communication network.
  • program modules may be located on a local or remote computing system storage medium including storage devices.
  • the technical solution for realizing remote sensing image recognition may be a single chip microcomputer, an FPGA (Field Programmable Gate Array), a microprocessor, a smart mobile phone, a notebook computer, a tablet computer, a desktop computer or a server.
  • An electronic device capable of running a computer program (which may also be referred to as program code) may be implemented, and the computer program may be stored in a computer readable storage medium such as a flash memory, a cache, a hard disk, or an optical disk.
  • FIG. 1 is a schematic flow chart of an embodiment of a remote sensing image recognition method according to the present application.
  • the remote sensing image recognition method of the present embodiment mainly includes an operation S100, an operation S110, an operation S120, an operation S130, and an operation S140. The respective operations of the present application will be described in detail below.
  • the present application may perform a resolution reduction process on the remote sensing image in a plurality of manners.
  • the present application is preset with a reduced scale set, and the reduced scale set includes at least two different reduced values.
  • An unreduced reduction value may be selected from the reduced scale set, and the remote sensing image is subjected to a resolution reduction process according to the selected reduced value.
  • the present application when selecting a reduced value that has not been used, the present application usually preferentially selects the smallest reduced value that has not been used.
  • the present application can also perform resolution reduction processing on the remote sensing image to be recognized by means of pooling processing or downsampling processing. The present application does not limit the implementation of the resolution reduction processing of the remote sensing image to be recognized.
  • S100 may be executed by a processor invoking a corresponding instruction stored in a memory, or may be performed by a down-resolution module 500 that is executed by the processor.
  • the present application generally divides the remote sensing image obtained by reducing the resolution processing into a plurality of remote sensing image blocks according to the size requirement of the neural network, and the size of all the remote sensing image blocks that are sliced is usually It's exactly the same.
  • Each remote sensing image block that is segmented corresponds to one region in the remote sensing image to be identified, and the corresponding regions of different remote sensing image blocks are different.
  • the size of the region corresponding to the remote sensing image block is usually related to the reduction value and the size requirement of the neural network for the input image; in one example, the size of the input image in the neural network is 561 ⁇ 561, and the resolution is reduced by two.
  • the size of the remote sensing image block corresponding to the remote sensing image to be identified in the present application is 1122 ⁇ 1122.
  • the first remote sensing image block that is segmented corresponds to the remote sensing image to be identified.
  • the last remote sensing image block that is segmented corresponds to the area of the lower right corner 1122 ⁇ 1122 in the remote sensing image to be identified.
  • the application does not limit the number of segmented remote sensing image blocks and the size of the remote sensing image block.
  • S101 may be executed by a processor invoking a corresponding instruction stored in a memory, or may be performed by a remote sensing image segmentation module 510 that is executed by the processor.
  • the present application can treat all the remote sensing image blocks as the remote sensing image blocks to be processed, and not in the case of performing the resolution reduction processing for the first time.
  • the number of remote sensing image blocks to be processed determined by the present application is usually smaller than the number of all remote sensing image blocks currently segmented.
  • the present application may select the at least one remote sensing image block that is segmented according to the current confidence of the pixel of the remote sensing image to be identified, and select The remote sensing image block is the remote sensing image block to be processed.
  • the present application first determines a current confidence of at least one pixel according to current classification probability information of at least one pixel in the remote sensing image to be identified, for example, calculating a maximum of the pixel for a pixel in the remote sensing image to be identified.
  • the ratio of the classification probability value to the sub-classification probability value, and the calculated ratio is used as the current confidence of the pixel, according to which the current confidence of all pixels in the remote sensing image to be identified can be obtained. Then, the number of pixels whose current confidence level of the pixel in the remote sensing image region to be identified corresponding to the remote sensing image block reaches a predetermined confidence level is counted, and the remote sensing image block whose number of pixels reaching the predetermined confidence level does not meet the predetermined requirement is determined as the remote sensing to be processed.
  • the remote sensing image block determines whether the ratio of the number of pixels reaching the predetermined confidence level to the number of all pixels in the remote sensing image region to be recognized corresponding to the remote sensing image block is less than a predetermined ratio. If the ratio of the number of pixels reaching the predetermined confidence level to the number of all pixels in the remote sensing image region to be recognized corresponding to the remote sensing image block is less than a predetermined ratio, The remote sensing image block is determined as a remote sensing image block to be processed.
  • the present application can obtain current classification probability information of at least one pixel of the remote sensing image to be identified by using the following two methods:
  • the classification probability information of at least one pixel in each remote sensing image block to be processed output by the neural network is mapped into the remote sensing image to be recognized, and the remote sensing image to be recognized is obtained.
  • Classification probability information for at least one pixel may use the classification probability information of at least one pixel of the currently obtained remote sensing image to be recognized as the current classification probability information of at least one pixel of the remote sensing image to be identified.
  • the present application obtains current classification probability information of at least one pixel of the remote sensing image to be identified.
  • S102 may be executed by a processor invoking a corresponding instruction stored in a memory, or may be performed by a selected image block module 520 executed by the processor.
  • the neural network in the present application may be a convolutional neural network, and the present application does not limit the representation of the neural network.
  • the present invention reduces the resolution processing of the remote sensing image to be recognized, and inputs the image block to be processed in the remote sensing image after the resolution reduction processing into the neural network, thereby improving the receptive field of the neuron in the neural network, thereby It is helpful to avoid the misidentification of the pixels of the remote sensing image by the neural network.
  • the neural network in the present application is trained using remote sensing image samples having a plurality of different resolutions.
  • the resolution of each of the remote sensing image samples belongs to a resolution corresponding to each of the reduced values in the reduced scale set.
  • the scaling set includes: a first reduced value, a second reduced value, and a third reduced value, the first reduced value corresponds to the first resolution, the second reduced value corresponds to the second resolution, and the third reduced value corresponds to the first
  • the present application can use the remote sensing image samples having the first resolution, the remote sensing image samples having the second resolution, and the remote sensing image samples having the third resolution to train the trained neural network, and successfully training.
  • Neural networks can be used for remote sensing image recognition of the present application.
  • S103 may be executed by a processor invoking a corresponding instruction stored in a memory, or may be performed by a classification processing module 530 executed by the processor.
  • the application may perform the remote sensing image block according to the to-be-processed image block in response to the non-existing reduced value in the reduced scale set or the undetectable remote sensing image block from the at least one remote sensing image block.
  • the operation of the classification probability information of the pixel in the determination of the recognition result of the remote sensing image to be recognized for example, for a pixel in the remote sensing image to be identified, determining the probability value belonging to each category in the current classification probability information of the pixel
  • the maximum probability value, the category corresponding to the maximum probability value and the maximum probability value is used as the recognition result of the pixel.
  • S104 may be executed by a processor invoking a corresponding instruction stored in a memory, or may be performed by a determination recognition result module 540 that is executed by the processor.
  • FIG. 2 is a schematic flow chart of another embodiment of a remote sensing image recognition method according to the present application.
  • the remote sensing image recognition method of the present embodiment mainly includes an operation S200, an operation S210, an operation S220, and an operation S230.
  • S200 Perform a resolution reduction process on the remote sensing image to be recognized according to the minimum reduced value that is not used in the reduced ratio set.
  • the reduced scale set is formed by a plurality of reduced values, each of which has a value greater than zero and no more than one.
  • the reduced value, 0.25 means that the resolution of the remote sensing image to be identified is reduced to a quarter of the original resolution
  • 0.5 means that the resolution of the remote sensing image to be identified is reduced to one-half of the original resolution
  • 1 represents the remote sensing to be recognized. The resolution of the image.
  • This embodiment takes 1 as a special reduction value, for example, it can be considered as a reduction value that keeps the remote sensing image to be recognized at the original resolution.
  • the present application does not limit the magnitude of the reduction value included in the reduced scale set and the number of reduced values included.
  • the present application should select a reduction value from the reduced scale set each time according to the arrangement order of the reduction values from small to large, and perform resolution reduction processing on the remote sensing image according to the selected reduced value.
  • the selected reduction value becomes the used reduction value.
  • S200 may be executed by a processor invoking a corresponding instruction stored in a memory, or may be performed by a down-resolution module 500 that is executed by the processor.
  • the remote sensing image after the resolution reduction processing is divided into at least one remote sensing image block.
  • the present application generally divides the remote sensing image obtained by reducing the resolution processing into a plurality of remote sensing image blocks according to the size requirement of the neural network, and the size of all the remote sensing image blocks that are sliced is usually It's exactly the same.
  • Each remote sensing image block that is segmented corresponds to one region in the remote sensing image to be identified, and the corresponding regions of different remote sensing image blocks are different.
  • the size of the region corresponding to the remote sensing image block is usually related to the reduction value and the size requirement of the neural network for the input image; in one example, the size of the input image in the neural network is 561 ⁇ 561, and the resolution is reduced by two.
  • the size of the remote sensing image block corresponding to the remote sensing image to be identified in the present application is 1122 ⁇ 1122.
  • the first remote sensing image block that is segmented corresponds to the remote sensing image to be identified.
  • the last remote sensing image block that is segmented corresponds to the area of the lower right corner 1122 ⁇ 1122 in the remote sensing image to be identified.
  • the application does not limit the number of segmented remote sensing image blocks and the size of the remote sensing image block.
  • S201 may be executed by a processor invoking a corresponding instruction stored in a memory, or may be performed by a remote sensing image segmentation module 510 that is executed by the processor.
  • S220 Select a remote sensing image block from the at least one remote sensing image block according to a current confidence of the pixel of the remote sensing image to be identified, and obtain, by using a neural network, classification probability information of at least one pixel in each selected remote sensing image block.
  • the selected remote sensing image block is the above-mentioned remote sensing image block to be processed.
  • the present application in the case of performing the resolution reduction processing for the first time, generally uses all of the segmented remote sensing image blocks as the selected remote sensing image blocks.
  • the present application can be considered that in the case of performing the resolution reduction processing for the first time, the current confidence of the pixel of the remote sensing image to be recognized is initialized to 0, and therefore, all the remote sensing image blocks that are segmented are selected, and respectively Input into the neural network, so that the classification probability information of all the pixels of all the remote sensing image blocks that are cut out can be obtained via the neural network, and the classification probability information of all the pixels of all the remote sensing image blocks obtained this time can be used for updating The current confidence of the pixels of the remote sensing image is identified.
  • the neural network of the present application can be classified for C categories. After each remote sensing image block is respectively transmitted via a neural network, the present application can obtain all pixels in all remote sensing image blocks for C respectively. Classification probability information for the category.
  • the present application generally selects all the remote sensing image blocks that are cut out according to the current confidence of the pixels of the remote sensing image to be identified.
  • Remote sensing image block It should be specially noted that the remote sensing image blocks selected this time may be all the remote sensing image blocks that have been segmented this time, and it is also possible to partially segment the remote sensing image blocks in all the remote sensing image blocks; It is also possible that no remote sensing image block can be selected from all the remote sensing image blocks that have been cut out this time.
  • the present application reduces the resolution of the remote sensing image to be recognized, and inputs the image block in the remote sensing image after the resolution reduction into a neural network (such as a convolutional neural network), which is beneficial to improving the neural network (such as a convolutional neural network).
  • a neural network such as a convolutional neural network
  • the receptive field of the neuron is beneficial to avoid the misidentification of the pixels of the remote sensing image by the neural network.
  • S202 may be executed by a processor invoking a corresponding instruction stored in a memory, or may be performed by a selected image block module 520 and a classification processing module 530 that are executed by the processor.
  • FIG. 3 An optional example of selecting a remote sensing image block (ie, a remote sensing image block to be processed) according to the current confidence of the pixel of the remote sensing image to be identified is shown in FIG. 3 .
  • 300 determining a current confidence level of at least one pixel according to current classification probability information of at least one pixel of the remote sensing image to be identified.
  • the current classification probability information of at least one pixel of the remote sensing image to be identified in the present application is based on a classification of at least one pixel in the remote sensing image block output by the neural network for at least one remote sensing image block input in the past. Formed by probability information.
  • the present application needs to maintain the current classification probability information of at least one pixel of the remote sensing image (such as updating the current classification probability information of at least one pixel of the remote sensing image to be recognized).
  • the manner in which the present application maintains the current classification probability information of at least one pixel of the remote sensing image can be referred to the following description for FIG. 4.
  • the present application determines the pixel according to the current classification probability information of the pixel (i, j) of the remote sensing image to be identified (i, j).
  • the current confidence can be set by: the neural network can identify C categories from the remote sensing image, and the current classification probability information of the pixel (i, j) will include C probability values, which can be from this C Among the probability values, the maximum probability value and the second largest probability value (ie, the second largest probability value) are selected, and the ratio of the selected maximum classification probability value to the second largest classification probability value is calculated, and the ratio may be used as the ratio in the present application.
  • the current confidence of the pixel (i, j) the maximum probability value and the second largest probability value (ie, the second largest probability value) are selected, and the ratio of the selected maximum classification probability value to the second largest classification probability value is calculated, and the ratio may be used as the ratio in the present application.
  • the present application may also determine the current confidence of at least one pixel of the remote sensing image to be identified according to the current classification probability information of the at least one pixel of the remote sensing image to be identified.
  • the present application does not limit the current determination of at least one pixel of the remote sensing image to be identified. The way to confidence.
  • S300 may be executed by a processor invoking a corresponding instruction stored in a memory, or may be performed by a first unit that is executed by the processor.
  • the predetermined confidence level in the present application may be set according to actual needs, for example, the predetermined confidence may be set to 4.
  • the present application can respectively determine whether the current confidence of 1122 ⁇ 1122 pixels is reached. The judgment of the confidence is predetermined, so that the number of pixels reaching the predetermined confidence can be counted.
  • S301 may be executed by a processor invoking a corresponding instruction stored in a memory or by a second unit being executed by the processor.
  • a remote sensing image block that does not meet the predetermined requirement for the number of pixels with a predetermined degree of confidence is used as the selected remote sensing image block.
  • the remote sensing image block may be input to the neural network as the selected remote sensing image block, thereby obtaining classification probability information of at least one pixel in the remote sensing image block via the neural network; if the ratio reaches a predetermined ratio , the remote sensing image block does not need to be input into the neural network.
  • the predetermined ratio in the present application may be set according to actual needs, for example, the predetermined ratio may be set to 0.97 or the like.
  • the present application may also use other methods to determine a remote sensing image block that does not meet the predetermined requirement, for example, determining the number of pixels that reach a predetermined confidence level in the remote sensing image region to be identified corresponding to the remote sensing image block and the region. If the difference between the number of pixels that have not reached the predetermined confidence reaches a predetermined difference, the remote sensing image block may be input to the neural network as the selected remote sensing image block. Otherwise, the remote sensing image block is not required.
  • the remote sensing image block Inputting into the neural network; for example, determining that the ratio of the number of pixels reaching the predetermined confidence level in the remote sensing image region to be identified corresponding to the remote sensing image block to the number of pixels in the region that does not reach the predetermined confidence level does not reach the predetermined value In the case of the ratio, the remote sensing image block can be input into the neural network as the selected remote sensing image block. Otherwise, the remote sensing image block does not need to be input into the neural network.
  • the present application does not limit the implementation of remote sensing image blocks that do not meet the predetermined requirements.
  • S302 may be executed by a processor invoking a corresponding instruction stored in a memory, or may be performed by a third unit that is executed by the processor.
  • FIG. 1 An example of maintaining the current classification probability information of at least one pixel of the remote sensing image to be recognized in the present application is shown in FIG.
  • the remote sensing image recognition method of the present application begins.
  • the present application may initialize current classification probability information of at least one pixel of the remote sensing image to be recognized when the remote sensing image recognition method starts, such as setting current classification probability information of at least one pixel of the remote sensing image to be identified to 0.
  • current classification probability information of at least one pixel of the remote sensing image to be recognized it is completely feasible not to perform an initialization operation on the current classification probability information of at least one pixel of the remote sensing image to be recognized.
  • the execution of the operation S440 is triggered.
  • any of the remote sensing images to be identified is obtained.
  • Classification probability information of one pixel (i, j) It can be expressed as:
  • the upper corner mark 1 indicates the first execution of the resolution reduction process
  • C indicates the number of categories in which the neural network performs classification. It is indicated that, in the case of performing the resolution reduction processing for the first time, any pixel (i, j) in the remote sensing image to be identified belongs to the classification probability information of the first category, It is indicated that, in the case where the resolution reduction processing is performed for the first time, any pixel (i, j) in the remote sensing image to be identified belongs to the classification probability information of the second category, It is indicated that in the case where the resolution reduction processing is performed for the first time, any pixel (i, j) in the remote sensing image to be identified belongs to the classification probability information of the category C.
  • the classification probability information of at least one pixel of the remote sensing image to be identified obtained this time is used as the current classification probability information of the corresponding pixel of the remote sensing image to be identified, for example, the current of any pixel (i, j) of the remote sensing image to be identified.
  • S420 and S430 may be executed by a processor invoking a corresponding instruction stored in a memory, or may be performed by a fourth unit executed by the processor.
  • S450 Mapping classification probability information of at least one pixel in each remote sensing image block output by the neural network to the remote sensing image to be identified, thereby obtaining classification probability information of at least one pixel of the remote sensing image to be identified. Go to operation S460.
  • the classification probability information of all the pixels in each remote sensing image block output by the neural network is obtained, and the to-be-identified is obtained.
  • Classification probability information of any pixel (i, j) in remote sensing image It can be expressed as:
  • n indicates the nth execution of the resolution reduction process
  • C indicates the number of categories in which the neural network performs classification. Representing that in the case of performing the down resolution processing for the nth time, any pixel (i, j) in the remote sensing image to be identified belongs to the classification probability information of the first category, Representing that in the case of performing the resolution reduction processing for the nth time, any pixel (i, j) in the remote sensing image to be identified belongs to the classification probability information of the second category, It is indicated that in the case where the resolution reduction processing is performed for the nth time, any pixel (i, j) in the remote sensing image to be identified belongs to the classification probability information of the Cth category.
  • the present application can Set to 0.
  • S450 may be executed by a processor invoking a corresponding instruction stored in a memory, or may be performed by a fifth unit that is executed by the processor.
  • the present application is to classify all the pixels in each of the to-be-processed remote sensing image blocks respectively obtained by accumulating the previous times. Probability information to obtain current classification probability information of all pixels of the remote sensing image to be identified.
  • the classification probability information of any pixel (i, j) in the remote sensing image to be recognized obtained by the present application is set in the case where the resolution processing is performed for the first time.
  • the classification probability information of any pixel (i, j) in the remote sensing image to be recognized obtained by the present application based on the neural network may be expressed as Wherein, if the pixel (i, j) in the remote sensing image to be identified belongs to the pixel of the region in the remote sensing image to be recognized by the remote sensing image block selected by the second resolution reduction process, The classification probability information of the corresponding pixel outputted by the neural network, and if the pixel (i, j) in the remote sensing image to be identified does not belong to the second time resolution processing process, the remote sensing image block is mapped to the remote sensing image to be recognized.
  • the classification probability information of any pixel (i, j) in the remote sensing image to be recognized obtained based on the neural network may be expressed as Wherein, if the pixel (i, j) in the remote sensing image to be identified belongs to the pixel of the region in the remote sensing image to be recognized by the remote sensing image block selected by the third resolution reduction processing, The classification probability information of the corresponding pixel outputted by the neural network, and if the pixel (i, j) in the remote sensing image to be identified does not belong to the third time resolution processing process, the remote sensing image block is mapped to the remote sensing image to be recognized.
  • the process of updating the current classification probability information of all pixels of the remote sensing image to be recognized is actually performing the previous classification probability information. Fusion, for example, The fusion is performed to form current classification probability information of all pixels of the updated remote sensing image to be identified.
  • the current classification probability information of all pixels of the remote sensing image to be identified is the result of fusion of the previous classification probability information, for example, The result of the fusion.
  • S460 may be executed by a processor invoking a corresponding instruction stored in a memory, or may be performed by a sixth unit executed by the processor.
  • S230 may be executed by a processor invoking a corresponding instruction stored in a memory, or may be performed by a determination recognition result module 540 that is executed by the processor.
  • the remote sensing image samples used are generally remote sensing image samples including multiple resolutions, and the number of different resolutions is usually smaller than the reduced value in the reduced ratio set.
  • the number of correlations is, for example, in the case where the reduced scale set includes three kinds of reduced values, the sample data set should include at least the resolution of the remote sensing image samples corresponding to the three reduced values, and each of the remote sensing image samples has the labeled information (such as mask marking information, etc.).
  • the number of remotely sensed image samples of different resolutions is approximately the same in the sample data set.
  • the present application trains the neural network by using remote sensing image samples with different resolutions in the sample data set, which is beneficial to improving the accuracy of the neural network outputting the classification probability information for the input remote sensing image block to be processed, thereby facilitating the improvement of the neural network. Accuracy of remote sensing image recognition.
  • Remote sensing images are usually different from photos or pictures or video frames in the traditional sense.
  • the size of remote sensing images is usually large. For example, the length and width of remote sensing images can reach several thousand or even tens of thousands of pixels.
  • the remote sensing image is identified and processed.
  • the reliability of the recognition result may be adversely affected due to the limited receptive field of the neuron.
  • the computational cost and time cost of the neural network may be higher.
  • the present invention performs the resolution reduction processing on the remote sensing image to be identified, and the neural network processes the remote sensing image after the resolution reduction processing, which is beneficial to reducing the calculation cost and the time cost of the neural network.
  • the present invention selects a remote sensing image block from the remote sensing image after the resolution reduction processing and inputs the neural network, which is beneficial to reducing the calculation amount of the neural network for identifying the corresponding local region in the remote sensing image after the resolution reduction processing. Since the present application can select a remote sensing image block according to the current confidence of the pixels in the remote sensing image to be identified during the resolution reduction process, the neural network is used for some easily recognizable image regions in the remote sensing image to be identified. The accuracy of the recognition is high. Therefore, after each resolution reduction, the neural network is repeatedly classified for such image regions, and the contribution to the accuracy of the final classification result of the image region is not prominent.
  • the neural network is repeatedly classified for such image regions, and the contribution to the accuracy of the final classification result of the image region is prominent. Furthermore, in the process of reducing the resolution multiple times, the resolution of the remote sensing image after the previous resolution reduction is smaller than the resolution of the remote sensing image after the lower resolution processing, so that if the previous resolution is lowered After the accuracy of the recognition result of the image block processed by the rate has been satisfied, the application may not identify the image block in the corresponding region after the subsequent resolution reduction, thereby reducing the nerve to a large extent. The amount of calculation of the network and the processing time. It can be seen that the technical solution provided by the present application is advantageous for reducing the time and calculation cost of the neural network and improving the recognition efficiency of the remote sensing image of the neural network while ensuring the accuracy of the remote sensing image recognition.
  • Any remote sensing image recognition provided by the embodiments of the present application may be performed by any suitable device having data processing capabilities, including but not limited to: a terminal device, a server, and the like.
  • any remote sensing image recognition method provided by the embodiment of the present application may be executed by a processor.
  • the processor performs any remote sensing image recognition method mentioned in the embodiment of the present application by calling a corresponding instruction stored in the memory. This will not be repeated below.
  • the foregoing programs may be stored in a computer readable storage medium, and the program is executed when executed.
  • the operation of the foregoing method embodiment is included; and the foregoing storage medium includes at least one medium that can store program codes, such as a ROM, a RAM, a magnetic disk, or an optical disk.
  • FIG. 5 is a schematic structural diagram of an embodiment of a remote sensing image recognition apparatus according to the present application.
  • the apparatus of this embodiment mainly includes: a resolution reduction module 500, a remote sensing image segmentation module 510, a selected image block module 520, a classification processing module 530, and a determination recognition result module 540.
  • the device may further include: a training module 550.
  • the resolution resolution module 500 is mainly used for performing resolution reduction processing on the remote sensing image to be obtained, and obtaining a remote sensing image after the resolution reduction processing.
  • the down-resolution module 500 can be configured to perform a resolution reduction process on the remote sensing image to be recognized according to the reduced value that has not been used in the reduced scale set.
  • the reduced set of scales includes: at least two different reduced values, and the smaller reduced values will take precedence over the larger reduced values and are used by the reduced resolution module 500.
  • For the operations performed by the resolution reduction module 500 reference may be made to the descriptions of S100 in FIG. 1 and S200 in FIG. 2, and the description thereof will not be repeated here.
  • the remote sensing image segmentation module 510 is mainly used to segment at least one remote sensing image block from the remote sensing image after the resolution reduction module 500 performs the resolution reduction processing.
  • the remote sensing image segmentation module 510 reference may be made to the descriptions of S110 in FIG. 1 and S210 in FIG. 2, and the description thereof will not be repeated here.
  • the selected image block module 520 is mainly used to determine the remote sensing image block to be processed from the remote sensing image block segmented by the remote sensing image segmentation module 510.
  • the selected image block module 520 can use all the remote sensing image blocks that are segmented by the remote sensing image segmentation module 510 as the remote sensing image block to be processed.
  • the number of the remote sensing image blocks to be processed selected by the selected image block module 520 is less than or equal to the remote sensing image block currently segmented by the remote sensing image segmentation module. Quantity.
  • the selected image block module 520 may select the remote sensing image block that is segmented from the remote sensing image segmentation module 510 according to the current confidence of the pixel of the remote sensing image to be identified. Remote sensing image block to be processed.
  • the selected image block module 520 can include a first unit, a second unit, and a third unit.
  • the first unit is mainly configured to determine a current confidence level of the at least one pixel according to the current classification probability information of the at least one of the plurality of pixels included in the remote sensing image to be identified;
  • the second unit is mainly used to calculate the corresponding to the remote sensing image block.
  • the third unit is mainly configured to determine, as the to-be-processed remote sensing image block, the remote sensing image block that does not meet the predetermined requirement for the number of pixels that reach the predetermined confidence level.
  • the third unit will The remote sensing image block is determined as a remote sensing image block to be processed.
  • the operations performed by the first unit reference may be made to the description in S300 in FIG. 3, and the operations performed by the second unit may be referred to the description in S310 above for FIG. 3, and the operations performed by the third unit may be referred to the foregoing for FIG. 3.
  • the description in S320 is not repeated here.
  • the selected image block module 520 may further include: a fourth unit, a fifth unit, and a sixth unit.
  • the fourth unit is mainly used for mapping the classification probability information of at least one pixel in each remote sensing image block to be processed outputted by the neural network to the remote sensing image to be identified, and obtaining the to-be-identified image, in the case of performing the resolution-reduction processing for the first time.
  • Classification probability information of at least one pixel of the remote sensing image is used as current classification probability information of at least one pixel of the remote sensing image to be identified.
  • the fifth unit is mainly used to map the classification probability information of at least one pixel in each remote sensing image block to be processed outputted by the neural network to the remote sensing image to be recognized, without obtaining the first resolution processing.
  • the classification probability information of at least one pixel of the remote sensing image is identified.
  • the sixth unit is mainly configured to update current classification probability information of the corresponding pixel of the remote sensing image to be recognized according to the classification probability information of the at least one pixel of the remote sensing image to be identified obtained this time.
  • the sixth unit may be configured to calculate an average value of the classification probability information of the at least one pixel of the remote sensing image to be identified obtained and the current classification probability information of the corresponding pixel in the remote sensing image to be identified, and update the calculated average value.
  • the current classification probability information of the corresponding pixel in the remote sensing image to be identified For the operations performed by the fourth unit, reference may be made to the descriptions in S420 and S430 in FIG. 4, and the operations performed by the fifth unit may be referred to the description in S450 above for FIG. 4, and the operations performed by the sixth unit may be referred to the foregoing. The description in S460 in Fig. 4 will not be repeated here.
  • the classification processing module 530 is mainly used to obtain classification probability information of pixels in the remote sensing image block to be processed via the neural network.
  • the classification processing module 530 For the operations performed by the classification processing module 530, reference may be made to the related descriptions in S130 in FIG. 1 and S220 in FIG. 2, and the description thereof will not be repeated here.
  • the determination recognition result module 540 is mainly used to determine the recognition result of the remote sensing image to be recognized according to the classification probability information of the pixels in the remote sensing image block to be processed. For example, the determination recognition result module 540 may respond to the absence of the unused reduction value in the reduction ratio set or determine that there is no remote sensing image block to be processed in the remote sensing image block segmented by the remote sensing image segmentation module 510, and The classification probability information of the pixels in the remote sensing image block to be processed determines the recognition result of the remote sensing image to be identified.
  • the determination recognition result module 540 determines a maximum probability value among the probability values belonging to each category in the current classification probability information of the pixel, and the maximum probability value and the maximum The category corresponding to the probability value is used as the recognition result of the pixel.
  • the operation performed by the determination recognition result module 540 can be referred to the above description for S140 in FIG. 1 and S230 in FIG. 2, and the description thereof will not be repeated here.
  • the training module 550 is mainly used to train the neural network using remote sensing image samples.
  • the resolution of the remote sensing image sample includes: performing resolution reduction processing on the remote sensing image according to each of the reduced values in the reduced scale set, and obtaining the resolution of each remote sensing image.
  • the scaling set includes: a first reduced value, a second reduced value, and a third reduced value, the first reduced value corresponds to the first resolution, the second reduced value corresponds to the second resolution, and the third reduced value corresponds to the third resolution
  • the training module 550 can use the remote sensing image sample having the first resolution, the remote sensing image sample having the second resolution, and the remote sensing image sample having the third resolution to train the trained neural network, and successfully train the training.
  • Neural networks can be used for remote sensing image recognition of the present application.
  • FIG. 6 illustrates an exemplary device 600 suitable for implementing the present application, which may be a control system/electronic system configured in a car, a mobile terminal (eg, a smart mobile phone, etc.), a personal computer (PC, eg, a desktop computer Or a notebook computer, etc.), a tablet computer, a server, and the like.
  • a mobile terminal eg, a smart mobile phone, etc.
  • PC personal computer
  • tablet computer eg, a tablet computer, a server, and the like.
  • device 600 includes one or more processors, communication units, etc., which may be: one or more central processing units (CPUs) 601, and/or one or more utilized
  • the neural network performs an acceleration unit 613 for remote sensing image recognition, etc.
  • the acceleration unit 613 may include, but is not limited to, a GPU, an FPGA, other types of dedicated processors, etc.
  • the processor may execute executable instructions according to the read-only memory (ROM) 602.
  • ROM read-only memory
  • RAM random access memory
  • the communication unit 612 may include, but is not limited to, a network card, which may include, but is not limited to, an IB (Infiniband) network card.
  • the processor can communicate with read only memory 602 and/or random access memory 603 to execute executable instructions, connect to communication portion 612 via bus 604, and communicate with other target devices via communication portion 612 to accomplish the corresponding in this application. operating.
  • RAM 603 various programs and data required for the operation of the device can be stored.
  • the CPU 601 or the acceleration unit 613, the ROM 602, and the RAM 603 are connected to each other through a bus 604.
  • ROM 602 is an optional module.
  • the RAM 603 stores executable instructions, or writes executable instructions to the ROM 602 at runtime, the executable instructions causing the processor to perform operations corresponding to the methods described above.
  • An input/output (I/O) interface 605 is also coupled to bus 604.
  • the communication unit 612 may be integrated, or may be configured to have a plurality of sub-modules (for example, a plurality of IB network cards) and be respectively connected to the bus.
  • the following components are connected to the I/O interface 605: an input portion 606 including a keyboard, a mouse, etc.; an output portion 607 including, for example, a cathode ray tube (CRT), a liquid crystal display (LCD), and the like, and a storage portion 608 including a hard disk or the like. And a communication portion 609 including a network interface card such as a LAN card, a modem, or the like. The communication section 609 performs communication processing via a network such as the Internet.
  • Driver 610 is also coupled to I/O interface 605 as needed.
  • a removable medium 611 such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory or the like is mounted on the drive 610 as needed so that a computer program read therefrom is installed in the storage portion 608 as needed.
  • FIG. 6 is only an optional implementation manner.
  • the number and types of the components in FIG. 6 may be selected, deleted, added, or replaced according to actual needs.
  • separate implementations such as separate settings or integrated settings may also be employed.
  • the acceleration unit 613 and the CPU 601 may be separately configured.
  • the acceleration unit 613 may be integrated on the CPU 601, and the communication portion 612 may be separated.
  • the settings may also be integrated on the CPU 601 or the acceleration unit 613, and the like.
  • embodiments of the present application include a computer program product comprising tangibly embodied on a machine readable medium.
  • a computer program comprising program code for performing the operations illustrated by the flowcharts, the program code comprising instructions corresponding to performing the operations provided herein.
  • the computer program can be downloaded and installed from the network via communication portion 609, and/or installed from removable media 611.
  • the computer program is executed by the central processing unit (CPU) 601, the above-described instructions described in the present application are executed.
  • embodiments of the present disclosure also provide a computer program product for storing computer readable instructions that, when executed, cause a computer to perform the operations described in any of the above embodiments Remote sensing image recognition method.
  • the computer program product can be implemented by means of hardware, software or a combination thereof.
  • the computer program product is embodied as a computer storage medium.
  • the computer program product is embodied as a software product, such as a software development kit (SDK), etc. Wait.
  • SDK software development kit
  • embodiments of the present disclosure further provide another remote sensing image recognition method and corresponding apparatus and electronic device, computer storage medium, computer program, and computer program product, wherein the method includes: The first device transmits a remote sensing image recognition indication to the second device, the indication causing the second device to perform the remote sensing image recognition method in any of the possible embodiments described above; the first device receiving the remote sensing image recognition result transmitted by the second device.
  • the remote sensing image recognition indication may be specifically a call instruction
  • the first device may instruct the second device to perform a remote sensing image recognition operation by calling, and correspondingly, in response to receiving the call instruction, the second device may perform The operation and/or flow in any of the above remote sensing image recognition methods.
  • a plurality may mean two or more, and “at least one” may mean one, two or more.
  • the methods and apparatus, electronic devices, and computer readable storage media of the present application are possible in many ways.
  • the methods and apparatus, electronic devices, and computer readable storage media of the present application can be implemented in software, hardware, firmware, or any combination of software, hardware, or firmware.
  • the above-described sequence of operations for the method is for illustrative purposes only, and the operation of the method of the present application is not limited to the order specifically described above unless otherwise specifically stated.
  • the present application can also be implemented as a program recorded in a recording medium, the program including machine readable instructions for implementing the method according to the present application.
  • the present application also covers a recording medium storing a program for executing the method according to the present application.
  • FIG. 7 an application scenario in which an implementation may be implemented according to an embodiment of the present application is schematically illustrated.
  • the neural network 700 is a neural network for identifying a remote sensing image to be recognized (such as forming a snow mask or a cloud snow mask, etc.).
  • the sample data set used to train the neural network 700 typically includes a plurality of remote sensing image samples, and the resolution of all remote sensing image samples is not identical.
  • the successfully trained neural network 700 can be used to identify the remote sensing image to be recognized from the satellite, and the recognition result of the neural network 700 can form a cloud.
  • a mask such as a snow mask or a cloud snow mask.
  • the methods and apparatus, electronic devices, computer programs, and computer readable storage media of the present application are possible in many ways.
  • the methods and apparatus, electronic devices, computer programs, and computer readable storage media of the present application can be implemented by software, hardware, firmware or any combination of software, hardware, firmware.
  • the above-described sequence of operations for the method is for illustrative purposes only, and the operation of the method of the present application is not limited to the order specifically described above unless otherwise specifically stated.
  • the present application can also be implemented as a program recorded in a recording medium, the program including machine readable instructions for implementing the method according to the present application.
  • the present application also covers a recording medium storing a program for executing the method according to the present application.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • Multimedia (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • General Engineering & Computer Science (AREA)
  • Molecular Biology (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Medical Informatics (AREA)
  • Databases & Information Systems (AREA)
  • Astronomy & Astrophysics (AREA)
  • Remote Sensing (AREA)
  • Probability & Statistics with Applications (AREA)
  • Biodiversity & Conservation Biology (AREA)
  • Algebra (AREA)
  • Computational Mathematics (AREA)
  • Mathematical Analysis (AREA)
  • Mathematical Optimization (AREA)
  • Pure & Applied Mathematics (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Image Analysis (AREA)

Abstract

一种遥感图像识别方法、装置、设备、存储介质以及计算机程序，其中的遥感图像识别方法包括：对待识别遥感图像进行降分辨率处理，得到降分辨率处理后的遥感图像(S100)；从所述降分辨率处理后的遥感图像中切分出至少一个遥感图像块(S110)；从所述至少一个遥感图像块中确定待处理遥感图像块(S120)；将所述待处理遥感图像块输入神经网络中，经由神经网络获得输入的待处理遥感图像块中的像素的分类概率信息(S130)；根据所述待处理遥感图像块中的像素的分类概率信息确定所述待识别遥感图像的识别结果(S140)。

Description

遥感图像识别方法、装置、存储介质以及电子设备
本申请要求在2017年12月26日提交中国专利局、申请号为CN201711436470.2、发明名称为“遥感图像识别方法、装置、存储介质以及电子设备”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。
技术领域
本申请涉及计算机视觉技术，尤其是涉及一种遥感图像识别方法、遥感图像识别装置、计算机可读存储介质以及电子设备。
背景技术
随着深度学习技术在图像识别、目标检测以及图像分割等方面不断取得的突破性进展，卷积神经网络已经被应用于遥感图像识别领域中。
在遥感图像识别过程中，一个像素的预测结果通常是由该像素周边一定区域内的其他像素决定，该区域可以称为卷积神经网络神经元的感受野。
卷积神经网络神经元的感受野通常有限(如几百像素)，而遥感图像的尺寸通常较大(如遥感图像的长和宽可达几千甚至上万像素)，因此，神经元有限的感受野可能会无法获得足够的环境信息，这往往会导致卷积神经网络对遥感图像中的像素的误识别。
如何在尽量保证神经网络的计算量不会大幅提升的同时，提高遥感图像的识别准率性，是一个值得关注的技术问题。
发明内容
本申请实施方式提供一种遥感图像识别的技术方案。
根据本申请实施方式的其中一方面，提供了一种遥感图像识别方法，该方法主要包括：对待识别遥感图像进行降分辨率处理，得到降分辨率处理后的遥感图像；从所述降分辨率处理后的遥感图像中切分出至少一个遥感图像块；从所述至少一个遥感图像块中确定待处理遥感图像块；经由神经网络获得所述待处理遥感图像块中的像素的分类概率信息；根据所述待处理遥感图像块中的像素的分类概率信息确定所述待识别遥感图像的识别结果。
根据本申请实施方式的另一方面，提供了一种遥感图像识别装置，且该装置包括：降分辨率模块，用于对待识别遥感图像进行降分辨率处理，得到降分辨率处理后的遥感图像；遥感图像切分模块，用于从所述降分辨率处理后的遥感图像中切分出至少一个遥感图像块；选取图像块模块，用于从所述至少一个遥感图像块中确定待处理遥感图像块；分类处理模块，用于经由神经网络获得所述待处理遥感图像块中的像素的分类概率信息；确定识别结果模块，用于根据所述待处理遥感图像块中的像素的分类概率信息确定所述待识别遥感图像的识别结果。
根据本申请实施方式的再一个方面，提供了一种电子设备，包括：存储器，用于存储计算机程序；处理器，用于执行所述存储器中存储的计算机程序，且所述计算机程序被执行时，实现本申请任一实施方式所述的方法。
根据本申请实施方式的再一个方面，提供的一种计算机可读存储介质，其上存储有计算机程序，该计算机程序被处理器执行时，实现本申请任一实施方式所述的方法。
根据本申请实施方式的再一个方面，提供的一种计算机程序，该计算机程序在被设备中的处理器执行时，实现本申请任一实施方式所述的方法。
基于本申请提供的一种遥感图像识别方法、遥感图像识别装置、电子设备、计算机程序以及计算机可读存储介质，本申请通过从降分辨率处理后的遥感图像中选取遥感图像块，输入神经网络，避免了在每次降分辨率处理后，均将切分出的所有遥感图像块输入神经网络，而导致的神经网络的计算量较大的现象。
下面通过附图和实施方式，对本申请的技术方案做进一步的详细描述。
附图说明
构成说明书的一部分的附图描述了本申请的实施方式，并且连同描述一起用于解释本申请的原理。
参照附图，根据下面的详细描述，可以更加清楚地理解本申请，其中：
图1为本申请的遥感图像识别方法的一实施方式的流程示意图；
图2为本申请的遥感图像识别方法的另一实施方式的流程示意图；
图3为本申请的根据待识别遥感图像的像素的当前置信度选取遥感图像块的一实施方式的流程示意图；
图4为本申请的对待识别遥感图像的至少一个像素的当前分类概率信息进行维护的一实施方式的流程示意图；
图5为本申请的遥感图像识别装置的一实施方式示意图；
图6为实现本申请实施方式的一示例性设备的框图；
图7为本申请的一个应用场景示意图。
具体实施方式
现在将参照附图来详细描述本申请的各种示例性实施方式。应该注意到：除非另外具体说明，否则在这些实施方式中阐述的部件和步骤的相对布置、数字表达式和数值不限制本申请的范围。
同时，应当明白，为了便于描述，附图中所示出的各个部分的尺寸并不是按照实际的比例关系绘制的。
以下对至少一个示例性实施方式的描述实际上仅仅是说明性的，决不作为对本申请及其应用或使用的任何限制。
对于相关领域普通技术人员已知的技术、方法和设备可能不作详细讨论，但在适当情况下，所述技术、方法和设备应当被视为说明书的一部分。
应注意到：相似的标号和字母在下面的附图中表示类似项，因此，一旦某一项在一附图中被定义，则在随后的附图中不需要对其进行进一步讨论。
本申请实施方式可以应用于终端设备、计算机系统、服务器等电子设备，其可与众多其它通用或者专用计算系统环境或配置一起操作。适于与终端设备、计算机系统、服务器等电子设备一起使用的众所周知的计算系统、环境和/或配置的例子包括但不限于：个人计算机系统、服务器计算机系统、瘦客户 机、厚客户机、手持或膝上设备、基于微处理器的系统、机顶盒、可编程消费电子产品、网络个人电脑、小型计算机系统、大型计算机系统和包括上述任何系统的分布式云计算技术环境，等等。
终端设备、计算机系统、服务器等电子设备可以在由计算机系统执行的计算机系统可执行指令(诸如程序模块)的一般语境下描述。通常，程序模块可以包括例程、程序、目标程序、组件、逻辑以及数据结构等等，它们执行特定的任务或者实现特定的抽象数据类型。计算机系统/服务器可以在分布式云计算环境中实施，分布式云计算环境中，任务是由通过通信网络链接的远程处理设备执行的。在分布式云计算环境中，程序模块可以位于包括存储设备的本地或者远程计算系统存储介质上。
示例性实施例
本申请提供的用于实现遥感图像识别的技术方案可以由单片机、FPGA(FieldProgrammable Gate Array，现场可编程门阵列)、微处理器、智能移动电话、笔记型计算机、平板电脑、台式计算机或者服务器等能够运行计算机程序(也可以称为程序代码)的电子设备实现，且该计算机程序可以存储于闪存、缓存、硬盘或者光盘等计算机可读存储介质中。
下面结合图1至图7对本申请提供的用于实现遥感图像识别的技术方案进行说明。
图1为本申请的遥感图像识别方法的一实施方式的流程示意图。如图1所示，本实施方式的遥感图像识别方法主要包括：操作S100、操作S110、操作S120、操作S130以及操作S140。下面对本申请的各操作进行详细说明。
S100、对待识别遥感图像进行降分辨率处理，得到降分辨率处理后的遥感图像。
在一个可选示例中，本申请可以采用多种方式对待识别遥感图像进行降分辨率处理，例如，本申请预先设置有缩小比例集合，该缩小比例集合包括至少两个不同的缩小值，本申请可以从缩小比例集合中选取一个未被采用过的缩小值，并根据选取出的缩小值对待识别遥感图像进行降分辨率处理。另外，本申请在选取未被采用过的缩小值时，通常会优先选取未被采用过的最小的缩小值。再有，本申请还可以采用池化处理或者降采样处理等方式对待识别遥感图像进行降分辨率处理。本申请不限制对待识别遥感图像进行降分辨率处理的实现方式。
在一个可选示例中，S100可以由处理器调用存储器存储的相应指令执行，也可以由被处理器运行的降分辨率模块500执行。
S110，从降分辨率处理后的遥感图像中切分出至少一个遥感图像块。
在一个可选示例中，本申请通常是根据神经网络对输入图像的大小要求将降分辨率处理后获得的遥感图像切分为多个遥感图像块，切分出的所有遥感图像块的大小通常完全相同。切分出的每一个遥感图像块分别对应待识别遥感图像中的一个区域，不同遥感图像块对应区域不相同。遥感图像块对应的区域的大小通常与缩小值以及神经网络对输入图像的大小要求相关；一个例子，在神经网络对输入图像的大小要求为561×561，且降分辨率处理为降二分之一分辨率处理的情况下，本申请切分出的遥感图像块对应待识别遥感图像中的区域的大小为1122×1122，例如，切分出的第一个遥感图像块对应待识别遥感图像中的左上角1122×1122大小的区域，切分出的最后一个遥感图像块对应待识别遥感图像中的右下角1122×1122大小的区域。本申请不限制切分出的遥感图像块的数量以及遥感图像块的大小。
在一个可选示例中，S101可以由处理器调用存储器存储的相应指令执行，也可以由被处理器运行的遥感图像切分模块510执行。
S120，从至少一个遥感图像块中确定待处理遥感图像块。
在一个可选示例中，在第一次执行降分辨率处理的情况下，本申请可以将所有遥感图像块均作为待处理遥感图像块，而在非第一次执行降分辨率处理的情况下，本申请确定出的待处理遥感图像块的数量通常会小于当前切分出的所有遥感图像块的数量。
在一个可选示例中，在非第一次执行降分辨率处理的情况下，本申请可以根据待识别遥感图像的像素的当前置信度对切分出的至少一个遥感图像块进行选取，选取出的遥感图像块即为待处理遥感图像块。一个可选的例子，本申请先根据待识别遥感图像中的至少一个像素的当前分类概率信息确定至少一个像素的当前置信度，例如，针对待识别遥感图像中的一像素，计算该像素的最大分类概率值与次大分类概率值的比值，将计算出的该比值作为该像素的当前置信度，依据该处理可以获得待识别遥感图像中的所有像素的当前置信度。然后，统计遥感图像块所对应的待识别遥感图像区域中像素的当前置信度达到预定置信度的像素数量，并将达到预定置信度的像素数量不符合预定要求的遥感图像块确定为待处理遥感图像块，例如，针对一遥感图像块而言，在达到预定置信度的像素数量与该遥感图像块所对应的待识别遥感图像区域中所有像素的数量的比值小于预定比值的情况下，可以将该遥感图像块确定为待处理遥感图像块。
在一个可选示例中，本申请可以利用下述两种方式获得待识别遥感图像的至少一个像素的当前分类概率信息：
方式一，在第一次执行降分辨率处理的情况下，将神经网络输出的每个待处理遥感图像块中至少一个像素的分类概率信息映射到待识别遥感图像中，获得待识别遥感图像的至少一个像素的分类概率信息。本申请可以将当前获得的待识别遥感图像的至少一个像素的分类概率信息作为待识别遥感图像的至少一个像素的当前分类概率信息。
方式二、在非第一次执行降分辨率处理的情况下，将神经网络输出的每个待处理遥感图像块中至少一个像素的分类概率信息映射到待识别遥感图像中，获得待识别遥感图像的至少一个个像素的分类概率信息，并根据本次获得的待识别遥感图像的至少一个像素的分类概率信息更新待识别遥感图像的相应像素的当前分类概率信息，例如，计算本次获得的待识别遥感图像的至少一个像素的分类概率信息与待识别遥感图像中相应像素的当前分类概率信息的平均值，并利用计算出的平均值更新待识别遥感图像中相应像素的当前分类概率信息。在更新处理后，本申请获得待识别遥感图像的至少一个像素的当前分类概率信息。
在一个可选示例中，S102可以由处理器调用存储器存储的相应指令执行，也可以由被处理器运行的选取图像块模块520执行。
S130，经由神经网络获得待处理遥感图像块中的像素的分类概率信息。
在一个可选示例中，本申请中的神经网络可以为卷积神经网络，本申请不限制神经网络的表现形式。本申请通过对待识别遥感图像进行降分辨率处理，并将降分辨率处理后的遥感图像中的待处理的图像块输入神经网络中，有利于提高神经网络中的神经元的感受野，从而有利于避免神经网络对遥感图像的像素误识别。
在一个可选示例中，本申请中的神经网络是利用具有多种不同分辨率的遥感图像样本训练获得的。例如，所有遥感图像样本各自的分辨率，属于缩小比例集合中的每个缩小值分别对应的分辨率。作为一个例子，缩放比例集合包括：第一缩小值、第二缩小值和第三缩小值，第一缩小值对应第一分辨率，第二缩小值对应第二分辨率，第三缩小值对应第三分辨率，则本申请可以利用具有第一分辨率的遥感图像 样本、具有第二分辨率的遥感图像样本以及具有第三分辨率的遥感图像样本来对待训练的神经网络进行训练，成功训练的神经网络可以用于本申请的遥感图像识别。
在一个可选示例中，S103可以由处理器调用存储器存储的相应指令执行，也可以由被处理器运行的分类处理模块530执行。
S140，根据待处理遥感图像块中的像素的分类概率信息确定待识别遥感图像的识别结果。
在一个可选示例中，本申请会响应于缩小比例集合中不存在未被采用过的缩小值或者从至少一个遥感图像块中无法确定出待处理遥感图像块，从而执行根据待处理遥感图像块中的像素的分类概率信息确定待识别遥感图像的识别结果的操作，例如，针对待识别遥感图像中的一像素而言，确定该像素的当前分类概率信息中的属于每个类别的概率值中的最大概率值，将该最大概率值以及该最大概率值对应的类别，作为该像素的识别结果。
在一个可选示例中，S104可以由处理器调用存储器存储的相应指令执行，也可以由被处理器运行的确定识别结果模块540执行。
图2为本申请的遥感图像识别方法的另一实施方式的流程示意图。
如图2所示，本实施方式的遥感图像识别方法主要包括：操作S200、操作S210、操作S220以及操作S230。
S200，根据缩小比例集合中未被使用过的最小缩小值对待识别遥感图像进行降分辨率处理。
在一个可选示例中，缩小比例集合由多个缩小值形成，每一个缩小值的取值均大于零且不超过1。在通常情况下，缩小比例集合包括至少两个缩小值，一个可选的例子，Sr＝[0.25，0.5，1.0]，Sr表示缩小比例集合，其中的0.25、0.5和1为缩小比例集合中的缩小值，0.25表示将待识别遥感图像的分辨率降低到原分辨率的四分之一，0.5表示将待识别遥感图像的分辨率降低到原分辨率的二分之一，1表示待识别遥感图像的分辨率。本实施例将1作为是一种特殊的缩小值，例如，可以认为是指使待识别遥感图像保持原分辨率的缩小值。本申请不限制缩小比例集合中所包含的缩小值的取值大小以及所包含的缩小值的数量。
在一个可选示例中，本申请应按照缩小值从小到大的排列顺序，每次从缩小比例集合中选取一个缩小值，并根据选取的缩小值对待识别遥感图像进行降分辨率处理。被选取的缩小值即成为被使用过的缩小值。
在一个可选示例中，S200可以由处理器调用存储器存储的相应指令执行，也可以由被处理器运行的降分辨率模块500执行。
S210，将降分辨率处理后的遥感图像切分为至少一个遥感图像块。
在一个可选示例中，本申请通常是根据神经网络对输入图像的大小要求将降分辨率处理后获得的遥感图像切分为多个遥感图像块，切分出的所有遥感图像块的大小通常完全相同。切分出的每一个遥感图像块分别对应待识别遥感图像中的一个区域，不同遥感图像块对应区域不相同。遥感图像块对应的区域的大小通常与缩小值以及神经网络对输入图像的大小要求相关；一个例子，在神经网络对输入图像的大小要求为561×561，且降分辨率处理为降二分之一分辨率处理的情况下，本申请切分出的遥感图像块对应待识别遥感图像中的区域的大小为1122×1122，例如，切分出的第一个遥感图像块对应待识别遥感图像中的左上角1122×1122大小的区域，切分出的最后一个遥感图像块对应待识别遥感图像中的右下角1122×1122大小的区域。本申请不限制切分出的遥感图像块的数量以及遥感图像块的大小。
在一个可选示例中，S201可以由处理器调用存储器存储的相应指令执行，也可以由被处理器运行的遥感图像切分模块510执行。
S220，根据待识别遥感图像的像素的当前置信度从至少一个遥感图像块中选取遥感图像块，并经由神经网络获得选取出的每个遥感图像块中至少一个像素的分类概率信息。其中，选取出的遥感图像块即为上述待处理遥感图像块。
在一个可选示例中，在第一次执行降分辨率处理的情况下，本申请通常会将切分出的所有遥感图像块均作为选取出的遥感图像块。本申请可以认为在第一次执行降分辨率处理的情况下，待识别遥感图像的像素的当前置信度被初始化为0，因此，切分出的所有遥感图像块均会被选取出来，并分别输入到神经网络中，从而经由神经网络可以获得本次切分出的所有遥感图像块的所有像素的分类概率信息，本次获得的所有遥感图像块的所有像素的分类概率信息可以用于更新待识别遥感图像的像素的当前置信度。一个可选的例子，设定本申请的神经网络可以针对C个类别进行分类处理，在每个遥感图像块分别经由神经网络之后，本申请可以获得所有遥感图像块中的所有像素分别针对C个类别的分类概率信息。
在一个可选示例中，在非第一次执行降分辨率处理的情况下，本申请通常会根据待识别遥感图像的像素的当前置信度，从本次切分出的所有遥感图像块中选取遥感图像块。需要特别说明的是，本次选取出的遥感图像块有可能为本次切分出的所有遥感图像块，也有可能为本次切分出的所有遥感图像块中的部分遥感图像块；当然，还有可能无法从本次切分出的所有遥感图像块中选取出任何一个遥感图像块。
本申请通过对待识别遥感图像进行降分辨率处理，并将降分辨率处理后的遥感图像中的图像块输入神经网络(如卷积神经网络)中，有利于提高神经网络(如卷积神经网络)神经元的感受野，从而有利于避免神经网络对遥感图像的像素误识别。
在一个可选示例中，S202可以由处理器调用存储器存储的相应指令执行，也可以由被处理器运行的选取图像块模块520和分类处理模块530执行。
本申请根据待识别遥感图像的像素的当前置信度选取遥感图像块(即待处理遥感图像块)的一个可选例子如图3所示。
图3中，300，根据待识别遥感图像的至少一个像素的当前分类概率信息确定至少一个像素的当前置信度。
在一个可选示例中，本申请中的待识别遥感图像的至少一个像素的当前分类概率信息，是基于神经网络针对历次输入的至少一个遥感图像块输出的遥感图像块中的至少一个像素的分类概率信息而形成的。本申请在每一次执行降分辨率处理的过程中，都需要对待识别遥感图像的至少一个像素的当前分类概率信息进行维护(如更新待识别遥感图像的至少一个像素的当前分类概率信息)。本申请对待识别遥感图像的至少一个像素的当前分类概率信息进行维护的方式可以参见下述针对图4的描述。
在一个可选示例中，针对待识别遥感图像中的任一像素(i，j)，本申请根据待识别遥感图像的该像素(i，j)的当前分类概率信息确定该像素(i，j)的当前置信度的方式可以为：设定神经网络能够从遥感图像中识别出C个类别，像素(i，j)的当前分类概率信息中会包括C个概率值，本申请可以从这C个概率值中，挑选出最大概率值和次大概率值(即第2大概率值)，并计算挑选出的最大分类概率值与次大分类概率值的比值，本申请可以将该比值作为该像素(i，j)的当前置信度。本申请也可以根据待识别遥感图像的至少一个像素的当前分类概率信息采用其他方式确定待识别遥感图像的至少一个像素的当前置信度，本申请不限制确定待识别遥感图像的至少一个像素的当前置信度的方式。
在一个可选示例中，S300可以由处理器调用存储器存储的相应指令执行，也可以由被处理器运行的第一单元执行。
310，统计遥感图像块所对应的待识别遥感图像区域中至少一个像素的当前置信度达到预定置信度的像素数量。
在一个可选示例中，本申请中的预定置信度的大小可以根据实际需求设置，例如，预定置信度可以设置为4。续前例，在遥感图像块对应待识别遥感图像中的区域的大小为1122×1122的情况下，针对一个遥感图像块而言，本申请可以对1122×1122个像素的当前置信度分别进行是否达到预定置信度的判断，从而可以统计出达到预定置信度的像素数量。
在一个可选示例中，S301可以由处理器调用存储器存储的相应指令执行，也可以由被处理器运行的第二单元执行。
320，将达到预定置信度的像素数量不符合预定要求的遥感图像块作为选取出的遥感图像块。
在一个可选示例中，针对任一遥感图像块而言，如果该遥感图像块所对应的待识别遥感图像区域中达到预定置信度的像素数量与该区域中的至少一个像素的数量的比值未达到预定比值，则可以将该遥感图像块作为选取出的遥感图像块，输入至神经网络中，从而经由该神经网络获得该遥感图像块中至少一个像素的分类概率信息；如果上述比值达到预定比值，则不需要将该遥感图像块输入至神经网络中。本申请中的预定比值可以根据实际需求设置，例如，该预定比值可以设置为0.97等。
另外，本申请也可以采用其他方式来确定出不符合预定要求的遥感图像块，例如，在判断出该遥感图像块所对应的待识别遥感图像区域中达到预定置信度的像素数量与该区域中的未达到预定置信度的像素数量的差值达到预定差值的情况下，则可以将该遥感图像块作为选取出的遥感图像块，输入至神经网络中，否则，不需要将该遥感图像块输入至神经网络中；再例如，在判断出该遥感图像块所对应的待识别遥感图像区域中达到预定置信度的像素数量与该区域中的未达到预定置信度的像素数量的比值未达到预定比值的情况下，则可以将该遥感图像块作为选取出的遥感图像块，输入神经网络中，否则，不需要将该遥感图像块输入至神经网络中。本申请不限制确定出不符合预定要求的遥感图像块的实现方式。
在一个可选示例中，S302可以由处理器调用存储器存储的相应指令执行，也可以由被处理器运行的第三单元执行。
本申请对待识别遥感图像的至少一个像素的当前分类概率信息进行维护的一个例子如图4所示。
图4中，S400，本申请的遥感图像识别方法开始。可选的，本申请可以在遥感图像识别方法开始时，初始化待识别遥感图像的至少一个像素的当前分类概率信息，如将待识别遥感图像的至少一个像素的当前分类概率信息均设置为0。当然，不针对待识别遥感图像的至少一个像素的当前分类概率信息执行初始化操作，也是完全可行的。
在本申请第一次执行降分辨率处理的情况下，触发操作S410的执行。
在本申请第二次或者第三次等非第一次执行降分辨率处理的情况下，触发操作S440的执行。
S410，将本次降分辨率处理后的遥感图像所切分出的所有遥感图像块均输入至神经网络中，从而神经网络会针对切分出的所有遥感图像分别输出每个遥感图像块中至少一个像素的分类概率信息。到操作S420。
S420，将神经网络输出的每个遥感图像块中的至少一个像素的分类概率信息分别映射到待识别遥 感图像中，从而获得待识别遥感图像的至少一个像素的分类概率信息。到操作S430。
在一个可选示例中，在本申请第一次执行降分辨率处理的情况下，基于神经网络输出的每个遥感图像块中的所有像素的分类概率信息，而获得待识别遥感图像中的任一像素(i，j)的分类概率信息
Figure PCTCN2018123807-appb-000001
可以表示为：
Figure PCTCN2018123807-appb-000002
其中，
Figure PCTCN2018123807-appb-000003
和
Figure PCTCN2018123807-appb-000004
中的上角标处1表示第一次执行降分辨率处理，C表示神经网络进行分类的类别数量，
Figure PCTCN2018123807-appb-000005
表示在第一次执行降分辨率处理的情况下，待识别遥感图像中的任一像素(i，j)属于第一类别的分类概率信息，
Figure PCTCN2018123807-appb-000006
表示在第一次执行降分辨率处理的情况下，待识别遥感图像中的任一像素(i，j)属于第二类别的分类概率信息，
Figure PCTCN2018123807-appb-000007
表示在第一次执行降分辨率处理的情况下，待识别遥感图像中的任一像素(i，j)属于第C类别的分类概率信息。
S430，将本次获得的待识别遥感图像的至少一个像素的分类概率信息作为待识别遥感图像的相应像素的当前分类概率信息，例如，待识别遥感图像的任一像素(i，j)的当前分类概率信息为：P ij＝[P ij,1,P ij,2,......P ij,C]。
在一个可选示例中，S420和S430可以由处理器调用存储器存储的相应指令执行，也可以由被处理器运行的第四单元执行。
S440，从本次降分辨率处理后的遥感图像所切分出的所有遥感图像块中选取遥感图像块，将选取出的遥感图像块输入至神经网络中。到操作S450。
本申请选取遥感图像块的过程的一个例子可以参见上述针对图3的描述，在此不再重复说明。
S450，将神经网络输出的每个遥感图像块中的至少一个像素的分类概率信息分别映射到待识别遥感图像中，从而获得待识别遥感图像的至少一个像素的分类概率信息。到操作S460。
在一个可选示例中，在本申请第n次(n大于1)执行降分辨率处理的情况下，基于神经网络输出的每个遥感图像块中的所有像素的分类概率信息，而获得待识别遥感图像中的任一像素(i，j)的分类概率信息
Figure PCTCN2018123807-appb-000008
可以表示为：
Figure PCTCN2018123807-appb-000009
其中，
Figure PCTCN2018123807-appb-000010
和
Figure PCTCN2018123807-appb-000011
中的上角标处n表示第n次执行降分辨率处理，C表示神经网络进行分类的类别数量，
Figure PCTCN2018123807-appb-000012
表示在第n次执行降分辨率处理的情况下，待识别遥感图像中的任一像素(i，j)属于第一类别的分类概率信息，
Figure PCTCN2018123807-appb-000013
表示在第n次执行降分辨率处理的情况下，待识别遥感图像中的任一像素(i，j)属于第二类别的分类概率信息，
Figure PCTCN2018123807-appb-000014
表示在第n次执行降分辨率处理的情况下，待识别遥感图像中的任一像素(i，j)属于第C类别的分类概率信息。需要特别说明的是，如果待识别遥感图像中的像素(i，j)属于本次降分辨率处理过程选取出的遥感图像块映射到待识别遥感图像中的区 域的像素，则
Figure PCTCN2018123807-appb-000015
为神经网络输出的相应像素的分类概率信息，而如果像素(i，j)不属于本次降分辨率处理过程选取出的遥感图像块映射到待识别遥感图像中的区域的像素，则本申请可以将
Figure PCTCN2018123807-appb-000016
设置为0。
在一个可选示例中，S450可以由处理器调用存储器存储的相应指令执行，也可以由被处理器运行的第五单元执行。
S460，利用本次获得的待识别遥感图像的至少一个像素的分类概率信息更新待识别遥感图像的相应像素的当前分类概率信息。
在一个可选示例中，在第二次或第三次等非第一次执行降分辨率处理的情况下，本申请是通过累积历次分别获得的每个待处理遥感图像块中所有像素的分类概率信息来获得待识别遥感图像的所有像素的当前分类概率信息的。
设定在第一次执行降分辨率处理的情况下，本申请获得的待识别遥感图像中的任一像素(i，j)的分类概率信息为
Figure PCTCN2018123807-appb-000017
从而本申请将
Figure PCTCN2018123807-appb-000018
记录为待识别遥感图像中的任一像素(i，j)的当前分类概率信息P ij＝[P ij,1,P ij,2,......P ij,C]；
在第二次执行降分辨率处理的情况下，本申请基于神经网络获得的待识别遥感图像中的任一像素(i，j)的分类概率信息可以表示为
Figure PCTCN2018123807-appb-000019
其中，如果待识别遥感图像中的像素(i，j)属于第二次降分辨率处理过程选取出的遥感图像块映射到待识别遥感图像中的区域的像素，则
Figure PCTCN2018123807-appb-000020
为神经网络输出的相应像素的分类概率信息，而如果待识别遥感图像中的像素(i，j)不属于第二次降分辨率处理过程选取出的遥感图像块映射到待识别遥感图像中的区域的像素，则
Figure PCTCN2018123807-appb-000021
为0；本申请可以针对非0的
Figure PCTCN2018123807-appb-000022
计算
Figure PCTCN2018123807-appb-000023
和
Figure PCTCN2018123807-appb-000024
的平均值，并利用计算出的平均值更新待识别遥感图像中的相应像素(i，j)的当前分类概率信息P ij＝[P ij,1,P ij,2,......P ij,C]。
在第三次执行降分辨率处理的情况下，本申请基于神经网络获得的待识别遥感图像中的任一像素(i，j)的分类概率信息可以表示为
Figure PCTCN2018123807-appb-000025
其中，如果待识别遥感图像中的像素(i，j)属于第三次降分辨率处理过程选取出的遥感图像块映射到待识别遥感图像中的区域的像素，则
Figure PCTCN2018123807-appb-000026
为神经网络输出的相应像素的分类概率信息，而如果待识别遥感图像中的像素(i，j)不属于第三次降分辨率处理过程选取出的遥感图像块映射到待识别遥感图像中的区域的像素，则
Figure PCTCN2018123807-appb-000027
为0；本申请可以针对非0的
Figure PCTCN2018123807-appb-000028
计算
Figure PCTCN2018123807-appb-000029
和
Figure PCTCN2018123807-appb-000030
的平均值，并利用计算出的平均值更新待识别遥感图像中的相应像素(i，j)的当前分类概率信息P ij＝[P ij,1,P ij,2,......P ij,C]。
以此类推，根据上述针对在第二次以及第三次执行降分辨率处理的情况下，所执行的更新待识别遥感图像的所有像素的当前分类概率信息的过程，可以得知在第四次或第五次等非第一次(第n次)执行 降分辨率处理的情况下，所执行的更新待识别遥感图像的所有像素的当前分类概率信息的过程其实是对历次的分类概率信息进行融合，例如，对
Figure PCTCN2018123807-appb-000031
进行融合，形成更新后的待识别遥感图像的所有像素的当前分类概率信息。然而，本申请在选取需要输入神经网络的遥感图像块时，所依据的待识别遥感图像的所有像素的当前分类概率信息是针对之前历次的分类概率信息进行融合的结果，例如，针对
Figure PCTCN2018123807-appb-000032
进行融合的结果。
在一个可选示例中，S460可以由处理器调用存储器存储的相应指令执行，也可以由被处理器运行的第六单元执行。
S230，在缩小比例集合中不存在未被使用过的最小缩小值或者从至少一个遥感图像块中选取不出遥感图像块的情况下，根据基于每个选取出的遥感图像块中至少一个像素的分类概率信息形成的待识别遥感图像中至少一个像素的分类概率信息确定待识别遥感图像的识别结果，否则，返回上述S200。
在一个可选示例中，在缩小比例集合中不存在未被使用过的最小缩小值的情况下，本申请可以根据待识别遥感图像中所有像素的当前分类概率信息确定待识别遥感图像的识别结果，即本申请针对
Figure PCTCN2018123807-appb-000033
……以及
Figure PCTCN2018123807-appb-000034
进行融合后，获得P ij＝[P ij,1,P ij,2,......P ij,C]，将P ij，1，P ij，2，......P ij，C中的最大概率值作为待识别遥感图像中像素(i，j)的识别结果。
在一个可选示例中，在从至少一个遥感图像块中选取不出遥感图像块的情况下，本申请可以根据待识别遥感图像中所有像素的当前分类概率信息确定待识别遥感图像的识别结果，即本申请针对
Figure PCTCN2018123807-appb-000035
……以及
Figure PCTCN2018123807-appb-000036
进行融合后，获得P ij＝[P ij,1,P ij,2,......P ij,C]，将P ij,1,P ij,2,......P ij,C中的最大概率值作为待识别遥感图像中像素(i，j)的识别结果。
在一个可选示例中，S230可以由处理器调用存储器存储的相应指令执行，也可以由被处理器运行的确定识别结果模块540执行。
在一个可选示例中，本申请在对神经网络进行训练时，所采用的遥感图像样本通常是包括多种分辨率的遥感图像样本，且不同分辨率的数量通常与缩小比例集合中的缩小值的数量相关，例如，在缩小比例集合包括三种缩小值的情况下，样本数据集中应至少包括这三种缩小值所对应的分辨率的遥感图像样本，每一个遥感图像样本均具有标注信息(如掩膜标注信息等)。通常情况下，在样本数据集中，不同分辨率的遥感图像样本的数量大致相同。本申请通过利用样本数据集中的具有不同分辨率的遥感图像样本对神经网络进行训练，有利于提高神经网络针对输入的待处理遥感图像块而输出分类概率信息的准确度，从而有利于提高神经网络对遥感图像识别的准确率。
遥感图像通常不同于传统意义上的照片或者图片或者视频帧等图像，遥感图像的尺寸通常较大，如遥感图像的长和宽可达几千甚至上万像素，如果直接利用神经网络对待识别的遥感图像进行识别处理，一方面，可能会由于神经元有限的感受野而使识别结果的可靠性受到不良影响，另一方面，神经网络的计算成本以及时间成本可能会较高。本申请通过对待识别的遥感图像进行降分辨率处理，并使神经网络针对降分辨率处理后的遥感图像进行处理，有利于降低神经网络的计算成本以及时间成本。本申请通过从降分辨率处理后的遥感图像中选取遥感图像块，输入神经网络，有利于降低神经网络针对降分辨率处 理后的遥感图像中的相应的局部区域进行识别处理的计算量。由于本申请在降分辨率处理过程中，可以根据待识别遥感图像中的像素的当前置信度来选取遥感图像块，从而对于待识别遥感图像中的一些易于辨识的图像区域而言，由于神经网络对其辨识的准确度较高，因此，在每次降分辨率处理后，均使神经网络针对这样的图像区域反复进行分类处理，对于该图像区域的最终分类结果的准确度的贡献并不突出；而对于待识别遥感图像中的一些不易辨识的复杂的图像区域而言，使神经网络针对这样的图像区域反复进行分类处理，对于该图像区域的最终分类结果的准确度的贡献较为突出。再有，在多次降分辨率的过程中，通过使前一次降分辨率处理后的遥感图像的分辨率小于后一次降分辨率处理后的遥感图像的分辨率，这样，如果前一次降分辨率处理后的一图像块的识别结果的准确度，已经满足要求，则本申请可以不对后一次降分辨率处理后的相应区域中的图像块进行识别处理，从而可以在较大程度上减少神经网络的计算量以及处理时间。由此可知，本申请提供的技术方案有利于在保证遥感图像识别准确率的情况下，降低神经网络的时间和计算成本，提高神经网络的遥感图像识别效率。
在一个实际测试过程中，在由209张高分1号(GF-1)卫星的PMS(Panchromatic and Multispectral Scanners，全色多光谱图像)的遥感图像验证集上，使用本申请的技术方案，在缩小比例集合中的缩小值为0.5和1.0的情况下，相对于现有技术中的遥感图像识别而言，计算量大致可以减少35％，而识别精度却略有提升，例如，mIoU可以由0.8366提升至0.8383。
本申请实施例提供的任一种遥感图像识别可以由任意适当的具有数据处理能力的设备执行，包括但不限于：终端设备和服务器等。或者，本申请实施例提供的任一种遥感图像识别方法可以由处理器执行，如处理器通过调用存储器存储的相应指令来执行本申请实施例提及的任一种遥感图像识别方法。下文不再赘述。
本领域普通技术人员可以理解：实现上述方法实施例的全部或部分操作可以通过程序指令相关的硬件来完成，前述的程序可以存储于一计算机可读取存储介质中，该程序在执行时，执行包括上述方法实施例的操作；而前述的存储介质包括：ROM、RAM、磁碟或者光盘等至少一个种可以存储程序代码的介质。
图5为本申请的遥感图像识别装置的一实施例的结构示意图。如图5所示，该实施例的装置主要包括：降分辨率模块500、遥感图像切分模块510、选取图像块模块520、分类处理模块530以及确定识别结果模块540。可选的，该装置还可以包括：训练模块550。
降分辨率模块500主要用于对待识别遥感图像进行降分辨率处理，得到降分辨率处理后的遥感图像。例如，降分辨率模块500可以用于根据缩小比例集合中未被采用过的缩小值对待识别遥感图像进行降分辨率处理。该缩小比例集合包括：至少两个不同的缩小值，且较小的缩小值会优先于较大的缩小值，而被降分辨率模块500所采用。降分辨率模块500执行的操作可以参见上述针对图1中的S100以及图2中的S200的描述，在此不再重复说明。
遥感图像切分模块510主要用于从降分辨率模块500执行降分辨率处理后的遥感图像中切分出至少一个遥感图像块。遥感图像切分模块510执行的操作可以参见上述针对图1中的S110以及图2中的S210的描述，在此不再重复说明。
选取图像块模块520主要用于从遥感图像切分模块510切分出的遥感图像块中确定待处理遥感图像块。可选的，在本申请的装置第一次执行降分辨率处理的情况下，选取图像块模块520可以将遥感图像切分模块510切分出的所有遥感图像块均作为待处理遥感图像块。在本申请的装置非第一次执行降分辨 率处理的情况下，选取图像块模块520选取出的待处理遥感图像块的数量小于或等于遥感图像切分模块当前切分出的遥感图像块的数量。例如，在非第一次执行降分辨率处理的情况下，选取图像块模块520可以根据待识别遥感图像的像素的当前置信度从遥感图像切分模块510切分出的遥感图像块中选取出待处理遥感图像块。
在一个可选示例中，选取图像块模块520可以包括：第一单元、第二单元以及第三单元。其中的第一单元主要用于根据待识别遥感图像包括的多个像素中至少一个像素的当前分类概率信息确定至少一个像素的当前置信度；第二单元主要用于统计遥感图像块所对应的待识别遥感图像区域中至少一个像素的当前置信度达到预定置信度的像素数量；第三单元主要用于将达到预定置信度的像素数量不符合预定要求的遥感图像块确定为待处理遥感图像块。例如，针对一遥感图像块而言，在达到预定置信度的像素数量与该遥感图像块所对应的待识别遥感图像区域中至少一个像素的数量的比值小于预定比值的情况下，第三单元将该遥感图像块确定为待处理遥感图像块。第一单元执行的操作可以参见上述针对图3中的S300中的描述，第二单元执行的操作可以参见上述针对图3中的S310中的描述，第三单元执行的操作可以参见上述针对图3中的S320中的描述，在此不再重复说明。
在一个可选示例中，选取图像块模块520还可以包括：第四单元、第五单元以及第六单元。第四单元主要用于在第一次执行降分辨率处理的情况下，将神经网络输出的每个待处理遥感图像块中至少一个像素的分类概率信息映射到待识别遥感图像中，获得待识别遥感图像的至少一个像素的分类概率信息。其中的待识别遥感图像的至少一个像素的分类概率信息被作为待识别遥感图像的至少一个像素的当前分类概率信息。第五单元主要用于在非第一次执行降分辨率处理的情况下，将神经网络输出的每个待处理遥感图像块中至少一个像素的分类概率信息映射到待识别遥感图像中，获得待识别遥感图像的至少一个像素的分类概率信息。第六单元主要用于根据本次获得的待识别遥感图像的至少一个像素的分类概率信息更新所述待识别遥感图像的相应像素的当前分类概率信息。例如，第六单元可以用于计算本次获得的待识别遥感图像的至少一个像素的分类概率信息与待识别遥感图像中相应像素的当前分类概率信息的平均值，并利用计算出的平均值更新待识别遥感图像中相应像素的当前分类概率信息。第四单元执行的操作可以参见上述针对图4中的S420以及S430中的描述，第五单元执行的操作可以参见上述针对图4中的S450中的描述，第六单元执行的操作可以参见上述针对图4中的S460中的描述，在此不再重复说明。
分类处理模块530主要用于经由神经网络获得待处理遥感图像块中的像素的分类概率信息。分类处理模块530执行的操作可以参见上述针对图1中的S130以及图2中的S220中的相关描述，在此不再重复说明。
确定识别结果模块540主要用于根据待处理遥感图像块中的像素的分类概率信息确定待识别遥感图像的识别结果。例如，确定识别结果模块540可以响应于缩小比例集合中不存在未被采用过的缩小值或者确定出遥感图像切分模块510切分出的遥感图像块中不存在待处理遥感图像块，而根据待处理遥感图像块中的像素的分类概率信息确定所述待识别遥感图像的识别结果。例如，针对待识别遥感图像中的一像素而言，确定识别结果模块540确定该像素的当前分类概率信息中的属于每个类别的概率值中的最大概率值，将该最大概率值以及该最大概率值对应的类别，作为该像素的识别结果。确定识别结果模块540执行的操作可以参见上述针对图1中的S140以及图2中的S230中的描述，在此不再重复说明。
训练模块550主要用于利用遥感图像样本对神经网络进行训练。其中的遥感图像样本的分辨率包 括：根据缩小比例集合中的每个缩小值分别对遥感图像进行降分辨率处理，所获得的每个遥感图像的分辨率。例如，缩放比例集合包括：第一缩小值、第二缩小值和第三缩小值，第一缩小值对应第一分辨率，第二缩小值对应第二分辨率，第三缩小值对应第三分辨率，则训练模块550可以利用具有第一分辨率的遥感图像样本、具有第二分辨率的遥感图像样本以及具有第三分辨率的遥感图像样本，来对待训练的神经网络进行训练，成功训练的神经网络可以用于本申请的遥感图像识别。
示例性设备
图6示出了适于实现本申请的示例性设备600，设备600可以是汽车中配置的控制系统/电子系统、移动终端(例如，智能移动电话等)、个人计算机(PC，例如，台式计算机或者笔记型计算机等)、平板电脑以及服务器等。图6中，设备600包括一个或者多个处理器、通信部等，所述一个或者多个处理器可以为：一个或者多个中央处理单元(CPU)601，和/或，一个或者多个利用神经网络进行遥感图像识别的加速单元613等，加速单元613可包括但不限于GPU、FPGA、其他类型的专用处理器等，处理器可以根据存储在只读存储器(ROM)602中的可执行指令或者从存储部分608加载到随机访问存储器(RAM)603中的可执行指令而执行各种适当的动作和处理。通信部612可以包括但不限于网卡，所述网卡可以包括但不限于IB(Infiniband)网卡。处理器可与只读存储器602和/或随机访问存储器603中通信以执行可执行指令，通过总线604与通信部612相连、并经通信部612与其他目标设备通信，从而完成本申请中的相应操作。
上述各指令所执行的操作可以参见上述方法实施例中的相关描述，在此不再详细说明。
此外，在RAM 603中，还可以存储有装置操作所需的各种程序以及数据。CPU601或加速单元613、ROM602以及RAM603通过总线604彼此相连。在有RAM603的情况下，ROM602为可选模块。RAM603存储可执行指令，或在运行时向ROM602中写入可执行指令，可执行指令使处理器执行上述方法对应的操作。输入/输出(I/O)接口605也连接至总线604。通信部612可以集成设置，也可以设置为具有多个子模块(例如，多个IB网卡)，并分别与总线连接。
以下部件连接至I/O接口605：包括键盘、鼠标等的输入部分606；包括诸如阴极射线管(CRT)、液晶显示器(LCD)等以及扬声器等的输出部分607；包括硬盘等的存储部分608；以及包括诸如LAN卡、调制解调器等的网络接口卡的通信部分609。通信部分609经由诸如因特网的网络执行通信处理。驱动器610也根据需要连接至I/O接口605。可拆卸介质611，诸如磁盘、光盘、磁光盘、半导体存储器等等，根据需要安装在驱动器610上，以便于从其上读出的计算机程序根据需要被安装在存储部分608中。
需要特别说明的是，如图6所示的架构仅为一种可选实现方式，在具体实践过程中，可根据实际需要对上述图6的部件数量和类型进行选择、删减、增加或替换；在不同功能部件设置上，也可采用分离设置或集成设置等实现方式，例如，加速单元613和CPU601可分离设置，再如理，可将加速单元613集成在CPU601上，通信部612可分离设置，也可集成设置在CPU601或加速单元613上等。这些可替换的实施方式均落入本申请的保护范围。
特别地，根据本申请的实施方式，下文参考流程图描述的过程可以被实现为计算机软件程序，例如，本申请实施方式包括一种计算机程序产品，其包含有形地包含在机器可读介质上的计算机程序，计算机程序包含用于执行流程图所示的操作的程序代码，程序代码可包括对应执行本申请提供的操作对应的指令。
在这样的实施方式中，该计算机程序可以通过通信部分609从网络上被下载及安装，和/或从可拆卸介质611被安装。在该计算机程序被中央处理单元(CPU)601执行时，执行本申请中记载的上述指令。
在一个或多个可选实施方式中，本公开实施例还提供了一种计算机程序程序产品，用于存储计算机可读指令，所述指令被执行时使得计算机执行上述任意实施例中所述的遥感图像识别方法。
该计算机程序产品可以具体通过硬件、软件或其结合的方式实现。在一个可选例子中，所述计算机程序产品具体体现为计算机存储介质，在另一个可选例子中，所述计算机程序产品具体体现为软件产品，例如软件开发包(Software Development Kit，SDK)等等。
在一个或多个可选实施方式中，本公开实施例还提供了另一种遥感图像识别方法及其对应的装置和电子设备、计算机存储介质、计算机程序以及计算机程序产品，其中的方法包括：第一装置向第二装置发送遥感图像识别指示，该指示使得第二装置执行上述任一可能的实施例中的遥感图像识别方法；第一装置接收第二装置发送的遥感图像识别结果。
在一些实施例中，该遥感图像识别指示可以具体为调用指令，第一装置可以通过调用的方式指示第二装置执行遥感图像识别操作，相应地，响应于接收到调用指令，第二装置可以执行上述遥感图像识别方法中的任意实施例中的操作和/或流程。
应理解，本公开实施例中的“第一”、“第二”等术语仅仅是为了区分，而不应理解成对本公开实施例的限定。
还应理解，在本公开中，“多个”可以指两个或两个以上，“至少一个”可以指一个、两个或两个以上。
还应理解，对于本公开中提及的任一部件、数据或结构，在没有明确限定或者在前后文给出相反启示的情况下，一般可以理解为一个或多个。
还应理解，本公开对至少一个个实施例的描述着重强调至少一个个实施例之间的不同之处，其相同或相似之处可以相互参考，为了简洁，不再一一赘述。
可能以许多方式来实现本申请的方法和装置、电子设备以及计算机可读存储介质。例如，可通过软件、硬件、固件或者软件、硬件、固件的任何组合来实现本申请的方法和装置、电子设备以及计算机可读存储介质。用于方法的操作的上述顺序仅是为了进行说明，本申请的方法的操作不限于以上具体描述的顺序，除非以其它方式特别说明。此外，在一些实施方式中，还可将本申请实施为记录在记录介质中的程序，这些程序包括用于实现根据本申请的方法的机器可读指令。因而，本申请还覆盖存储用于执行根据本申请的方法的程序的记录介质。
本申请的描述是为了示例和描述起见而给出的，而并不是无遗漏的或者将本申请限于所公开的形式。很多修改和变化对于本领域的普通技术人员而言是显然的。选择以及描述实施方式是为了更好说明本申请的原理和实际应用，并且使本领域的普通技术人员能够理解本申请实施例可以从而设计适于特定用途的带有各种修改的各种实施方式。
示例性应用场景
参考图7，示意性地示出了根据本申请实施方式的可以在其中实现的一个应用场景。
图7中，神经网络700为用于针对待识别遥感图像进行识别(如形成云雪掩膜或者云雪水掩膜等)的神经网络。用于训练神经网络700的样本数据集通常包括多个遥感图像样本，且所有遥感图像样本的分辨率并不完全相同。在利用样本数据集中具有不同分辨率的遥感图像样本对神经网络700成功训练 后，可以利用成功训练后的神经网络700对来自卫星的待识别遥感图像进行识别，神经网络700的识别结果可以形成云雪掩膜或者云雪水掩膜等掩膜。
然而，本领域技术人员完全可以理解，本申请实施方式的适用场景不受到该框架任何方面的限制。
可能以许多方式来实现本申请的方法和装置、电子设备、计算机程序以及计算机可读存储介质。例如，可通过软件、硬件、固件或者软件、硬件、固件的任何组合来实现本申请的方法和装置、电子设备、计算机程序以及以及计算机可读存储介质。用于方法的操作的上述顺序仅是为了进行说明，本申请的方法的操作不限于以上具体描述的顺序，除非以其它方式特别说明。此外，在一些实施方式中，还可将本申请实施为记录在记录介质中的程序，这些程序包括用于实现根据本申请的方法的机器可读指令。因而，本申请还覆盖存储用于执行根据本申请的方法的程序的记录介质。
本申请的描述是为了示例和描述起见而给出的，而并不是无遗漏的或者将本申请限于所公开的形式。很多修改和变化对于本领域的普通技术人员而言是显然的。选择和描述实施方式是为了更好说明本申请的原理和实际应用，并且使本领域的普通技术人员能够理解本申请从而设计适于特定用途的带有各种修改的各种实施方式。

Claims (31)

  1. 一种遥感图像识别方法，其特征在于，所述方法包括：
    对待识别遥感图像进行降分辨率处理，得到降分辨率处理后的遥感图像；
    从所述降分辨率处理后的遥感图像中切分出至少一个遥感图像块；
    从所述至少一个遥感图像块中确定待处理遥感图像块；
    经由神经网络获得所述待处理遥感图像块中的像素的分类概率信息；
    根据所述待处理遥感图像块中的像素的分类概率信息确定所述待识别遥感图像的识别结果。
  2. 根据权利要求1所述的方法，其特征在于，所述对待识别遥感图像进行降分辨率处理包括：
    根据缩小比例集合中未被采用过的缩小值对待识别遥感图像进行降分辨率处理。
  3. 根据权利要求2所述的方法，其特征在于，所述根据所述待处理遥感图像块中的像素的分类概率信息确定所述待识别遥感图像的识别结果包括：
    响应于缩小比例集合中不存在未被采用过的缩小值或者确定所述至少一个遥感图像块中不存在待处理遥感图像块，根据所述待处理遥感图像块中的像素的分类概率信息确定所述待识别遥感图像的识别结果。
  4. 根据权利要求2至3中任一项所述的方法，其特征在于，所述缩小比例集合包括：至少两个不同的缩小值，且较小的缩小值相对较大的缩小值先被采用。
  5. 根据权利要求1至4中任一项所述的方法，其特征在于，所述从所述至少一个遥感图像块中确定待处理遥感图像块包括：
    在第一次执行降分辨率处理的情况下，将所有遥感图像块均作为待处理遥感图像块。
  6. 根据权利要求1至5中任一项所述的方法，其特征在于，所述从所述至少一个遥感图像块中确定待处理遥感图像块包括：
    在非第一次执行降分辨率处理的情况下，待处理遥感图像块的数量小于或等于当前切分出的遥感图像块的数量，和/或，所述待识别遥感图像经当前执行降分辨率处理后的图像分辨率大于经上一次执行降分辨率处理后的图像分辨率。
  7. 根据权利要求6所述的方法，其特征在于，在非第一次执行降分辨率处理的情况下，根据待识别遥感图像的像素的当前置信度从所述至少一个遥感图像块中确定待处理遥感图像块。
  8. 根据权利要求7所述的方法，其特征在于，所述根据待识别遥感图像的像素的当前置信度从所述至少一个遥感图像块中确定待处理遥感图像块，包括：
    根据待识别遥感图像包括的多个像素中至少一个像素的当前分类概率信息确定所述至少一个像素的当前置信度；
    统计遥感图像块所对应的待识别遥感图像区域中所述至少一个像素的当前置信度达到预定置信度的像素数量；
    将所述达到预定置信度的像素数量不符合预定要求的遥感图像块确定为待处理遥感图像块。
  9. 根据权利要求8所述的方法，其特征在于，所述将所述达到预定置信度的像素数量不符合预定要求的遥感图像块确定为待处理遥感图像块包括：
    针对一遥感图像块而言，在达到预定置信度的像素数量与该遥感图像块所对应的待识别遥感图像区域中所述至少一个像素的数量的比值小于预定比值的情况下，将该遥感图像块确定为待处理遥感图像块。
  10. 根据权利要求8至9中任一项所述的方法，其特征在于，所述待识别遥感图像的至少一个像素的当前分类概率信息的获得方式包括：
    在第一次执行降分辨率处理的情况下，将神经网络输出的每个待处理遥感图像块中至少一个像素的分类概率信息映射到待识别遥感图像中，获得待识别遥感图像的至少一个像素的分类概率信息；
    其中，所述待识别遥感图像的至少一个像素的分类概率信息被作为待识别遥感图像的至少一个像素的当前分类概率信息。
  11. 根据权利要求8至10中任一项所述的方法，其特征在于，所述待识别遥感图像的至少一个像素的当前分类概率信息的获得方式包括：
    在非第一次执行降分辨率处理的情况下，将神经网络输出的每个待处理遥感图像块中至少一个像素的分类概率信息映射到待识别遥感图像中，获得待识别遥感图像的至少一个像素的分类概率信息；
    根据本次获得的待识别遥感图像的至少一个像素的分类概率信息更新所述待识别遥感图像的相应像素的当前分类概率信息。
  12. 根据权利要求11所述的方法，其特征在于，所述根据本次获得的待识别遥感图像的至少一个像素的分类概率信息更新所述待识别遥感图像的相应像素的当前分类概率信息包括：
    计算本次获得的待识别遥感图像的至少一个像素的分类概率信息与待识别遥感图像中相应像素的当前分类概率信息的平均值；
    利用计算出的平均值更新待识别遥感图像中相应像素的当前分类概率信息。
  13. 根据权利要求12所述的方法，其特征在于，所述根据所述待处理遥感图像块中的像素的分类概率信息确定所述待识别遥感图像的识别结果包括：
    针对待识别遥感图像中的一像素而言，确定该像素的当前分类概率信息中的属于每个类别的概率值中的最大概率值，将该最大概率值以及该最大概率值对应的类别，作为该像素的识别结果。
  14. 根据权利要求2至13中任一项所述的方法，其特征在于，对所述神经网络进行训练的遥感图像样本的分辨率包括：根据缩小比例集合中的每个缩小值分别对遥感图像进行降分辨率处理，所获得的每个遥感图像的分辨率。
  15. 一种遥感图像识别装置，其特征在于，所述装置包括：
    降分辨率模块，用于对待识别遥感图像进行降分辨率处理，得到降分辨率处理后的遥感图像；
    遥感图像切分模块，用于从所述降分辨率处理后的遥感图像中切分出至少一个遥感图像块；
    选取图像块模块，用于从所述至少一个遥感图像块中确定待处理遥感图像块；
    分类处理模块，用于经由神经网络获得所述待处理遥感图像块中的像素的分类概率信息；
    确定识别结果模块，用于根据所述待处理遥感图像块中的像素的分类概率信息确定所述待识别遥感图像的识别结果。
  16. 根据权利要求15所述的装置，其特征在于，所述降分辨率模块用于：
    根据缩小比例集合中未被采用过的缩小值对待识别遥感图像进行降分辨率处理。
  17. 根据权利要求16所述的装置，其特征在于，所述确定识别结果模块用于：
    响应于缩小比例集合中不存在未被采用过的缩小值或者确定所述至少一个遥感图像块中不存在待处理遥感图像块，根据所述待处理遥感图像块中的像素的分类概率信息确定所述待识别遥感图像的识别结果。
  18. 根据权利要求16至17中任一项所述的装置，其特征在于，所述缩小比例集合包括：至少两个不同的缩小值，且较小的缩小值相对较大的缩小值先被采用。
  19. 根据权利要求15至18中任一项所述的装置，其特征在于，所述选取图像块模块用于：
    在第一次执行降分辨率处理的情况下，将所有遥感图像块均作为待处理遥感图像块。
  20. 根据权利要求15至19中任一项所述的装置，其特征在于：
    在非第一次执行降分辨率处理的情况下，选取图像块模块确定出的待处理遥感图像块的数量小于或等于遥感图像切分模块当前切分出的遥感图像块的数量，且降分辨率模块当前执行的降分辨率大于上一次执行的降分辨率。
  21. 根据权利要求20所述的装置，其特征在于，在非第一次执行降分辨率处理的情况下，选取图像块模块根据待识别遥感图像的像素的当前置信度从所述至少一个遥感图像块中确定待处理遥感图像块。
  22. 根据权利要求21所述的装置，其特征在于，所述选取图像块模块包括：
    第一单元，用于根据待识别遥感图像包括的多个像素中至少一个像素的当前分类概率信息确定所述至少一个像素的当前置信度；
    第二单元，用于统计遥感图像块所对应的待识别遥感图像区域中所述至少一个像素的当前置信度达到预定置信度的像素数量；
    第三单元，用于将所述达到预定置信度的像素数量不符合预定要求的遥感图像块确定为待处理遥感图像块。
  23. 根据权利要求22所述的装置，其特征在于，所述第三单元用于：针对一遥感图像块而言，在达到预定置信度的像素数量与该遥感图像块所对应的待识别遥感图像区域中所述至少一个像素的数量的比值小于预定比值的情况下，将该遥感图像块确定为待处理遥感图像块。
  24. 根据权利要求22至23中任一项所述的装置，其特征在于，所述选取图像块模块还包括：
    第四单元，用于在第一次执行降分辨率处理的情况下，将神经网络输出的每个待处理遥感图像块中至少一个像素的分类概率信息映射到待识别遥感图像中，获得待识别遥感图像的至少一个像素的分类概率信息；
    其中，所述待识别遥感图像的至少一个像素的分类概率信息被作为待识别遥感图像的至少一个像素的当前分类概率信息。
  25. 根据权利要求22至23中任一项所述的装置，其特征在于，所述选取图像块模块还包括：
    第五单元，用于在非第一次执行降分辨率处理的情况下，将神经网络输出的每个待处理遥感图像块中至少一个像素的分类概率信息映射到待识别遥感图像中，获得待识别遥感图像的至少一个像素的分类概率信息；
    第六单元，用于根据本次获得的待识别遥感图像的至少一个像素的分类概率信息更新所述待识别遥感图像的相应像素的当前分类概率信息。
  26. 根据权利要求25所述的装置，其特征在于：
    所述第六单元用于：
    计算本次获得的待识别遥感图像的至少一个像素的分类概率信息与待识别遥感图像中相应像素的当前分类概率信息的平均值；
    利用计算出的平均值更新待识别遥感图像中相应像素的当前分类概率信息。
  27. 根据权利要求26所述的装置，其特征在于，所述确定识别结果模块用于：
    针对待识别遥感图像中的一像素而言，确定该像素的当前分类概率信息中的属于每个类别的概率值中的最大概率值，将该最大概率值以及该最大概率值对应的类别，作为该像素的识别结果。
  28. 根据权利要求16至27中任一项所述的装置，其特征在于，所述装置还包括：
    训练模块，用于利用遥感图像样本对待训练的神经网络进行训练；
    其中，所述遥感图像样本的分辨率包括：根据缩小比例集合中的每个缩小值分别对遥感图像进行降分辨率处理，所获得的每个遥感图像的分辨率。
  29. 一种电子设备，包括：
    存储器，用于存储计算机程序；
    处理器，用于执行所述存储器中存储的计算机程序，且所述计算机程序被执行时，实现权利要求1-15中任一项所述的方法。
  30. 一种计算机可读存储介质，其上存储有计算机程序，该计算机程序被处理器执行时，实现上述权利要求1-15中任一项所述的方法。
  31. 一种计算机程序，包括计算机指令，当所述计算机指令在设备的处理器中运行时，实现上述权利要求1-15中任一项所述的方法。
PCT/CN2018/123807 2017-12-26 2018-12-26 遥感图像识别方法、装置、存储介质以及电子设备 WO2019129032A1 (zh)

Priority Applications (4)

Application Number Priority Date Filing Date Title
KR1020207021508A KR20200106909A (ko) 2017-12-26 2018-12-26 원격 감지 이미지 인식 방법, 장치, 저장 매체 및 전자 기기
SG11202006035RA SG11202006035RA (en) 2017-12-26 2018-12-26 Remote sensing image recognition method and apparatus, storage medium and electronic device
JP2020535252A JP7080978B2 (ja) 2017-12-26 2018-12-26 リモートセンシング画像認識方法、装置、記憶媒体及び電子機器
US16/909,291 US11074445B2 (en) 2017-12-26 2020-06-23 Remote sensing image recognition method and apparatus, storage medium and electronic device

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201711436470.2 2017-12-26
CN201711436470.2A CN108304775B (zh) 2017-12-26 2017-12-26 遥感图像识别方法、装置、存储介质以及电子设备

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US16/909,291 Continuation US11074445B2 (en) 2017-12-26 2020-06-23 Remote sensing image recognition method and apparatus, storage medium and electronic device

Publications (1)

Publication Number Publication Date
WO2019129032A1 true WO2019129032A1 (zh) 2019-07-04

Family

ID=62867455

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/123807 WO2019129032A1 (zh) 2017-12-26 2018-12-26 遥感图像识别方法、装置、存储介质以及电子设备

Country Status (6)

Country Link
US (1) US11074445B2 (zh)
JP (1) JP7080978B2 (zh)
KR (1) KR20200106909A (zh)
CN (1) CN108304775B (zh)
SG (1) SG11202006035RA (zh)
WO (1) WO2019129032A1 (zh)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110866493A (zh) * 2019-11-14 2020-03-06 生态环境部南京环境科学研究所 基于遥感影像的规模化生猪养殖场识别方法
CN111866427A (zh) * 2019-09-06 2020-10-30 阳光学院 一种基于psvm和crnn的遥感影像分类显示装置
CN112232249A (zh) * 2020-10-22 2021-01-15 中国科学院空天信息创新研究院 一种基于深度特征的遥感图像变化检测方法及装置
CN113553938A (zh) * 2021-07-19 2021-10-26 黑芝麻智能科技(上海)有限公司 安全带检测方法、装置、计算机设备和存储介质

Families Citing this family (15)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108304775B (zh) 2017-12-26 2021-06-11 北京市商汤科技开发有限公司 遥感图像识别方法、装置、存储介质以及电子设备
CN108711159B (zh) * 2018-03-30 2019-06-18 长安大学 基于Landsat8含雪影像的云检测方法
CN110189328B (zh) * 2019-06-11 2021-02-23 北华航天工业学院 一种卫星遥感图像处理系统及其处理方法
CN112560577A (zh) * 2020-11-13 2021-03-26 空间信息产业发展股份有限公司 一种基于语义分割的遥感图像地物分类方法
CN112579816B (zh) * 2020-12-29 2022-01-07 二十一世纪空间技术应用股份有限公司 遥感图像的检索方法、装置、电子设备及存储介质
CN112857268B (zh) 2021-01-21 2023-10-31 北京百度网讯科技有限公司 对象面积测量方法、装置、电子设备和存储介质
CN112818840A (zh) * 2021-01-29 2021-05-18 中国科学院微小卫星创新研究院 无人机在线检测系统和方法
CN113570536B (zh) * 2021-07-31 2022-02-01 中国人民解放军61646部队 基于cpu和gpu协同处理的全色和多光谱影像实时融合方法
CN114202691B (zh) * 2021-12-09 2022-10-04 广东省科学院广州地理研究所 基于遥感图像的水稻识别方法、装置以及设备
CN114676907B (zh) * 2022-01-17 2022-09-20 中国地质大学（北京） 一种区域地质灾害预警方法、装置、存储介质及设备
CN114663412B (zh) * 2022-04-01 2023-02-10 中国科学院地理科学与资源研究所 基于陆表水域生态红线的长连续遥感图像处理方法及装置
CN115019198B (zh) * 2022-08-08 2022-10-18 中国科学院空天信息创新研究院 高频次遥感观测影像迭代更新制图方法
CN115080925B (zh) * 2022-08-19 2022-11-11 北京数慧时空信息技术有限公司 基于区块链的遥感成果知识产权管理方法及系统
CN115346124B (zh) * 2022-09-02 2023-05-30 浙大城市学院 一种水稻制图的方法、装置、设备及可读存储介质
CN115641445B (zh) * 2022-12-23 2023-03-14 西南石油大学 一种非对称内卷积与Transformer相融合的遥感图像阴影检测方法

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101826157A (zh) * 2010-04-28 2010-09-08 华中科技大学 一种地面静止目标实时识别跟踪方法
CN102186069A (zh) * 2011-01-14 2011-09-14 王慧 一种保持量测性能的遥感影像数据压缩方法
CN103093182A (zh) * 2011-11-03 2013-05-08 北京师范大学 一种基于多敏感性策略的遥感影像层次分类识别方法
CN106548192A (zh) * 2016-09-23 2017-03-29 北京市商汤科技开发有限公司 基于神经网络的图像处理方法、装置和电子设备
CN108304775A (zh) * 2017-12-26 2018-07-20 北京市商汤科技开发有限公司 遥感图像识别方法、装置、存储介质以及电子设备

Family Cites Families (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
DE19955919C1 (de) * 1999-11-20 2001-05-31 Daimler Chrysler Ag Verfahren zur Erkennung von Objekten in Bildern auf der Bildpixelebene
JP2006011967A (ja) * 2004-06-28 2006-01-12 Fuji Xerox Co Ltd 文字認識装置、文字認識プログラム
US7359555B2 (en) * 2004-10-08 2008-04-15 Mitsubishi Electric Research Laboratories, Inc. Detecting roads in aerial images using feature-based classifiers
US8339470B2 (en) * 2009-11-30 2012-12-25 Indian Institute Of Technology Madras Method and system for generating a high resolution image
CN103150561A (zh) * 2013-03-19 2013-06-12 华为技术有限公司 人脸识别方法和设备
CN103310195B (zh) * 2013-06-09 2016-12-28 西北工业大学 车辆高分遥感图像的基于llc特征的弱监督识别方法
CN104463873B (zh) * 2014-12-10 2017-11-07 西安电子科技大学 一种基于局部均匀缩放的图像目标重定位方法
CN105046231A (zh) * 2015-07-27 2015-11-11 小米科技有限责任公司 人脸检测方法和装置
CN106778821B (zh) * 2016-11-25 2019-05-03 西安电子科技大学 基于slic和改进的cnn的极化sar图像分类方法

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101826157A (zh) * 2010-04-28 2010-09-08 华中科技大学 一种地面静止目标实时识别跟踪方法
CN102186069A (zh) * 2011-01-14 2011-09-14 王慧 一种保持量测性能的遥感影像数据压缩方法
CN103093182A (zh) * 2011-11-03 2013-05-08 北京师范大学 一种基于多敏感性策略的遥感影像层次分类识别方法
CN106548192A (zh) * 2016-09-23 2017-03-29 北京市商汤科技开发有限公司 基于神经网络的图像处理方法、装置和电子设备
CN108304775A (zh) * 2017-12-26 2018-07-20 北京市商汤科技开发有限公司 遥感图像识别方法、装置、存储介质以及电子设备

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111866427A (zh) * 2019-09-06 2020-10-30 阳光学院 一种基于psvm和crnn的遥感影像分类显示装置
CN110866493A (zh) * 2019-11-14 2020-03-06 生态环境部南京环境科学研究所 基于遥感影像的规模化生猪养殖场识别方法
CN112232249A (zh) * 2020-10-22 2021-01-15 中国科学院空天信息创新研究院 一种基于深度特征的遥感图像变化检测方法及装置
CN112232249B (zh) * 2020-10-22 2023-08-15 中国科学院空天信息创新研究院 一种基于深度特征的遥感图像变化检测方法及装置
CN113553938A (zh) * 2021-07-19 2021-10-26 黑芝麻智能科技(上海)有限公司 安全带检测方法、装置、计算机设备和存储介质
CN113553938B (zh) * 2021-07-19 2024-05-14 黑芝麻智能科技(上海)有限公司 安全带检测方法、装置、计算机设备和存储介质

Also Published As

Publication number Publication date
KR20200106909A (ko) 2020-09-15
CN108304775B (zh) 2021-06-11
CN108304775A (zh) 2018-07-20
JP7080978B2 (ja) 2022-06-06
US20200320273A1 (en) 2020-10-08
US11074445B2 (en) 2021-07-27
JP2021508123A (ja) 2021-02-25
SG11202006035RA (en) 2020-07-29

Similar Documents

Publication Publication Date Title
WO2019129032A1 (zh) 遥感图像识别方法、装置、存储介质以及电子设备
US11321593B2 (en) Method and apparatus for detecting object, method and apparatus for training neural network, and electronic device
US11334763B2 (en) Image processing methods, training methods, apparatuses, devices, media, and programs
US11657602B2 (en) Font identification from imagery
US11062453B2 (en) Method and system for scene parsing and storage medium
CN108229322B (zh) 基于视频的人脸识别方法、装置、电子设备及存储介质
CN108154222B (zh) 深度神经网络训练方法和系统、电子设备
US20190279014A1 (en) Method and apparatus for detecting object keypoint, and electronic device
CN110431560B (zh) 目标人物的搜索方法和装置、设备和介质
WO2020020146A1 (zh) 激光雷达稀疏深度图的处理方法、装置、设备及介质
WO2018010657A1 (zh) 结构化文本检测方法和系统、计算设备
CN108229591B (zh) 神经网络自适应训练方法和装置、设备、程序和存储介质
US11538244B2 (en) Extraction of spatial-temporal feature representation
CN109409288B (zh) 图像处理方法、装置、电子设备和存储介质
US11164004B2 (en) Keyframe scheduling method and apparatus, electronic device, program and medium
CN108154153B (zh) 场景分析方法和系统、电子设备
CN108229680B (zh) 神经网络系统、遥感图像识别方法、装置、设备及介质
CN115861400A (zh) 目标对象检测方法、训练方法、装置以及电子设备
CN113469025B (zh) 应用于车路协同的目标检测方法、装置、路侧设备和车辆
US11881044B2 (en) Method and apparatus for processing image, device and storage medium
CN115273148B (zh) 行人重识别模型训练方法、装置、电子设备及存储介质
CN113361536B (zh) 图像语义分割模型训练、图像语义分割方法及相关装置
CN111311604A (zh) 用于分割图像的方法和装置
CN115497112B (zh) 表单识别方法、装置、设备以及存储介质
US11837000B1 (en) OCR using 3-dimensional interpolation

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: 18897848

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 2020535252

Country of ref document: JP

Kind code of ref document: A

NENP Non-entry into the national phase

Ref country code: DE

ENP Entry into the national phase

Ref document number: 20207021508

Country of ref document: KR

Kind code of ref document: A

122 Ep: pct application non-entry in european phase

Ref document number: 18897848

Country of ref document: EP

Kind code of ref document: A1