WO2020062088A1 - 图像识别方法和设备、存储介质和处理器 - Google Patents

图像识别方法和设备、存储介质和处理器 Download PDF

Info

Publication number
WO2020062088A1
WO2020062088A1 PCT/CN2018/108461 CN2018108461W WO2020062088A1 WO 2020062088 A1 WO2020062088 A1 WO 2020062088A1 CN 2018108461 W CN2018108461 W CN 2018108461W WO 2020062088 A1 WO2020062088 A1 WO 2020062088A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
detection target
candidate
acquired image
screening
Prior art date
Application number
PCT/CN2018/108461
Other languages
English (en)
French (fr)
Inventor
万涛
赵永生
徐海青
吴立刚
陈是同
徐唯耀
梁翀
Original Assignee
安徽继远软件有限公司
国网信息通信产业集团有限公司
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 安徽继远软件有限公司, 国网信息通信产业集团有限公司 filed Critical 安徽继远软件有限公司
Priority to PCT/CN2018/108461 priority Critical patent/WO2020062088A1/zh
Publication of WO2020062088A1 publication Critical patent/WO2020062088A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition

Definitions

  • This application relates to image recognition technology, and in particular, to an image recognition method and device, a storage medium, and a processor.
  • the identification of the defects appearing in the collected images at least includes the identification of the defects of the transmission line body and the channel defects that appear in the collected images.
  • the body defects include at least defects related to transmission lines such as insulator self-explosion, broken wire strands, and bolts lacking pins; channel defects include at least foreign bodies (bird nests, kites, plastics, etc.) on the body, smoky mountains, and ultra-high mechanical operations. And so on.
  • the detection of defects on power lines is performed manually.
  • the recognition device collects images on the power line and performs defect detection on the entire image.
  • the defects on the power line may be displayed very small on the collected images.
  • the current recognition algorithm is relatively rough. For the small areas where the defects are displayed on the collected images, it is possible that The accuracy of the recognition results cannot be guaranteed.
  • embodiments of the present application provide an image recognition method and device, a storage medium, and a processor.
  • An embodiment of the present application provides an image recognition method.
  • the method includes:
  • the candidate area where the detection target may exist on the acquired image is processed to obtain a second screening result, where the second screening result includes at least Collect the location information of the image.
  • the second screening result further includes at least a type of a detection target existing on the candidate region.
  • the preliminary screening of a plurality of candidate regions in the collected image based on the feature image of the collected image to obtain a first screening result includes:
  • candidate regions that may have detection targets on the acquired images are obtained.
  • the processing the candidate area where the detection target may exist on the collected image based on the feature information of at least one detection target, to obtain a second screening result including:
  • the position information of the candidate area where the detection target exists on the acquired image is determined.
  • the method includes:
  • a characteristic image of the acquired image is obtained by at least one convolution process on the acquired image.
  • An embodiment of the present application provides an image recognition device, where the device includes:
  • a first acquisition unit configured to acquire a captured image
  • a second acquisition unit configured to acquire a characteristic image of the acquired image
  • a first screening unit is configured to perform preliminary screening on a plurality of candidate regions in the acquired image based on a characteristic image of the acquired image to obtain a first screening result, where the screening result includes at least a detection target that may exist on the acquired image Candidate area
  • a third obtaining unit configured to obtain characteristic information of at least one detection target
  • a second screening unit configured to process a candidate region where a detection target may exist on the collected image based on feature information of at least one detection target, to obtain a second screening result, where the second screening result is included in at least the candidate Location information of the detection target in the area where the image is collected.
  • the second screening unit is further configured to:
  • the first screening unit is further configured to:
  • candidate regions that may have detection targets on the acquired images are obtained.
  • the second screening unit is further configured to:
  • the position information of the candidate area where the detection target exists on the acquired image is determined.
  • the second obtaining unit is configured to:
  • a characteristic image of the acquired image is obtained by at least one convolution process on the acquired image.
  • An embodiment of the present application provides a storage medium including a stored program, and at least the foregoing image recognition method is executed when the foregoing program is run.
  • An embodiment of the present application provides a processor for running a program, and when the program is executed by the processor, at least the foregoing image recognition method is performed.
  • the image recognition method and device, the storage medium and the processor provided in the embodiments of the present application wherein the method includes: acquiring a captured image; acquiring a characteristic image of the acquired image; The candidate regions are preliminarily screened to obtain a first screening result, and the screening results include at least candidate regions that may have detection targets on the acquired image; obtain characteristic information of at least one detection target; and based on the at least one detection target, Feature information, processing a candidate area where a detection target may exist on the collected image to obtain a second screening result, where the second screening result includes at least position information of the detection target existing on the candidate area in the collected image .
  • the detection target is identified on the acquired image and the position of the target on the acquired image is identified. At least more accurate data is provided for the identification of the defect. Guaranteeing the accuracy of the recognition results can also improve the recognition accuracy. In addition, this solution is not only aimed at identifying whether a defect exists on the acquired image, but also the position of the defect on the acquired image. The recognition results are more abundant, which makes it easier for maintenance personnel to visualize the recognition results.
  • FIG. 1 is a schematic flowchart of implementation of a first embodiment of an image recognition method provided by this application;
  • FIG. 2 is a schematic flowchart of a second embodiment of an image recognition method provided by this application.
  • FIG. 3 is a schematic diagram of an implementation principle of an embodiment of the present application.
  • FIG. 4 is a schematic structural diagram of an embodiment of an image recognition device provided by this application.
  • the technical solutions of the embodiments of the present application can be applied to image recognition technology to realize the recognition of objects that should not appear in the collected images and / or the identification of the main body defects.
  • the embodiments of the present application can be applied to power transmission line technology, and can be used to identify whether there is a bird's nest, insulator self-explosion, wire broken strands, missing bolts, and channel smoke for the images collected from the transmission line The occurrence of wildfires, that is, whether there is a defect on the transmission line (whether the transmission line is an abnormal transmission line).
  • the technical solution of the embodiment of the present application can not only detect whether there is a defect in the acquired image, but also detect which defect is specific, and the position of the defect on the acquired image. It can be seen that, compared with the recognition algorithm in the related art, the recognition result in the embodiment of the present application is more accurate, and the content of the recognition result is more abundant.
  • the first embodiment of the image recognition method provided in the present application as shown in FIG. 1, the method includes:
  • Step 101 acquiring a captured image
  • the captured image may be captured by a camera in the image recognition device, or may be manually collected from the scene and entered into the image recognition device, or the image recognition device may be received from another outdoor shooting machine.
  • the outdoor shooting machine is responsible for collecting images of the transmission line scene.
  • the specific implementation of how the acquired image is acquired by the image recognition device is not described in detail, and any other conceivable content may also be included.
  • Step 102 Obtain a characteristic image of the collected image.
  • step 102 the image recognition device obtains a characteristic image of the acquired image by performing convolution processing on the acquired image at least once.
  • Step 103 Based on the feature image of the collected image, preliminary screening is performed on a plurality of candidate regions in the collected image to obtain a first screening result, where the screening result includes at least candidate regions where a detection target may exist on the collected image;
  • Step 104 acquiring characteristic information of at least one detection target
  • the detection target can be considered as all possible defect situations that should not appear in the acquired image. For example, bird nests on power transmission lines, smoky mountain fires, self-destructive insulators, broken wires, missing bolts, etc. Regarding these defects as different types of defects, considering that each defect is different in terms of size, shape, volume, color, etc., these features of these types of defects can be extracted in advance and stored in the image recognition device You can just read or call these feature information when you need it.
  • Step 105 Based on the feature information of at least one detection target, processing the candidate area where the detection target may exist on the collected image to obtain a second screening result, where the second screening result includes at least the The location information of the detection target in the acquired image.
  • the subject that performs steps 101 to 105 is an image recognition device.
  • the (first) screening is performed in step 103, and the (screening) processing is performed on the result of the previous screening in step 105, that is, the solution has been identified through at least two screening processes.
  • a detection target exists on the acquired image and the position of the detection target on the acquired image is identified.
  • This method of screening and processing at least twice provides at least more accurate data for the identification of defects, which can ensure the accuracy of the recognition results and improve the recognition accuracy.
  • this solution not only identifies whether a defect exists in the acquired image, but also identifies the position of the defect in the acquired image.
  • the recognition results are more abundant, which makes it easier for maintenance personnel to visualize the recognition results.
  • step 105 may also be replaced by step 205:
  • the candidate area where the detection target may exist on the collected image is processed to obtain a second screening result.
  • the second screening result includes at least the detection target existing on the candidate area.
  • the location information where the image is located and the type of the detection target It can be understood that, through the schemes of steps 101 to 104 and step 205, not only can there be a defect in the acquired image, but also the type of the defect in the acquired image, and the defect can be identified in the acquired image. On location. In this way, the recognition results are more abundant, and it is more convenient for maintenance personnel to visually recognize the recognition results when the recognition results are displayed (displayed).
  • performing preliminary screening on multiple candidate regions in the collected image to obtain a first screening result including:
  • candidate regions that may have detection targets on the acquired images are obtained.
  • an anchor point may represent a center point of a certain area size on the collected image (original image).
  • the area of a certain size is regarded as a candidate area of the acquired image.
  • the feature image of the collected image is input to the proposed network model (RPN) to obtain the anchor point information of each candidate region.
  • RPN proposed network model
  • a classifier is used to separate the foreground image and the background image in the image of the candidate region to obtain the collected image.
  • the background image can be seen as the image part that should appear in the collected image, for example, the image of the power transmission line in the collected image; and the foreground image can be seen as the part of the image that should not appear in the collected image, for example, the captured image is covered by the power transmission.
  • Images of foreign objects on the line such as bird nests, kites, etc.
  • this is a preliminary screening of candidate areas where detection targets may exist. Because the adopted RPN network model has strong robustness and robustness, the preliminary screening provides a relatively basic data base for the second screening process. Accuracy provides a certain basis for the scheme to guarantee or provide the accuracy of recognition.
  • processing the candidate area where the detection target may exist on the collected image to obtain a second screening result includes:
  • the position information of the candidate area where the detection target exists on the acquired image is determined.
  • the embodiments of the present application can identify defects existing on the transmission line body and identify channel defects. Based on the deep learning network model, specifically combined with the Faster-RCNN (Fast-CNN-based area detection) network model shown in Figure 3, how to use the Faster-RCNN network model in this solution to realize the defects included in the channel defects, Specifically, the identification of foreign objects such as bird nests, kites, and plastics on the output line, and the identification of the positions of the foreign objects on the collected images. Identification process of defects on the transmission line body and / or identification of other defects included in channel defects other than bird nests, kites, plastics and other foreign objects existing on the output line, and identification processes of the following foreign objects Roughly the same.
  • the Faster-RCNN Faster-RCNN-based area detection
  • composition of the Faster-RCNN network model can roughly include the following main parts: input layer, convolutional layer, excitation layer, pooling layer, fully connected layer, and RPN model.
  • input layer input layer
  • convolutional layer convolutional layer
  • excitation layer excitation layer
  • pooling layer pooling layer
  • RPN model RPN model
  • P * Q pixels are input to the input layer of the Faster-RCNN network model.
  • the input layer at least preprocesses the acquired images, including de-averaging. , Normalization, using principal component analysis algorithm PCA (principal component analysis) or singular value decomposition algorithm SVD (Singular value decomposition) for dimension reduction processing, etc.
  • PCA principal component analysis
  • SVD singular value decomposition
  • the input layer preprocesses the acquired image with P * Q pixels as the acquired image with M * N pixels.
  • the acquired image with pixels M * N is input to the convolution layer.
  • P, Q, M, and N are all positive integers, and P * Q is greater than M * N.
  • 13 convolutional layers are used.
  • the first convolutional layer among the 13 convolutional layers performs a convolution operation on the collected images input to the convolutional layer.
  • the first 1 The output of each convolution layer is used for convolution operation.
  • the collected image is subjected to the convolution operation of each convolution layer in the 13 convolution layers to obtain the characteristic image of the collected image, that is, the convolution layer is equivalent to the characteristic image of the collected image.
  • Feature extraction is performed to obtain a feature image.
  • the feature image passes through the excitation layer.
  • the number of layers in the excitation layer is usually the same as the number of layers in the convolution layer. In this solution, the excitation layer is also 13 layers.
  • the dimensionality of the feature image obtained through the excitation layer is high, and it needs to undergo dimensionality reduction processing through the pooling layer, so that the size of the matrix data corresponding to the feature image after dimensionality reduction is adapted to the computing capability of the RPN network.
  • the matrix data corresponding to the feature image output from the excitation layer is input to the RPN model.
  • a 3 * 3 size convolution kernel is used to perform a convolution operation on the matrix data corresponding to the feature image to obtain the anchor point information of each candidate region of the collected image, that is, the center point of each candidate region of the collected image. information.
  • the pixels of the acquired image processed by the input layer, the convolutional layer, the excitation layer, and the pooling layer are compressed compared to the original image.
  • an anchor point is equivalent to a certain size in the original image
  • a region such as 1 anchor point is equivalent to 16 * 16 pixels in the original image.
  • the softmax classifier in the RPN model is used to separate the foreground image and the background image on each candidate area of the collected image to obtain candidate areas where foreign objects may exist.
  • the role of the suggestion frame generation module in the RPN model is to obtain the size of the candidate area where foreign objects may exist.
  • the feature dimension changing module between the softmax classifier and the proposal generating module is equivalent to performing dimension reduction processing on the matrix data.
  • the input of the proposal generation module has two inputs, one is the pixel information of the original image, and the other is the coordinate information of each anchor point. From this, the coordinate position of each anchor point on the original image can be obtained. Combined with the size of the proposal, It can be seen that the approximate position of the candidate region that may have foreign objects after preliminary screening is on the original image. It can be considered that the generated proposal is a candidate region where foreign matter may exist. This concludes the RPN model.
  • the above-mentioned execution process in the RPN model can be regarded as a preliminary screening process. Because the adopted RPN network model has strong robustness and robustness, this preliminary screening is a second screening process performed by a candidate fully connected layer. Provides more accurate data, which provides a certain basis for the scheme to guarantee or provide recognition accuracy.
  • the output information of the RPN model is input to the ROI (Region Of Interest) pooling layer.
  • the ROI pooling layer processes the proposals of each size into the same-sized proposal and inputs them to the fully connected layer.
  • the processing of the fully connected layer is divided into two parts: one part (branch one) is based on the proposal input by the ROI pooling layer and the feature information of all foreign objects, and identifies candidates with foreign objects in the candidate regions where foreign objects may exist. Area, delete candidate areas where there are no foreign objects.
  • a softmax classifier is used to identify foreign objects in the candidate area, and to identify what type of foreign object is the foreign object, such as a bird's nest or a channel smoke mountain fire.
  • the other part (branch 2) is based on the approximate position of the candidate area where foreign objects may exist on the original image input from the ROI pooling layer and the recognition results of the remaining candidate areas with foreign objects obtained from branch one. The existence and foreign objects are obtained.
  • the position of the candidate region on the original image is trained on Smoothing Box Regression using Smooth L1 Loss (detection frame regression algorithm) to get the coordinates of foreign objects on the original image.
  • the type of foreign body and the position information of the candidate area on the original image and / or the position of the foreign body on the collected image can be obtained based on the combined training of Softmax Loss on classification probability and Smooth L1 Loss on border regression, and specific implementation Please refer to related description without specific description.
  • a Faster-RCNN network model such as how to use a convolutional layer to obtain a feature image, and how the excitation layer excites the output of the convolutional layer, please refer to the relevant description, which is not described here.
  • the aforementioned process includes an incentive layer.
  • the output of the convolutional layer can be directly input to the pooling layer without the participation of the incentive layer.
  • the Faster-RCNN network model has been used for at least two screening processes to identify the presence of foreign objects on the acquired image and the position of the foreign objects on the acquired image.
  • This solution using the Faster-RCNN network model to perform at least two screening processes provides more accurate data for foreign object recognition, which can ensure the accuracy of the recognition results and improve the recognition accuracy.
  • this solution not only identifies whether a foreign object is present on the acquired image, but also identifies the position of the foreign object on the acquired image.
  • the recognition results are more abundant, which is more convenient for maintenance personnel to visually recognize the recognition results, it is also convenient for maintenance personnel to quickly locate the position of foreign objects on the input line, and it is also convenient for maintenance personnel to handle foreign objects.
  • an embodiment of the present application provides an image recognition device.
  • the device includes: a first obtaining unit 41, a second obtaining unit 42, a first filtering unit 43, and a third obtaining A unit 44 and a second screening unit 45;
  • a first acquiring unit 41 configured to acquire a captured image
  • a second acquiring unit 42 configured to acquire a characteristic image of the acquired image
  • a first screening unit 43 is configured to perform preliminary screening on a plurality of candidate regions in the acquired image based on a characteristic image of the acquired image, to obtain a first screening result, where the screening result includes at least detection that may exist on the acquired image Target candidate area;
  • a third obtaining unit 44 configured to obtain characteristic information of at least one detection target
  • a second screening unit 45 is configured to process a candidate region where a detection target may exist on the collected image based on the feature information of at least one detection target, to obtain a second screening result, where the second screening result includes at least Location information of the detection target on the candidate area where the image is collected.
  • the second screening unit 45 is further configured to:
  • the first screening unit 43 is further configured to:
  • candidate regions that may have detection targets on the acquired images are obtained.
  • the second screening unit 45 is further configured to:
  • the position information of the candidate area where the detection target exists on the acquired image is determined.
  • the second obtaining unit 42 is configured to:
  • a characteristic image of the acquired image is obtained by at least one convolution process on the acquired image.
  • the image recognition device in the embodiment of the present application has a principle similar to that of the foregoing method. Therefore, the implementation process and implementation principle of the device can refer to the implementation of the foregoing method. The process and implementation principles are described, and the duplicates are not repeated here.
  • An embodiment of the present application provides a storage medium including a stored program, and when the program runs, at least the foregoing image recognition method and the foregoing image recognition method shown in FIG. 1 or FIG. 2 are executed.
  • the storage medium may be a memory.
  • the memory may be implemented by any type of volatile or non-volatile storage device, or a combination thereof.
  • the non-volatile memory may be a read-only memory (ROM, Read Only Memory), a programmable read-only memory (PROM, Programmable Read-Only Memory), or an erasable programmable read-only memory (EPROM, Erasable Programmable Read- Only Memory), Electrically Erasable Programmable Read-Only Memory (EEPROM), Ferromagnetic Random Access Memory (FRAM), Flash Memory, Flash Surface Memory , Compact disc, or read-only compact disc (CD-ROM, Compact Disc-Read-Only Memory);
  • the magnetic surface memory can be a disk memory or a tape memory.
  • the volatile memory may be random access memory (RAM, Random Access Memory), which is used as an external cache.
  • RAM random access memory
  • RAM Random Access Memory
  • many forms of RAM are available, such as Static Random Access Memory (SRAM, Static Random Access Memory), Synchronous Static Random Access Memory (SSRAM, Static Random Access, Memory), Dynamic Random Access DRAM (Dynamic Random Access Memory), Synchronous Dynamic Random Access Memory (SDRAM), Double Data Rate Synchronous Dynamic Random Access Memory (DDRSDRAM, Double Data Rate Synchronous Dynamic Random Access Memory), enhanced Type Synchronous Dynamic Random Access Memory (ESDRAM, Enhanced Random Dynamic Access Memory), Synchronous Link Dynamic Random Access Memory (SLDRAM, SyncLink Dynamic Random Access Memory), Direct Memory Bus Random Access Memory (DRRAM, Direct Rambus Random Access Memory) ).
  • SRAM Static Random Access Memory
  • SSRAM Synchronous Static Random Access Memory
  • SDRAM Synchronous Dynamic Random Access Memory
  • DDRSDRAM Double Data Rate Synchronous Dynamic Random Access Memory
  • An embodiment of the present application further provides a processor for running a program, and when the program is executed by the processor, at least the foregoing image recognition method, or the foregoing image recognition method shown in FIG. 1 or FIG. 2 is executed.
  • the processor may be a central processing unit (CPU, Central Processing Unit), or digital signal processing (DSP, Digital Signal Processor), or a microprocessor (MPU, Micro Processor or Unit), or field programmable Gate array (FPGA, Field Programmable Gate Array) and so on.
  • CPU Central Processing Unit
  • DSP Digital Signal Processor
  • MPU Micro Processor or Unit
  • FPGA Field Programmable Gate Array
  • the disclosed device and method may be implemented in other ways.
  • the device embodiments described above are only schematic.
  • the division of the unit is only a logical function division.
  • there may be another division manner such as multiple units or components may be combined, or Can be integrated into another system, or some features can be ignored or not implemented.
  • the displayed or discussed components are coupled, or directly coupled, or communicated with each other through some interfaces.
  • the indirect coupling or communication connection of the device or unit may be electrical, mechanical, or other forms. of.
  • the units described above as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, which may be located in one place or distributed to multiple network units; Some or all of the units may be selected according to actual needs to achieve the objective of the solution of this embodiment.
  • each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may be separately used as a unit, or two or more units may be integrated into one unit; the above integration
  • the unit can be implemented in the form of hardware, or in the form of hardware plus software functional units.
  • the foregoing program may be stored in a computer-readable storage medium.
  • the program is executed, the program is executed.
  • the foregoing storage medium includes: a mobile storage device, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, etc.
  • ROM read-only memory
  • RAM random access memory
  • magnetic disk or an optical disk etc.
  • the above-mentioned integrated unit of the present invention is implemented in the form of a software functional module and sold or used as an independent product, it may also be stored in a computer-readable storage medium.
  • the computer software product is stored in a storage medium and includes several instructions for A computer device (which may be a personal computer, a server, or a network device, etc.) is caused to execute all or part of the methods described in the embodiments of the present invention.
  • the foregoing storage medium includes: various types of media that can store program codes, such as a mobile storage device, a ROM, a RAM, a magnetic disk, or an optical disc.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • Image Analysis (AREA)

Abstract

一种图像识别方法和设备、存储介质和处理器，其中，所述方法包括：获取采集图像（S101）；获取采集图像的特征图像（S102）；基于采集图像的特征图像，对采集图像中的多个候选区域进行初步筛选，得到第一筛选结果，所述筛选结果中至少包括所述采集图像上可能存在有检测目标的候选区域（S103）；获取至少一种检测目标的特征信息（S104）；基于至少一种检测目标的特征信息，对所述采集图像上可能存在有检测目标的候选区域进行处理，得到第二筛选结果，所述第二筛选结果至少包括在候选区域上存在的检测目标在采集图像所处的位置信息（S105）。

Description

图像识别方法和设备、存储介质和处理器 技术领域
本申请涉及图像识别技术，具体涉及到图像识别方法和设备、存储介质和处理器。
背景技术
在图像处理技术中，针对采集到的图像，可实现对采集图像中本不属于应在该采集图像中出现的物体和/或采集图像中出现的本体的缺陷的识别。例如，以在输电线路上采集到的采集图像，对该采集图像中出现的缺陷的识别至少包括对采集图像中出现的输电线路本体的缺陷识别和通道缺陷的识别。其中，本体缺陷至少包括绝缘子自爆、导线断股、螺栓缺销子等与输电线路有关的缺陷；通道缺陷至少包括本体上出现异物(鸟巢、风筝、塑料等)、烟雾山火、超高机械作业等情况。
在识别初期阶段，输电线路上的缺陷的检测通过人工进行。随着技术的发展，从人工检测发展到及智能识别。在智能识别中，识别设备采集输电线路上的图像，并对整张图像进行缺陷检测。考虑到在实际应用中，输电线路上存在的缺陷之处可能在采集图像上显示得很小，目前的识别算法检测得较为粗糙，针对缺陷之处在采集图像上显示的很小的情况，可能无法保证识别结果的正确性。
发明内容
为解决现有存在的技术问题，本申请实施例提供一种图像识别方法和设备、存储介质和处理器。
本申请实施例的技术方案是这样实现的：
本申请实施例提供一种图像识别方法，所述方法包括：
获取采集图像；
获取采集图像的特征图像；
基于采集图像的特征图像，对采集图像中的多个候选区域进行初步筛选，得到第一筛选结果，所述筛选结果中至少包括所述采集图像上可能存在有检测目标的候选区域；
获取至少一种检测目标的特征信息；
基于至少一种检测目标的特征信息，对所述采集图像上可能存在有检测目标的候选区域进行处理，得到第二筛选结果，所述第二筛选结果至少包括在候选区域上存在的检测目标在采集图像所处的位置信息。
上述方案中，所述第二筛选结果至少还包括在候选区域上存在的检测目标的种类。
上述方案中，所述基于采集图像的特征图像，对采集图像中的多个候选区域进行初步筛选，得到第一筛选结果，包括：
基于采集图像的特征图像，得到采集图像各个候选区域的锚点信息；
基于各个候选区域的锚点信息，对采集图像的各个候选区域上的前景图像和背景图像进行分离；
基于分离结果，得到采集图像上可能存在有检测目标的候选区域。
上述方案中，所述基于至少一种检测目标的特征信息，对所述采集图像上可能存在有检测目标的候选区域进行处理，得到第二筛选结果，包括：
基于至少一种检测目标的特征信息，对所述采集图像上可能存在有检测目标的候选区域中的各个候选区域进行检测目标的识别，确定存在有检测目标的候选区域；
获取存在有检测目标的候选区域在采集图像上的位置信息；
依据存在有检测目标的候选区域在采集图像上的位置信息，确定在存在有检测目标的候选区域中的检测目标在采集图像上所处的位置信息。
上述方案中，所述方法包括：
至少通过对所述采集图像的至少一次卷积处理而得到所述采集图像的特征图像。
本申请实施例提供一种图像识别设备，所述设备包括：
第一获取单元，用于获取采集图像；
第二获取单元，用于获取采集图像的特征图像；
第一筛选单元，用于基于采集图像的特征图像，对采集图像中的多个候选区域进行初步筛选，得到第一筛选结果，所述筛选结果中至少包括所述采集图像上可能存在有检测目标的候选区域；
第三获取单元，用于获取至少一种检测目标的特征信息；
第二筛选单元，用于基于至少一种检测目标的特征信息，对所述采集图像上可能存在有检测目标的候选区域进行处理，得到第二筛选结果，所述第二筛选结果至少包括在候选区域上存在的检测目标在采集图像所处的位置信息。
上述方案中，所述第二筛选单元，还用于：
筛选出在候选区域上存在的检测目标的种类。
上述方案中，所述第一筛选单元，还用于：
基于采集图像的特征图像，得到采集图像各个候选区域的锚点信息；
基于各个候选区域的锚点信息，对采集图像的各个候选区域上的前景图像和背景图像进行分离；
基于分离结果，得到采集图像上可能存在有检测目标的候选区域。
上述方案中，所述第二筛选单元，还用于：
基于至少一种检测目标的特征信息，对所述采集图像上可能存在有检 测目标的候选区域中的各个候选区域进行检测目标的识别，确定存在有检测目标的候选区域；
获取存在有检测目标的候选区域在采集图像上的位置信息；
依据存在有检测目标的候选区域在采集图像上的位置信息，确定在存在有检测目标的候选区域中的检测目标在采集图像上所处的位置信息。
上述方案中，所述第二获取单元，用于：
至少通过对所述采集图像的至少一次卷积处理而得到所述采集图像的特征图像。
本申请实施例提供一种存储介质，该存储介质包括存储的程序，其中，上述程序运行时至少执行前述的图像识别方法。
本申请实施例提供一种处理器，该处理器用于运行程序，其中，该程序被处理器运行时至少执行前述的图像识别方法。
本申请实施例提供的图像识别方法和设备、存储介质和处理器，其中，所述方法包括：获取采集图像；获取采集图像的特征图像；基于采集图像的特征图像，对采集图像中的多个候选区域进行初步筛选，得到第一筛选结果，所述筛选结果中至少包括所述采集图像上可能存在有检测目标的候选区域；获取至少一种检测目标的特征信息；基于至少一种检测目标的特征信息，对所述采集图像上可能存在有检测目标的候选区域进行处理，得到第二筛选结果，所述第二筛选结果至少包括在候选区域上存在的检测目标在采集图像所处的位置信息。
本申请实施例中，经过至少两次筛选处理而识别出采集图像上存在有检测目标并识别出了目标在采集图像上的位置，至少为缺陷之处的识别提供了更为精准的数据，可保证识别结果的准确性，也可提高识别准确率。此外，本方案不仅针对采集图像上识别出是否存在缺陷，还可以识别出该缺陷在采集图像上的位置。识别结果更为丰富，更加便于维护人员对识别 结果的直观观看。
附图说明
为了更清楚地说明本申请实施例或现有技术中的技术方案，下面将对实施例或现有技术描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本发明的实施例，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据提供的附图获得其他的附图。
图1为本申请提供的图像识别方法的第一实施例的实现流程示意图；
图2为本申请提供的图像识别方法的第二实施例的实现流程示意图；
图3为本申请实施例的实现原理示意图；
图4为本申请提供的图像识别设备实施例的组成结构示意图。
具体实施方式
为使本申请的目的、技术方案和优点更加清楚明白，下面将结合本申请实施例中的附图，对本申请实施例中的技术方案进行清楚、完整地描述，显然，所描述的实施例仅仅是本发明一部分实施例，而不是全部的实施例。基于本发明中的实施例，本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例，都属于本发明保护的范围。在不冲突的情况下，本申请中的实施例及实施例中的特征可以相互任意组合。在附图的流程图示出的步骤可以在诸如一组计算机可执行指令的计算机系统中执行。并且，虽然在流程图中示出了逻辑顺序，但是在某些情况下，可以以不同于此处的顺序执行所示出或描述的步骤。
本领域技术人员应该而知，本申请实施例的技术方案可应用于图像识别技术中，实现对本不应出现在采集图像中的物体的识别和/或本体缺陷的识别。例如，本申请实施例可应用在输电线路技术中，用于针对从输电线路上采集到的图像可识别出该采集图像上是否存在鸟巢、绝缘子自爆、导 线断股、螺栓缺销子、通道烟雾山火等情况的发生，也即输电线路上是否存在缺陷(输电线路是否为非正常输电线路)。
可以理解，本申请实施例的技术方案不仅可以检测到在采集图像上是否存在缺陷之处，还可以检测到具体是哪种缺陷、以及该缺陷之处在采集图像上的位置。可见，与相关技术中的识别算法相比，本申请实施例的识别结果更加准确，识别结果内容更加丰富。
本申请提供的图像识别方法的第一实施例，如图1所示，所述方法包括：
步骤101：获取采集图像；
在步骤101中，采集图像可以为通过图像识别设备中的摄像头而拍摄到的，还可以是人工从现场采集并录入至图像识别设备中的，也可以是图像识别设备从其它室外拍摄机器中接收到的。其中，室外拍摄机器负责对输电线路现场的图像进行采集。此处，对于图像识别设备如何获取到的采集图像的具体实现方式不做重点说明，还可以包括其它任何想到的内容。
步骤102：获取采集图像的特征图像；
在步骤102中，图像识别设备通过至少一次对采集图像进行卷积处理而得到采集图像的特征图像。
步骤103：基于采集图像的特征图像，对采集图像中的多个候选区域进行初步筛选，得到第一筛选结果，所述筛选结果中至少包括所述采集图像上可能存在有检测目标的候选区域；
步骤104：获取至少一种检测目标的特征信息；
在步骤104中，检测目标可以认为是本不该在采集图像中出现的所有可能出现的缺陷情况。例如，输电线路上的鸟巢、通道烟雾山火、绝缘子自爆、导线断股、螺栓缺销子等。视这些缺陷情况为不同种类的缺陷，考虑到每种缺陷本身在大小、形状、体积、色彩等几种方面均存在不同，可 以预先对这几种缺陷的这些特征进行提取并存储到图像识别设备中，待到需要时读取或调用这些特征信息即可。
步骤105：基于至少一种检测目标的特征信息，对所述采集图像上可能存在有检测目标的候选区域进行处理，得到第二筛选结果，所述第二筛选结果至少包括在候选区域上存在的检测目标在采集图像所处的位置信息。
上述内容中，执行步骤101～105的主体为图像识别设备。
上述内容中，在步骤103中进行了(第)一次筛选，在步骤105中再前一次筛选的结果上又进行了一次(筛选)处理，也即本方案经过了至少两次筛选处理而识别出采集图像上存在有检测目标并识别出了检测目标在采集图像上的位置。这种通过至少两次筛选处理的方式，至少为缺陷之处的识别提供了更为精准的数据，可保证识别结果的准确性，也可提高识别准确率。此外，本方案不仅识别出了在采集图像上是否存在缺陷之处，还可以识别出该缺陷之处在采集图像上的位置。识别结果更为丰富，更加便于维护人员对识别结果的直观观看。
在一个可选的实施例中，如图2所示，步骤105还可以由步骤205来代替：
基于至少一种检测目标的特征信息，对所述采集图像上可能存在有检测目标的候选区域进行处理，得到第二筛选结果，所述第二筛选结果至少包括在候选区域上存在的检测目标在采集图像所处的位置信息以及所述检测目标的种类。可以理解，通过步骤101～104、步骤205的方案，不仅可以识别出了在采集图像上是否存在缺陷之处，还可以在采集图像上的缺陷之处的种类、以及识别出该缺陷在采集图像上的位置。如此，识别结果更为丰富，在进行识别结果的展示(显示)时，更加便于维护人员对识别结果的直观观看。
在一个可选的实施例中，所述基于采集图像的特征图像，对采集图像中的多个候选区域进行初步筛选，得到第一筛选结果，包括：
基于采集图像的特征图像，得到采集图像各个候选区域的锚点信息；
基于各个候选区域的锚点信息，对采集图像的各个候选区域上的前景图像和背景图像进行分离；
基于分离结果，得到采集图像上可能存在有检测目标的候选区域。
在此处可选的实施例中，可以理解，一个锚点可代表采集图像(原始图像)上一定区域大小的中心点。视所述一定大小的区域为采集图像的候选区域。将采集图像的特征图像输入至建议网络模型(RPN)得到各个候选区域的锚点信息，在RPN网络模型中利用分类器实现候选区域的图像中的前景图像和背景图像的分离，得到采集图像上可能存在有检测目标的候选区域。其中，背景图像可以看成采集图像中应该出现的图像部分，例如，采集图像中输电线路的图像；而前景图像可以看成本不该在采集图像中出现的图像部分，例如采集图像中覆盖在输电线路上的异物如鸟巢、风筝等的图像。此处，是对可能存在有检测目标的候选区域的初步筛选，由于所采用的RPN网络模型具有很强的健壮性和鲁棒性，所以该初步筛选为第二次筛选处理提供的数据基础较为准确，为本方案能够保证或提供识别准确率提供了一定的基础。
在一个可选的实施例中，所述基于至少一种检测目标的特征信息，对所述采集图像上可能存在有检测目标的候选区域进行处理，得到第二筛选结果，包括：
基于至少一种检测目标的特征信息，对所述采集图像上可能存在有检测目标的候选区域中的各个候选区域进行检测目标的识别，确定存在有检测目标的候选区域；
获取存在有检测目标的候选区域在采集图像上的位置信息；
依据存在有检测目标的候选区域在采集图像上的位置信息，确定在存在有检测目标的候选区域中的检测目标在采集图像上所处的位置信息。
在此处可选的实施例中，可以理解，对于可能存在有检测目标的候选区域，先确定肯定存在检测目标的候选区域，然后确定肯定存在检测目标的候选区域在采集图像上的位置信息，并基于该位置信息确定检测目标在采集图像上所处的位置。这种识别出缺陷之处在采集图像上的位置的方案，可方便维护人员对输电线路上的缺陷情况的定位，以避免由于缺陷的存在使得输电线路无法正常工作的问题。
本领域技术人员应该而知，在前述的几种可选实施例中，可以互相组合使用，以尽可能覆盖本申请技术方案的各种情况。
本领域技术人员应该理解，本申请实施例能够实现对输电线路本体上存在的缺陷进行识别和对通道缺陷进行识别。下面基于深度学习网络模型、具体是结合图3所示的Faster-RCNN(快速-基于CNN的区域检测)网络模型，对本方案中如何基于Faster-RCNN网络模型，实现对通道缺陷中包括的缺陷、具体是输出线路上存在的鸟巢、风筝、塑料等异物的识别、以及异物在采集图像上所处位置的识别。对输电线路本体上的缺陷的识别过程和/或对通道缺陷中包括的除了输出线路上存在的鸟巢、风筝、塑料等异物的识别之外的其它缺陷的识别过程与对下述异物的识别过程大致相同。
在Faster-RCNN网络模型的组成可大致包括以下几个主要部分：输入层、卷积层、激励层、池化层、全连接层和RPN模型。对于各部分的作用的说明请参见如下描述。
针对从输电线路上采集到的采集图像，为P*Q个像素点(原始图像)，将原始图像输入至Faster-RCNN网络模型的输入层，输入层至少对采集图像进行预处理，包括去均值、归一化、利用主成分分析算法PCA(principal component analysis)或奇异值分解算法SVD(Singular value decomposition) 进行降维处理等。输入层将输入为P*Q个像素点的采集图像预处理为像素点为M*N的采集图像。像素点为M*N的采集图像输入至卷积层。其中，P、Q、M和N均为正整数，且P*Q大于M*N。本方案中采用13个卷积层，13个卷积层中的第1个卷积层执行对输入至卷积层的采集图像进行卷积运算，从第2个卷积层开始，针对前1个卷积层的输出进行卷积运算，采集图像经过13个卷积层中各个卷积层逐一的卷积运算可得到采集图像的特征图像，也即卷积层相当于对采集图像的特征图像进行特征提取得到特征图像。特征图像经过激励层，激励层的层数通常取值为与卷积层的层数相同，本方案中取激励层也为13层。本领域技术人员可以理解：所谓的激励，实际上是对卷积层的输出结果做一次非线性映射或线性映射。如果激励层不用激励函数(其实就相当于激励函数是f(x)＝x)，这种情况下，每一激励层的输出都是相应卷积层输入的线性函数。
经过激励层得出的特征图像的维度较高，需要经过池化层进行降维处理，以使得降维之后的特征图像对应的矩阵数据的大小适配RPN网络的运算能力。将从激励层输出的特征图像对应的矩阵数据输入至RPN模型。在RPN模型中，采用3*3大小的卷积核对特征图像对应的矩阵数据进行卷积运算，得到采集图像的各个候选区域的锚点信息，也即得到采集图像的各个候选区域的中心点的信息。其中，可以这样理解，经过输入层、卷积层、激励层、池化层处理后的采集图像的像素相比于原始图像被压缩，在RPN模型中1个锚点相当于原始图像中一定大小的区域如1个锚点相当于原始图像中的16*16个像素点。锚点信息经过特征变维之后，经过softmax分类器，在RPN模型中的softmax分类器其作用是对采集图像的各个候选区域上的前景图像和背景图像进行分离，得到可能存在有异物的候选区域，RPN模型中的建议框生成模块的作用即是得到可能存在有异物的候选区域的大小。softmax分类器和建议框(proposal)生成模块之间的特征变维模块相 当于对矩阵数据进行降维处理。proposal生成模块的输入有二个输入，一个输入是原始图像的像素信息，一个输入是各个锚点的坐标信息，由此可以得到各个锚点在原始图像上的坐标位置，结合proposal的大小，就可知经过初步筛选出的可能存在有异物的候选区域在原始图像上的大致位置。可认为生成的proposal即为可能存在有异物的候选区域。至此，RPN模型结束。
上述在RPN模型中执行过程可视为初步筛选过程，由于所采用的RPN网络模型具有很强的健壮性和鲁棒性，所以该初步筛选为候选的全连接层执行的第二次筛选处理过程提供了较为准确的数据，为本方案能够保证或提供识别准确率提供了一定的基础。
RPN模型的输出信息输入至ROI(感兴趣区域，Region Of Interest)池化层。由于从RPN模型得到的各个proposal的大小并不相同，为方便后续全连接层的计算，ROI池化层将各个大小的proposal处理为相同大小的proposal，并输入至全连接层。全连接层的处理分为两个部分：其中一个部分(分支一)是基于ROI池化层输入的proposal和所有异物的特征信息，在可能存在有异物的候选区域中识别出存在有异物的候选区域，删除不存在异物的候选区域。并利用softmax分类器对候选区域上存在的异物进行识别，识别出该异物是何种类型的异物，如是鸟巢或通道烟雾山火等。另一部分(分支二)是基于ROI池化层输入的可能存在有异物的候选区域在原始图像上的大致位置和分支一得到的保留下来的存在有异物的候选区域的识别结果，得到存在与异物的候选区域在原始图像上的位置，利用Smooth L1 Loss(探测边框回归算法)对边框回归(Bounding box regression)进行训练，得到异物在原始图像上的坐标。
其中，异物的种类、以及候选区域在原始图像上的位置信息和/或异物在采集图像上所处的位置可以基于Softmax Loss对分类概率和Smooth L1  Loss对边框回归的联合训练而得到，具体实现不做具体描述请参见相关说明。对于利用Faster-RCNN网络模型实现本方案的细节描述，如如何利用卷积层得到特征图像，激励层如何对卷积层的输出进行激励的具体实现过程请参见相关说明，此处不赘述。前述流程是包括激励层的，此外还可以卷积层的输出直接输入至池化层，不需要激励层的参与。
由上述方案可知，利用Faster-RCNN网络模型经过了至少两次筛选处理而识别出采集图像上存在有异物并识别出了异物在采集图像上的位置。这种利用Faster-RCNN网络模型执行至少两次筛选处理的方案，为异物的识别提供了更为精准的数据，可保证识别结果的准确性，也可提高识别准确率。此外，本方案不仅识别出了在采集图像上是否存在异物，还可以识别出该异物在采集图像上的位置。识别结果更为丰富，更加便于维护人员对识别结果的直观观看，也方便维护人员快速定位出异物在输入线路上的位置，也方便维护人员对异物的处理。
基于前述的图像识别方法，本申请实施例提供一种图像识别设备，如图4所示，所述设备包括：第一获取单元41、第二获取单元42、第一筛选单元43、第三获取单元44、第二筛选单元45；其中，
第一获取单元41，用于获取采集图像；
第二获取单元42，用于获取采集图像的特征图像；
第一筛选单元43，用于基于采集图像的特征图像，对采集图像中的多个候选区域进行初步筛选，得到第一筛选结果，所述筛选结果中至少包括所述采集图像上可能存在有检测目标的候选区域；
第三获取单元44，用于获取至少一种检测目标的特征信息；
第二筛选单元45，用于基于至少一种检测目标的特征信息，对所述采集图像上可能存在有检测目标的候选区域进行处理，得到第二筛选结果，所述第二筛选结果至少包括在候选区域上存在的检测目标在采集图像所处 的位置信息。
其中，所述第二筛选单元45，还用于：
筛选出在候选区域上存在的检测目标的种类。
其中，所述第一筛选单元43，还用于：
基于采集图像的特征图像，得到采集图像各个候选区域的锚点信息；
基于各个候选区域的锚点信息，对采集图像的各个候选区域上的前景图像和背景图像进行分离；
基于分离结果，得到采集图像上可能存在有检测目标的候选区域。
其中，所述第二筛选单元45，还用于：
基于至少一种检测目标的特征信息，对所述采集图像上可能存在有检测目标的候选区域中的各个候选区域进行检测目标的识别，确定存在有检测目标的候选区域；
获取存在有检测目标的候选区域在采集图像上的位置信息；
依据存在有检测目标的候选区域在采集图像上的位置信息，确定在存在有检测目标的候选区域中的检测目标在采集图像上所处的位置信息。
其中，所述第二获取单元42，用于：
至少通过对所述采集图像的至少一次卷积处理而得到所述采集图像的特征图像。
需要说明的是，为实现上述图像识别方法，本申请实施例的图像识别设备，由于该设备解决问题的原理与前述的方法相似，因此，设备的实施过程及实施原理均可以参见前述方法的实施过程及实施原理描述，重复之处不再赘述。
本申请实施例提供一种存储介质，该存储介质包括存储的程序，其中，上述程序运行时至少执行前述的图像识别方法、如前述的图1或图2所示的图像识别方法。
该存储介质可以为存储器。存储器可以由任何类型的易失性或非易失性存储设备、或者它们的组合来实现。其中，非易失性存储器可以是只读存储器(ROM，Read Only Memory)、可编程只读存储器(PROM，Programmable Read-Only Memory)、可擦除可编程只读存储器(EPROM，Erasable Programmable Read-Only Memory)、电可擦除可编程只读存储器(EEPROM，Electrically Erasable Programmable Read-Only Memory)、磁性随机存取存储器(FRAM，Ferromagnetic Random Access Memory)、快闪存储器(Flash Memory)、磁表面存储器、光盘、或只读光盘(CD-ROM，Compact Disc Read-Only Memory)；磁表面存储器可以是磁盘存储器或磁带存储器。易失性存储器可以是随机存取存储器(RAM，Random Access Memory)，其用作外部高速缓存。通过示例性但不是限制性说明，许多形式的RAM可用，例如静态随机存取存储器(SRAM，Static Random Access Memory)、同步静态随机存取存储器(SSRAM，Synchronous Static Random Access Memory)、动态随机存取存储器(DRAM，Dynamic Random Access Memory)、同步动态随机存取存储器(SDRAM，Synchronous Dynamic Random Access Memory)、双倍数据速率同步动态随机存取存储器(DDRSDRAM，Double Data Rate Synchronous Dynamic Random Access Memory)、增强型同步动态随机存取存储器(ESDRAM，Enhanced Synchronous Dynamic Random Access Memory)、同步连接动态随机存取存储器(SLDRAM，SyncLink Dynamic Random Access Memory)、直接内存总线随机存取存储器(DRRAM，Direct Rambus Random Access Memory)。本申请实施例描述的存储器旨在包括但不限于这些和任意其它适合类型的存储器。
本申请实施例还提供一种处理器，该处理器用于运行程序，其中，该程序被处理器运行时至少执行前述的图像识别方法、如前述的图1或图2 所示的图像识别方法。
在实际应用中，所述处理器可以为中央处理单元(CPU，Central Processing Unit)、或数字信号处理(DSP，Digital Signal Processor)、或微处理器(MPU，Micro Processor Unit)、或现场可编程门阵列(FPGA，Field Programmable Gate Array)等。
在本申请所提供的几个实施例中，应该理解到，所揭露的设备和方法，可以通过其它的方式实现。以上所描述的设备实施例仅仅是示意性的，例如，所述单元的划分，仅仅为一种逻辑功能划分，实际实现时可以有另外的划分方式，如：多个单元或组件可以结合，或可以集成到另一个系统，或一些特征可以忽略，或不执行。另外，所显示或讨论的各组成部分相互之间的耦合、或直接耦合、或通信连接可以是通过一些接口，设备或单元的间接耦合或通信连接，可以是电性的、机械的或其它形式的。
上述作为分离部件说明的单元可以是、或也可以不是物理上分开的，作为单元显示的部件可以是、或也可以不是物理单元，即可以位于一个地方，也可以分布到多个网络单元上；可以根据实际的需要选择其中的部分或全部单元来实现本实施例方案的目的。
另外，在本发明各实施例中的各功能单元可以全部集成在一个处理单元中，也可以是各单元分别单独作为一个单元，也可以两个或两个以上单元集成在一个单元中；上述集成的单元既可以采用硬件的形式实现，也可以采用硬件加软件功能单元的形式实现。
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步骤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步骤；而前述的存储介质包括：移动存储设备、只读存储器(ROM，Read-Only Memory)、随机存取存储器(RAM，Random Access Memory)、磁碟或者光 盘等各种可以存储程序代码的介质。
或者，本发明上述集成的单元如果以软件功能模块的形式实现并作为独立的产品销售或使用时，也可以存储在一个计算机可读取存储介质中。基于这样的理解，本申请实施例的技术方案本质上或者说对现有技术做出贡献的部分可以以软件产品的形式体现出来，该计算机软件产品存储在一个存储介质中，包括若干指令用以使得一台计算机设备(可以是个人计算机、服务器、或者网络设备等)执行本发明各个实施例所述方法的全部或部分。而前述的存储介质包括：移动存储设备、ROM、RAM、磁碟或者光盘等各种可以存储程序代码的介质。
以上所述，仅为本发明的具体实施方式，但本发明的保护范围并不局限于此，任何熟悉本技术领域的技术人员在本发明揭露的技术范围内，可轻易想到变化或替换，都应涵盖在本发明的保护范围之内。因此，本发明的保护范围应以所述权利要求的保护范围为准。

Claims (12)

  1. 一种图像识别方法，其特征在于，所述方法包括：
    获取采集图像；
    获取采集图像的特征图像；
    基于采集图像的特征图像，对采集图像中的多个候选区域进行初步筛选，得到第一筛选结果，所述筛选结果中至少包括所述采集图像上可能存在有检测目标的候选区域；
    获取至少一种检测目标的特征信息；
    基于至少一种检测目标的特征信息，对所述采集图像上可能存在有检测目标的候选区域进行处理，得到第二筛选结果，所述第二筛选结果至少包括在候选区域上存在的检测目标在采集图像所处的位置信息。
  2. 根据权利要求1所述的方法，其特征在于，所述第二筛选结果至少还包括在候选区域上存在的检测目标的种类。
  3. 根据权利要求1所述的方法，其特征在于，所述基于采集图像的特征图像，对采集图像中的多个候选区域进行初步筛选，得到第一筛选结果，包括：
    基于采集图像的特征图像，得到采集图像各个候选区域的锚点信息；
    基于各个候选区域的锚点信息，对采集图像的各个候选区域上的前景图像和背景图像进行分离；
    基于分离结果，得到采集图像上可能存在有检测目标的候选区域。
  4. 根据权利要求1至3任一项所述的方法，其特征在于，所述基于至少一种检测目标的特征信息，对所述采集图像上可能存在有检测目标的候选区域进行处理，得到第二筛选结果，包括：
    基于至少一种检测目标的特征信息，对所述采集图像上可能存在有检 测目标的候选区域中的各个候选区域进行检测目标的识别，确定存在有检测目标的候选区域；
    获取存在有检测目标的候选区域在采集图像上的位置信息；
    依据存在有检测目标的候选区域在采集图像上的位置信息，确定在存在有检测目标的候选区域中的检测目标在采集图像上所处的位置信息。
  5. 根据权利要求4所述的方法，其特征在于，所述方法包括：
    至少通过对所述采集图像的至少一次卷积处理而得到所述采集图像的特征图像。
  6. 一种图像识别设备，其特征在于，所述设备包括：
    第一获取单元，用于获取采集图像；
    第二获取单元，用于获取采集图像的特征图像；
    第一筛选单元，用于基于采集图像的特征图像，对采集图像中的多个候选区域进行初步筛选，得到第一筛选结果，所述筛选结果中至少包括所述采集图像上可能存在有检测目标的候选区域；
    第三获取单元，用于获取至少一种检测目标的特征信息；
    第二筛选单元，用于基于至少一种检测目标的特征信息，对所述采集图像上可能存在有检测目标的候选区域进行处理，得到第二筛选结果，所述第二筛选结果至少包括在候选区域上存在的检测目标在采集图像所处的位置信息。
  7. 根据权利要求6所述的设备，其特征在于，所述第二筛选单元，还用于：
    筛选出在候选区域上存在的检测目标的种类。
  8. 根据权利要求6所述的设备，其特征在于，所述第一筛选单元，还用于：
    基于采集图像的特征图像，得到采集图像各个候选区域的锚点信息；
    基于各个候选区域的锚点信息，对采集图像的各个候选区域上的前景图像和背景图像进行分离；
    基于分离结果，得到采集图像上可能存在有检测目标的候选区域。
  9. 根据权利要求6至8任一项所述的设备，其特征在于，所述第二筛选单元，还用于：
    基于至少一种检测目标的特征信息，对所述采集图像上可能存在有检测目标的候选区域中的各个候选区域进行检测目标的识别，确定存在有检测目标的候选区域；
    获取存在有检测目标的候选区域在采集图像上的位置信息；
    依据存在有检测目标的候选区域在采集图像上的位置信息，确定在存在有检测目标的候选区域中的检测目标在采集图像上所处的位置信息。
  10. 根据权利要求9所述的设备，其特征在于，所述第二获取单元，用于：
    至少通过对所述采集图像的至少一次卷积处理而得到所述采集图像的特征图像。
  11. 一种存储介质，该存储介质包括存储的程序，其中，上述程序运行时至少执行前述的权利要求1～5任一项所述的图像识别方法。
  12. 一种处理器，该处理器用于运行程序，其中，该程序被处理器运行时至少执行前述的权利要求1～5任一项所述的图像识别方法。
PCT/CN2018/108461 2018-09-28 2018-09-28 图像识别方法和设备、存储介质和处理器 WO2020062088A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/CN2018/108461 WO2020062088A1 (zh) 2018-09-28 2018-09-28 图像识别方法和设备、存储介质和处理器

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2018/108461 WO2020062088A1 (zh) 2018-09-28 2018-09-28 图像识别方法和设备、存储介质和处理器

Publications (1)

Publication Number Publication Date
WO2020062088A1 true WO2020062088A1 (zh) 2020-04-02

Family

ID=69953265

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/108461 WO2020062088A1 (zh) 2018-09-28 2018-09-28 图像识别方法和设备、存储介质和处理器

Country Status (1)

Country Link
WO (1) WO2020062088A1 (zh)

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111738185A (zh) * 2020-06-28 2020-10-02 杭州海康威视数字技术股份有限公司 一种目标识别方法、装置及设备
CN112560743A (zh) * 2020-12-23 2021-03-26 平安银行股份有限公司 公共区域异物检测方法、装置、设备及存储介质
CN112634216A (zh) * 2020-12-16 2021-04-09 西安理工大学 一种基于深度学习模型的绝缘子自爆检测方法
CN112749644A (zh) * 2020-12-30 2021-05-04 大连海事大学 一种基于改进可变形卷积的Faster RCNN火灾烟雾检测方法
CN113674315A (zh) * 2021-07-21 2021-11-19 浙江大华技术股份有限公司 一种物体检测方法、设备及计算机可读存储介质
CN114627388A (zh) * 2022-03-23 2022-06-14 南方电网数字电网研究院有限公司 输电线路异物检测设备及其异物检测方法
CN115965646A (zh) * 2023-03-16 2023-04-14 深圳思谋信息科技有限公司 区域划分方法、装置、计算机设备及计算机可读存储介质
CN117152421A (zh) * 2023-10-31 2023-12-01 南方电网数字电网研究院有限公司 输电线路异物检测方法、装置、计算机设备和存储介质
CN112560743B (zh) * 2020-12-23 2024-05-31 平安银行股份有限公司 公共区域异物检测方法、装置、设备及存储介质

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105067520A (zh) * 2015-07-28 2015-11-18 爱威科技股份有限公司 一种镜检识别方法及装置
CN107424150A (zh) * 2017-07-27 2017-12-01 济南浪潮高新科技投资发展有限公司 一种基于卷积神经网络的道路破损检测方法及装置
CN108109132A (zh) * 2016-11-25 2018-06-01 杭州海康威视数字技术股份有限公司 一种图像分析方法及装置
CN108537256A (zh) * 2018-03-26 2018-09-14 北京智芯原动科技有限公司 一种安全帽佩戴识别方法及装置

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105067520A (zh) * 2015-07-28 2015-11-18 爱威科技股份有限公司 一种镜检识别方法及装置
CN108109132A (zh) * 2016-11-25 2018-06-01 杭州海康威视数字技术股份有限公司 一种图像分析方法及装置
CN107424150A (zh) * 2017-07-27 2017-12-01 济南浪潮高新科技投资发展有限公司 一种基于卷积神经网络的道路破损检测方法及装置
CN108537256A (zh) * 2018-03-26 2018-09-14 北京智芯原动科技有限公司 一种安全帽佩戴识别方法及装置

Cited By (15)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111738185A (zh) * 2020-06-28 2020-10-02 杭州海康威视数字技术股份有限公司 一种目标识别方法、装置及设备
CN111738185B (zh) * 2020-06-28 2023-09-05 杭州海康威视数字技术股份有限公司 一种目标识别方法、装置及设备
CN112634216A (zh) * 2020-12-16 2021-04-09 西安理工大学 一种基于深度学习模型的绝缘子自爆检测方法
CN112634216B (zh) * 2020-12-16 2024-02-09 西安理工大学 一种基于深度学习模型的绝缘子自爆检测方法
CN112560743A (zh) * 2020-12-23 2021-03-26 平安银行股份有限公司 公共区域异物检测方法、装置、设备及存储介质
CN112560743B (zh) * 2020-12-23 2024-05-31 平安银行股份有限公司 公共区域异物检测方法、装置、设备及存储介质
CN112749644B (zh) * 2020-12-30 2024-02-27 大连海事大学 一种基于改进可变形卷积的Faster RCNN火灾烟雾检测方法
CN112749644A (zh) * 2020-12-30 2021-05-04 大连海事大学 一种基于改进可变形卷积的Faster RCNN火灾烟雾检测方法
CN113674315A (zh) * 2021-07-21 2021-11-19 浙江大华技术股份有限公司 一种物体检测方法、设备及计算机可读存储介质
CN114627388A (zh) * 2022-03-23 2022-06-14 南方电网数字电网研究院有限公司 输电线路异物检测设备及其异物检测方法
CN114627388B (zh) * 2022-03-23 2024-03-29 南方电网数字电网研究院有限公司 输电线路异物检测设备及其异物检测方法
CN115965646A (zh) * 2023-03-16 2023-04-14 深圳思谋信息科技有限公司 区域划分方法、装置、计算机设备及计算机可读存储介质
CN115965646B (zh) * 2023-03-16 2023-07-04 深圳思谋信息科技有限公司 区域划分方法、装置、计算机设备及计算机可读存储介质
CN117152421B (zh) * 2023-10-31 2024-03-22 南方电网数字电网研究院股份有限公司 输电线路异物检测方法、装置、计算机设备和存储介质
CN117152421A (zh) * 2023-10-31 2023-12-01 南方电网数字电网研究院有限公司 输电线路异物检测方法、装置、计算机设备和存储介质

Similar Documents

Publication Publication Date Title
WO2020062088A1 (zh) 图像识别方法和设备、存储介质和处理器
CN109447169B (zh) 图像处理方法及其模型的训练方法、装置和电子系统
WO2021000524A1 (zh) 孔位保护门检测方法、装置、计算机设备和存储介质
US10909682B2 (en) Method and device for detecting pulmonary nodule in computed tomography image, and computer-readable storage medium
US20200364842A1 (en) Surface defect identification method and apparatus
US10055843B2 (en) System and methods for automatic polyp detection using convulutional neural networks
CN110678901B (zh) 信息处理设备、信息处理方法和计算机可读存储介质
CN111680746B (zh) 车损检测模型训练、车损检测方法、装置、设备及介质
US11847775B2 (en) Automated machine vision-based defect detection
CN112001902A (zh) 缺陷检测方法和相关装置、设备、存储介质
CN111325739A (zh) 肺部病灶检测的方法及装置，和图像检测模型的训练方法
CN107679495B (zh) 一种输电线路周边活动工程车辆的检测方法
CN115294117B (zh) Led灯珠的缺陷检测方法及相关装置
CN109726678B (zh) 一种车牌识别的方法及相关装置
CN108734684B (zh) 动态照明场景的图像背景减除
JP4674179B2 (ja) 影認識方法及び影境界抽出方法
CN112396053A (zh) 一种基于级联神经网络的环视鱼眼图像目标检测方法
CN112464850A (zh) 图像处理方法、装置、计算机设备及介质
CN112906483A (zh) 一种目标重识别方法、装置及计算机可读存储介质
JP2022545962A (ja) 団霧認識方法及び装置、電子機器、記憶媒体並びにコンピュータプログラム製品
CN111814852A (zh) 图像检测方法、装置、电子设备和计算机可读存储介质
JP7040624B2 (ja) 画像データの特徴データを判別する方法、装置、システム及びプログラム、並びに記憶媒体
CN111127358A (zh) 图像处理方法、装置及存储介质
CN112819834B (zh) 基于人工智能的胃部病理图像的分类方法和装置
CN108985294B (zh) 一种轮胎模具图片的定位方法、装置、设备及存储介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 18935176

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 18935176

Country of ref document: EP

Kind code of ref document: A1