WO2018112783A1 - 图像识别方法及装置 - Google Patents

图像识别方法及装置 Download PDF

Info

Publication number
WO2018112783A1
WO2018112783A1 PCT/CN2016/111270 CN2016111270W WO2018112783A1 WO 2018112783 A1 WO2018112783 A1 WO 2018112783A1 CN 2016111270 W CN2016111270 W CN 2016111270W WO 2018112783 A1 WO2018112783 A1 WO 2018112783A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
category
prior probability
sample
current scene
Prior art date
Application number
PCT/CN2016/111270
Other languages
English (en)
French (fr)
Inventor
柴伦绍
Original Assignee
深圳前海达闼云端智能科技有限公司
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 深圳前海达闼云端智能科技有限公司 filed Critical 深圳前海达闼云端智能科技有限公司
Priority to PCT/CN2016/111270 priority Critical patent/WO2018112783A1/zh
Priority to CN201680020414.8A priority patent/CN107690660B/zh
Publication of WO2018112783A1 publication Critical patent/WO2018112783A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/24Classification techniques
    • G06F18/241Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2415Classification 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
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/045Combinations of networks

Definitions

  • the present application relates to the field of detection and identification technologies, and in particular, to an image recognition method and apparatus.
  • CNN Convolutional Neural Network
  • a confidence level is assigned to each category of the image classification model. That is to say, when the existing image classification model inputs the sample image, the softmax layer outputs the image category confidence, that is, the probability size for the sample image belonging to each category.
  • the sample image belongs to a set of images that are identified after training to obtain a deep neural network, that is, a test set (English: testing set).
  • the present application discloses an image recognition method and device, which mainly solves the problem in the prior art.
  • the present application provides an image recognition method, including: acquiring a sample image; determining a prior probability distribution of each category image appearing in a current scene; and performing the sample image according to the prior probability distribution Identifying, obtaining image category confidence corresponding to the sample image.
  • the present application provides an image recognition apparatus, the apparatus comprising: an acquisition unit, configured to acquire a sample image; and a determining unit, configured to determine a prior probability distribution of each category image appearing in a current scene; and a processing unit, And determining, by the prior probability distribution determined by the determining unit, the sample image acquired by the acquiring unit, to obtain an image category confidence corresponding to the sample image.
  • the present application provides a computer storage medium for storing computer software instructions comprising program code designed to perform the method of the first aspect.
  • the present application provides a computer program that can be directly loaded into an internal memory of a computer and contains software code that can be implemented by loading and executing the computer program via a computer.
  • the application provides a computer device including: a memory, a communication interface, and a processor.
  • the memory is for storing computer executable code
  • the processor is for executing computer executable code control to perform the method of the first aspect
  • the communication interface is for data transmission between the computer device and the external device.
  • the present application provides a robot comprising at least the computer device of the fifth aspect.
  • the image recognition method and apparatus provided by the present application have the problem of low accuracy of the image recognition process due to the difference between the training set of the image classification model and the image distribution of each category in the test set compared with the prior art.
  • the acquired sample image may be identified according to the prior probability distribution of each category image in the current scene, and the image category confidence corresponding to the sample image is obtained. It can be seen that during the test, although it is impossible to effectively reduce the distribution of each category in the training set and the test set itself, The difference, but by considering the current application scenario, the prior probability distribution of each category image in different application scenarios can be taken as another input parameter in the image recognition process. In this way, based on the current scene, combined with the probability of occurrence of each category of images in the current scene, the prior probabilities of each category in the image recognition process are respectively adjusted, and the accuracy of the image recognition process is improved.
  • FIG. 1 is a schematic diagram of an image classification model provided by the present application.
  • FIG. 3 and FIG. 4 are flowcharts of another image recognition method provided by the present application.
  • FIG. 5 is a schematic structural diagram of an image recognition apparatus provided by the present application.
  • FIG. 6 and FIG. 7 are schematic structural diagrams of another image recognition apparatus provided by the present application.
  • an image classification model for identifying each category image in the sample image and a confidence level corresponding to each category image.
  • an image classification model as shown in FIG. 1 is provided.
  • the image classification model 10 at least an input parameter 11, a depth volume integration class network 12, and an output parameter 13 are included.
  • the input parameter 11 includes at least a sample map
  • the output parameter 13 is the recognition result of the sample image 111, and may specifically include the image category confidence 131 corresponding to the sample image 111, for example, 111 and the prior probability 112 of each category image appearing in the current scene. It should be noted that, in consideration of different scenarios, it may be necessary to weigh the influence of the image feature and the prior probability on the image recognition process. Therefore, in the present application, the input parameter 11 may further include the influence specific gravity 113.
  • the softmax layer has a scene-aware function, which means that the softmax layer can take the application scene as an influencing factor of image recognition, and the conventional softmax layer described in the background art is simply based on The image features extracted from the sample image are subjected to an image recognition process. It can be seen that the softmax layer with scene perception function increases the prior parameter probability of each type of image, and can comprehensively image the image and the prior probability distribution of each category image for more accurate image recognition.
  • the present application provides an image recognition method. As shown in FIG. 2, the method may be performed by the image classification model 10 described above, and the method includes:
  • Step 101 Acquire a sample image.
  • the acquired sample image may be part or all of the image in the test set, or may be an image captured by a current user through a device having an image capturing function such as a camera or a mobile phone, or may be stored locally in a database or at a terminal. image.
  • the source of the sample image is not limited, and it may be determined by the user which images are acquired as sample images.
  • Step 102 Determine a prior probability distribution of each category image appearing in the current scene.
  • the prior probabilities of images appearing in each category in different scenarios may not be identical or completely different.
  • the prior probability of each category image refers to the probability that the image portion belonging to each category appears in the currently input sample image.
  • the category image may be an image portion of the category in which the image body includes a plant, an image portion of the image body including an animal, and the like.
  • the category indicated by the category image may be preset by the user or the maintenance personnel. The manner of dividing the category is not limited in the present application. In this application, users or maintenance personnel can pre- First, according to the empirical value, and/or historical record, and/or the geographic location of the collected sample image, the prior probability of each type of image appearing in different scenes is set.
  • Step 103 Identify the sample image according to the prior probability distribution, and obtain the recognition result.
  • the recognition result includes at least an image category confidence corresponding to the sample image.
  • the image recognition process not only the prior probability but also the image features need to be considered, so that the more accurate image category confidence corresponding to the sample image can be obtained after the identification, that is, the image of each category corresponding to the sample image and each category The confidence of the image. It should be noted that the sum of the confidences of all the classified images outputted by the same sample image after the image recognition is completed is 1.
  • the present application can identify the acquired sample image according to the prior probability distribution of each category image in the current scene, and obtain the image category confidence corresponding to the sample image. It can be seen that in the test process, although the difference between the distribution of the various types of images in the training set and the test set cannot be effectively reduced, by considering the current application scenario, the prior probability distribution of each type of image in different application scenarios can be used as the prior probability distribution. Another input parameter in the image recognition process. In this way, based on the current scene, combined with the probability of occurrence of each category of images in the current scene, the prior probabilities of each category in the image recognition process are respectively adjusted, and the accuracy of the image recognition process is improved.
  • the sample image may be identified in conjunction with a formula.
  • the prior probability of each category image that changes with the change of the application scene is set in the formula. Therefore, on the basis of the implementation shown in FIG. 2, an implementation as shown in FIG. 3 can also be implemented.
  • the step 103 is to identify the sample image according to the prior probability distribution, and obtain the recognition result, which may be specifically implemented as step 201:
  • Step 201 according to the formula
  • the recognition result of the output of the softmax layer is calculated, and the confidence of each category image corresponding to the sample image is obtained.
  • y (i) is used to indicate the type of the category image
  • x (i) is used to represent the ith sample image
  • p(y (i) n
  • x (i) ; ⁇ ) is used to represent the sample image x ( i) the degree of confidence belonging to the nth category
  • p 0 (n) is used to indicate the prior probability of the nth category image appearing in the current scene
  • p 0 (j) is used to indicate that the jth category image appears in the current scene
  • the prior probability, n, j and i are positive integers greater than or equal to 1 and less than or equal to k
  • k is a positive integer greater than or equal to 1
  • is used to represent the basic parameters of the image classification model
  • is the impact weight .
  • the value of the influence specific gravity is positively correlated with the magnitude of the influence of the image feature on the recognition sample image, and the value of the influence specific gravity has a negative correlation with the magnitude of the influence of the prior probability on the recognition sample image.
  • the image category confidence can be obtained according to the following formula:
  • the formula referred to in the present application has the same part as the above formula, and the main difference between the two formulas is that p 0 (n) and ⁇ are added to the input parameters of the present application.
  • the influence ratio ⁇ can be preset by the user or the maintenance personnel.
  • can be any parameter of [0, + ⁇ ), which is mainly used to control the image extracted by the deep convolution optic network model.
  • the proportion of the two influencing factors of the feature and the prior probability of the object in the application scene, that is, the image feature in the image classification model and the prior probability of each classification image affect the size of the sample image recognition process.
  • the smaller the influence of image features on the image recognition result in the image recognition process, and the greater the influence of the prior probability of each category image on the image recognition result in the current scene.
  • the larger the value of ⁇ the greater the influence of image features on the image recognition result in the image recognition process, and the influence of the prior probability of each category image on the image recognition result in the current scene. The smaller.
  • This application can improve the existing softmax layer to make the image recognition process
  • the impact of environmental factors can be fully considered, that is, the prior probability of each category image determined for the current scene is introduced in the formula.
  • the image recognition process is closer to the current scene and meets the user's needs. In this way, a more accurate image recognition result can be obtained.
  • step 102 determines that the prior probability distribution of each category image appears in the current scenario, which may be specifically implemented as step 301 and step 302, or may be specifically implemented as step 401 and step 402:
  • Step 301 Determine, in the specified set, the proportion of the number of occurrences of each category image in the current scene to the total number of sample images belonging to the current scene.
  • the specified set includes at least part of the image or all of the images in the training set.
  • Step 302 Determine the ratio as a prior probability distribution.
  • the above implementation is mainly based on empirical values and/or historical records to determine the prior probability. It should be noted that the history record may be data generated by the same user or different users in performing image recognition during the historical time period.
  • Step 401 Determine a scene category corresponding to a geographical location where the current scene is located.
  • Step 402 Determine a prior probability distribution corresponding to the scene category.
  • the above implementation method mainly determines the prior probability according to the geographical location of the sample image.
  • the coin recognition application takes the coin recognition application as an example.
  • the user can use the image classification model shown in Figure 1 to identify commonly used coins.
  • common coins can include dollars, renminbi, euros, yen and other currencies.
  • the sample images corresponding to the above various currencies that is, the training samples, are necessarily included in the training set.
  • the practical application of the image classification model is mainly in China. Therefore, according to the domestic use requirements, it can be understood that the probability of each currency of the RMB appearing in the country is higher, while the US dollar and the euro are higher.
  • each currency of a foreign currency such as the Japanese yen and the Japanese currency will appear in the country will Right lower.
  • the probability of occurrence of each currency value of each currency can be estimated, and the above estimated probability is taken as an input parameter of the image classification model, and the value of the influence factor of the prior probability is set at the same time.
  • the final result is a foreign currency; if the probability that the currency is recognized as a foreign currency is not significantly greater than the probability of being recognized as RMB according to the image feature, the recognition result Tend to the yuan.
  • the deep neural network algorithm needs to train a new image classification model long enough, and in the actual training process, if the task complexity is high, it may take several days, weeks or even longer to complete the training.
  • the specific suppression scheme may be: assigning a value of 0 to the prior probability corresponding to the category image that does not appear in the current scene, and/or adding one to the prior probability corresponding to the category image that may appear in the current scene.
  • the existing image classification model reserve is fully utilized, and a large amount of training time is saved. Moreover, the input of various resources such as calculation, data, and manpower is reduced, that is, the limitation of time and resources in the use of the existing image classification model is avoided as much as possible, and the application range of the deep classification model such as the image classification model is effectively expanded. .
  • the application may divide the function module into the image recognition device according to the above method example.
  • each function module may be divided according to each function, or two or more functions may be integrated into one processing module.
  • the above integrated modules can be implemented in the form of hardware or in the form of software functional modules. It should be noted that the division of modules in the present application is schematic, and is only a logical function division, and may be further divided in actual implementation.
  • FIG. 5 is a schematic diagram showing a possible structure of the image recognition apparatus involved in the above embodiment, and the image recognition apparatus 20 includes an acquisition unit 21, a determination unit 22, and Processing unit 23.
  • the obtaining unit 21 is configured to support the image recognition device 20 to perform step 101 in FIGS. 2 to 4
  • the determining unit 22 is configured to support the image recognition device 20 to perform step 102 in FIG. 2 and FIG. 3, step 301 and steps in FIG. 302, and step 401 and step 402 in FIG. 4
  • the processing unit 23 is configured to support the image recognition device 20 to perform step 103 in FIGS. 2 and 4, and step 201 in FIG. All the related content of the steps involved in the foregoing method embodiments may be referred to the functional descriptions of the corresponding functional modules, and details are not described herein again.
  • FIG. 6 shows the involved in the above embodiment.
  • the image processing device 30 includes a processing module 31 and a communication module 32.
  • the processing module 31 is configured to control and control the action of the image recognition device 30.
  • the processing module 31 is configured to support the image recognition device 30 to perform step 102 in FIG. 2 and FIG. 3, step 301 and step 302 in FIG. Steps 401 and 402 in FIG. 4, step 103 in FIGS. 2 and 4, and step 201 in FIG.
  • communication module 32 is used to support
  • the image recognition device 30 can communicate with other network entities, for example, with other functional modules or network entities, and can also be used to support the image recognition device 30 to perform step 101 of FIGS. 2 through 4.
  • the image recognition device 30 may further include a storage module 33 for storing program codes and data of the image recognition device 30.
  • the processing module 31 can be a processor or a controller, for example, a central processing unit (CPU), a general-purpose processor, a digital signal processor (DSP), and an application specific integrated circuit. (English: application-specific integrated circuit, ASIC), field programmable gate array (FPGA) or other programmable logic device, transistor logic device, hardware component or any combination thereof. It is possible to implement or carry out the various illustrative logical blocks, modules and circuits described in connection with the present disclosure.
  • the processor may also be a combination of computing functions, for example, including one or more microprocessor combinations, a combination of a DSP and a microprocessor, and the like.
  • the communication module 32 can be a transceiver, a transceiver circuit, a communication interface, or the like.
  • the storage module 33 can be a memory.
  • the image recognition device 30 may be the image recognition device 40 shown in FIG.
  • the image recognition device 40 includes a processor 41, a communication interface 42, a memory 43, and a bus 44.
  • the processor 41, the communication interface 42 and the memory 43 can be connected to each other through a bus 44, which can be a peripheral component interconnect (PCI) bus or an extended industry standard structure (English: extended industry) Standard Architecture, EISA) bus, etc.
  • the bus can be divided into an address bus, a data bus, a control bus, and the like. For ease of representation, only one thick line is shown in Figure 7, but it does not mean that there is only one bus or one type of bus.
  • the steps of a method or algorithm described in connection with the present disclosure may be implemented in a hardware or may be implemented by a processor executing software instructions.
  • the present application also provides a storage medium, which may include a memory 43 for storing computer software instructions for use in the image recognition device 40, including program code designed to execute a human-machine hybrid decision method.
  • the software instructions may be composed of corresponding software modules, and the software modules may be stored in random access memory (RAM), flash memory, read only memory (English: read only memory, ROM), erasable In addition to programmable read only memory (EPROM), electrically erasable programmable read only memory (EEPROM), or any other form of storage medium known in the art.
  • An exemplary storage medium is coupled to processor 41 to enable processor 41 to read information from, and to write information to, the storage medium.
  • the storage medium may also be an integral part of the processor 41.
  • the processor 41 and the storage medium may be located in an ASIC. Additionally, the ASIC can be located in a core network interface device.
  • the processor 41 and the storage medium may also exist as discrete components in the core network interface device.
  • the present application also provides a computer program product that can be directly loaded into the memory 43 and contains software code that can be implemented by the computer and loaded to execute the obstacle detection method described above.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • Molecular Biology (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Probability & Statistics with Applications (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Image Analysis (AREA)

Abstract

一种图像识别方法及装置,涉及检测、识别技术领域。为了解决现有技术中存在的因训练集合与测试集合类别分布差异较大时导致的图像识别准确率下降的问题而发明。该图像识别方法,包括:获取样本图像(101);确定在当前场景下出现每个类别图像的先验概率分布(102);根据先验概率分布对样本图像进行识别,得到识别结果(103)。该方法应用在图像识别的过程中。

Description

图像识别方法及装置 技术领域
本申请涉及检测、识别技术领域,尤其涉及一种图像识别方法及装置。
背景技术
随着图像识别技术的发展,深度卷积神经网络(英文:Convolutional Neural Network,CNN)算法逐步成为诸如图像分类、物体检测等图像识别研究及应用过程中的主流算法。
以图像分类为例,在进行测试过程之前,往往需要针对现有图像分类模型进行深度神经网络训练,在训练集合(英文:training set)上进行最优化求解即最小化全局损失(英文:loss)。之后在测试过程中,作为深度卷积神经网络输出层的软最大化层(英文:softmax layer),会为图像分类模型的每个类别赋予一个置信度。也就意味着,在现有图像分类模型输入样本图像时,softmax层输出的是图像类别置信度,即针对该样本图像属于各个类别的概率大小。其中,该样本图像属于训练得到深度神经网络后识别的图像集合,即测试集合(英文:testing set)。
在实际应用过程中,由于图像分类模型在训练过程中,采集实际数据会耗费大量人力、物力和时间,因此往往利用现有数据集进行模型训练。因此,训练集合与测试集合的类别分布通常会存在不一致的情况。然而,对于一个已完成训练的图像分类模型而言,在训练过程中训练集合中各类别图像在数量上的差异本身就会影响到测试过程中图像识别过程的准确率,再加上实际测试过程中无法保证测试集合中各类别的分布情况,因此,采用上述方式进行图像识别,往往会降低图像识别过程中的准确率。
发明内容
本申请公开一种图像识别方法及装置,主要解决现有技术中存 在的因训练集合与测试集合类别分布差别较大而导致的图像识别过程中准确率降低的问题。
为达到上述目的,本申请采用如下技术方案:
第一方面,本申请提供一种图像识别方法,该方法包括:获取样本图像;确定在当前场景下出现每个类别图像的先验概率分布;根据所述先验概率分布对所述样本图像进行识别,得到所述样本图像对应的图像类别置信度。
第二方面,本申请提供一种图像识别装置,该装置包括:获取单元,用于获取样本图像;确定单元,用于确定在当前场景下出现每个类别图像的先验概率分布;处理单元,用于根据确定单元确定的先验概率分布对获取单元获取的样本图像进行识别,得到样本图像对应的图像类别置信度。
第三方面,本申请提供一种计算机存储介质,用于储存计算机软件指令,其包含执行第一方面的方法所设计的程序代码。
第四方面,本申请提供一种计算机程序,可直接加载到计算机的内部存储器中,并含有软件代码,计算机程序经由计算机载入并执行后能够实现第一方面的方法。
第五方面,本申请提供一种计算机设备,该计算机设备包括:存储器、通信接口和处理器。其中,存储器用于存储计算机可执行代码,处理器用于执行计算机可执行代码控制执行第一方面的方法,通信接口用于计算机设备与外部设备的数据传输。
第六方面,本申请提供一种机器人,该机器人至少包括第五方面的计算机设备。
本申请提供的图像识别方法及装置,相比较于现有技术中,因图像分类模型的训练集合与测试集合中各类别图像分布的差异,而导致的图像识别过程准确率低的问题,本申请可以根据先当前场景下出现每个类别图像的先验概率分布,对所获取的样本图像进行识别,得到样本图像对应的图像类别置信度。由此可见,在测试过程中,虽然无法有效缩小训练集合和测试集合中各类别图像分布本身 的差异,但是通过考虑当前的应用场景,可以将不同应用场景下各个类别图像的先验概率分布作为图像识别过程中另一输入参数。这样就能以当前场景为依据,结合在当前场景下出现各个类别图像的概率来分别调整图像识别过程中每个类别的先验概率,也就提升了图像识别过程的准确率。
附图说明
图1为本申请提供的一种图像分类模型的示意图;
图2为本申请提供的一种图像识别方法流程图;
图3、图4为本申请提供的另一种图像识别方法流程图;
图5为本申请提供的一种图像识别装置的结构示意图;
图6、图7为本申请提供的另一种图像识别装置的结构示意图。
具体实施方式
本申请描述的系统架构以及业务场景是为了更加清楚的说明本申请的技术方案,并不构成对于本申请提供的技术方案的限定,本领域普通技术人员可知,随着系统架构的演变和新业务场景的出现,本申请提供的技术方案对于类似的技术问题,同样适用。
需要说明的是,本申请中,“示例性的”或者“例如”等词用于表示作例子、例证或说明。本申请中被描述为“示例性的”或者“例如”的任何实施例或设计方案不应被解释为比其它实施例或设计方案更优选或更具优势。确切而言,使用“示例性的”或者“例如”等词旨在以具体方式呈现相关概念。
需要说明的是,本申请中,“的(英文:of)”,“相应的(英文:corresponding,relevant)”和“对应的(英文:corresponding)”有时可以混用,应当指出的是,在不强调其区别时,其所要表达的含义是一致的。
图像分类模型,用于识别样本图像中每个类别图像以及每个类别图像对应的置信度。在本申请中,提供了一种如图1所示的图像分类模型,在图像分类模型10中,至少包括输入参数11、深度卷积分类网络12和输出参数13。其中,输入参数11至少包括样本图 像111和当前场景下出现每个类别图像的先验概率112,输出参数13为样本图像111的识别结果,具体可以包括样本图像111对应的图像类别置信度131。需要说明的是,考虑到不同场景下可能需要权衡图像特征和先验概率对图像识别过程所产生的影响,因此,在本申请中,输入参数11还可以包括影响比重113。
在本申请中,softmax层具有场景感知的功能,也就意味着,softmax层可以将应用场景作为图像识别的一个影响因素考虑在内,而在背景技术中所阐述的传统的softmax层单纯是根据从样本图像中提取的图像特征进行图像识别过程的。由此可见,具有场景感知功能的softmax层增加了各个类别图像的先验概率这一输入参数,能够综合图像特征和各个类别图像的先验概率分布情况,进行更加准确地图像识别。
本申请提供了一种图像识别方法,如图2所示,该方法可以由上述图像分类模型10执行,该方法包括:
步骤101、获取样本图像。
所获取的样本图像可以为测试集合中的部分或是全部图像,也可以为当前用户通过诸如照相机、手机等具有图像采集功能的设备所采集的图像,还可以为数据库中或是终端本地存储的图像。在本申请中,对于样本图像的来源不进行限定,可以由用户确定将哪些图像作为样本图像来获取。
步骤102、确定在当前场景下出现每个类别图像的先验概率分布。
不同场景下出现每个类别图像的先验概率可以不完全相同,或是完全不同。其中,在当前场景下,每个类别图像的先验概率指的是属于各个类别的图像部分在当前输入的样本图像中出现的概率。比如,类别图像可以为图像主体包括植物这一类别的图像部分、图像主体包括动物这一类别的图像部分等。需要说明的是,类别图像所指的类别可以由用户或是维护人员预先设置,至于类别的划分方式,在本申请中不做限定。在本申请中,用户或是维护人员可以预 先根据经验值、和/或历史记录、和/或采集样本图像的地理位置来完成不同场景下出现每个类别图像的先验概率的设置,具体确定方式会在后文提出,在此不进行赘述。需要说明的是,在图像分类模型完成训练后,可以确定各个类别图像。之后只需要针对已经确定的类别图像确定每个类别图像在不同场景下出现的先验概率分布即可。
步骤103、根据先验概率分布对样本图像进行识别,得到识别结果。
其中,识别结果至少包括样本图像对应的图像类别置信度。
在图像识别过程中,不仅需要考虑先验概率,还需要考虑图像特征,这样在识别后才能得到样本图像对应的更加准确的图像类别置信度,即该样本图像对应的各个类别图像以及每个类别图像的置信度。需要说明的是,同一样本图像在完成图像识别之后,所输出的所有类别图像的置信度的总和为1。
本申请可以根据先当前场景下出现每个类别图像的先验概率分布,对所获取的样本图像进行识别,得到样本图像对应的图像类别置信度。由此可见,在测试过程中,虽然无法有效缩小训练集合和测试集合中各类别图像分布本身的差异,但是通过考虑当前的应用场景,可以将不同应用场景下各个类别图像的先验概率分布作为图像识别过程中另一输入参数。这样就能以当前场景为依据,结合在当前场景下出现各个类别图像的概率来分别调整图像识别过程中每个类别的先验概率,也就提升了图像识别过程的准确率。
为了得到更加符合当前场景的图像类别置信度,在本申请的一个实现方式中,可以结合公式来对样本图像进行识别。其中,在公式中设置了会随着应用场景的改变而变化的各个类别图像的先验概率。因此,在如图2所示的实现方式的基础上,还可以实现为如图3所示的实现方式。其中,步骤103根据先验概率分布对样本图像进行识别,得到识别结果,可以具体实现为步骤201:
步骤201、根据公式
Figure PCTCN2016111270-appb-000001
计算softmax层输出的识别结果,得到样本图像对应的每个类别图像的置信度。
其中,y(i)用于表示类别图像的种类,x(i)用于表示第i张样本图像,p(y(i)=n|x(i);θ)用于表示样本图像x(i)隶属于第n类别的置信度,p0(n)用于表示第n类别图像在当前场景下出现的先验概率,p0(j)用于表示第j类别图像在当前场景下出现的先验概率,n、j和i均为大于或等于1且小于或等于k的正整数,k为大于或等于1的正整数,θ用于表示图像分类模型的基础参数,∈为影响比重。
需要说明的是,影响比重的取值与图像特征对识别样本图像带来影响的大小呈正相关,影响比重的取值与先验概率对识别样本图像带来影响的大小呈负相关。
目前,在深度卷积视神经网络中,采用图像分类模型对样本图像进行识别的过程中,可以依据如下公式得到图像类别置信度:
Figure PCTCN2016111270-appb-000002
由此可见,本申请所涉及的公式与上述公式存在相同的部分,而两个公式主要存在的差别为:本申请的输入参数中增加了p0(n)和∈。其中,影响比重∈可以由用户或是维护人员进行预先设置,在本申请中,∈可以为[0,+∞)内任意取值的参数,主要用于控制深度卷积视神经网络模型抽取的图像特征和应用场景内物体出现的先验概率两个影响因素的比重,即图像分类模型中图像特征和各个分类图像的先验概率为样本图像识别过程带来影响的大小。其中,∈的取值越小,则在图像识别过程中,图像特征给图像识别结果带来的影响越小,而当前场景下各个类别图像的先验概率给图像识别结果带来的影响越大;同理,∈的取值越大,则在图像识别过程中,图像特征给图像识别结果带来的影响越大,而当前场景下各个类别图像的先验概率给图像识别结果带来的影响越小。
本申请可以通过对已有softmax层进行改进,使图像识别过程 能够充分考虑到环境因素所带来的影响,即在公式中引入针对于当前场景所确定的各个类别图像的先验概率。同时,通过考虑图像特征与先验概率在图像识别过程中所占的比重,使图像识别过程更贴近当前场景,且符合用户需求。这样一来,就可以得到更加准确的图像识别结果。
在本申请的一个实现方式中,提供了两种可行的用于确定先验概率的方式。因此,在如图2或图3所示的实现方式的基础上,以图2为例,还可以实现为如图4所示的实现方式。其中,步骤102确定在当前场景下出现每个类别图像的先验概率分布,可以具体实现为步骤301和步骤302,或者可以具体实现为步骤401和步骤402:
步骤301、在指定集合中,确定当前场景下出现每个类别图像的次数占属于当前场景的样本图像总数量的占比。
其中,指定集合至少包括训练集合中的部分图像或是全部图像。
步骤302、将占比确定为先验概率分布。
上述实现方式主要是根据经验值和/或历史记录来确定先验概率。需要说明的是,历史记录可以为同一用户或是不同用户在历史时间段内进行图像识别所产生的数据。
步骤401、确定当前场景所处地理位置对应的场景类别。
步骤402、确定与场景类别对应的先验概率分布。
上述实现方式主要是根据采集样本图像的地理位置来确定先验概率。
例如,以钱币识别应用为例。用户可以使用如图1所示的图像分类模型识别常用的钱币。其中,常用的钱币具体可以包括美元、人民币、欧元、日元等币种。在训练过程中,为了使图像分类模型能够识别各种币种,则在训练集合中必然包括上述各种币种所对应的样本图像,即训练样本。然而,在完成图像分类模型的训练之后,考虑到该图像分类模型的实际应用主要在国内,因此,根据国内的使用需求可以了解到人民币的各个币值在国内出现的概率较高,而美元、欧元、日元等币种的外币的各个币值在国内出现的概率会相 对较低。据此,可以估计出一个各个币种的各个币值出现的概率,并将上述估计的概率作为图像分类模型的输入参数,同时设定先验概率的影响因子的取值。这样一来,如果根据图像特征,币种很确定地被识别为外币,即最终结果为外币;如果根据图像特征,币种被识别为外币的概率不显著大于识别为人民币的概率,则识别结果倾向于人民币。
由此可见,在图像识别的实际使用过程中,环境因素往往会给样本图像的处理结果带来很大的影响。
需要说明的是,在实际确定先验概率的过程中,不仅限于上述步骤301和步骤302,以及步骤401和步骤402所示的实现方式,还可以采用其他能够确定各个类别图像在不同应用场景中出现的概率来确定先验概率,在此不做限定。
目前,深度神经网络算法训练新的图像分类模型需要足够长的时间,并且在实际训练过程中,若任务复杂度较高,则可能需要数天、数周甚至更长的时间来完成训练。为了实现上述内容,需要储备大量的计算资源、采集海量数据,同时还需要配备能够掌握深度学习算法的人才。因此,考虑到图像分类模型在训练过程中需要耗费大量的人力、物力,因此,采用本申请进行图像识别,则可以使用已完成训练的图像分类模型进行图像识别。由于考虑到环境因素给识别过程带来的影响,因此,即便在不对图像分类模型进行新的训练过程的情况下,也可以有效提高图像识别的精度。
由于预先训练好的图像分类模型中所涉及的类别图像的种类往往很多,其中,很多类别图像在特定场景出现的概率极小,因此,需要抑制该类别图像的结果。具体抑制方案可以为:对于当前场景中基本不会出现的类别图像所对应的先验概率赋值为0,和/或对于当前场景中可能出现的类别图像所对应的先验概率之上再加一个较大的常数,即步骤201的公式中所涉及的影响比重,比如,100。这样就可以有效调整各个类别图像在当前场景中出现的概率,并且,将上述内容输入如图1所示的深度卷积分类网络12后,可以尽可能 地提升当前场景下图像识别过程的准确性。
也就意味着,在本申请中,在效果损失可以接受的范围内,在不重新对已有图像分类模型进行训练的基础上,充分利用已有的图像分类模型储备,节省大量的训练时间,且减少计算、数据、人力等各项资源的投入,也就是尽可能规避已有图像分类模型在使用过程中对于时间、资源的限制,有效扩大了诸如图像分类模型等的深度分类模型的应用范围。
本领域技术人员应该很容易意识到,结合本文中所公开的实施例描述的各示例的单元及算法步骤,本申请能够以硬件或硬件和计算机软件的结合形式来实现。某个功能究竟以硬件还是计算机软件驱动硬件的方式来执行,取决于技术方案的特定应用和设计约束条件。专业技术人员可以对每个特定的应用来使用不同方法来实现所描述的功能,但是这种实现不应认为超出本申请的范围。
本申请可以根据上述方法示例对图像识别装置进行功能模块的划分,例如,可以对应各个功能划分各个功能模块,也可以将两个或两个以上的功能集成在一个处理模块中。上述集成的模块既可以采用硬件的形式实现,也可以采用软件功能模块的形式实现。需要说明的是,本申请中对模块的划分是示意性的,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。
在采用对应各个功能划分各个功能模块的情况下,图5示出了上述实施例中所涉及的图像识别装置的一种可能的结构示意图,图像识别装置20包括:获取单元21、确定单元22和处理单元23。获取单元21用于支持图像识别装置20执行图2至图4中的步骤101,确定单元22用于支持图像识别装置20执行图2和图3中的步骤102,图4中的步骤301和步骤302,以及图4中的步骤401和步骤402;处理单元23用于支持图像识别装置20执行图2和图4中的步骤103,以及图3中的步骤201。其中,上述方法实施例涉及的各步骤的所有相关内容均可以援引到对应功能模块的功能描述,在此不再赘述。
在采用集成的单元的情况下,图6示出了上述实施例中所涉及 的图像识别装置的一种可能的结构示意图。图像处理装置30包括:处理模块31和通信模块32。其中,处理模块31用于对图像识别装置30的动作进行控制管理,例如,处理模块31用于支持图像识别装置30执行图2和图3中的步骤102,图4中的步骤301和步骤302,图4中的步骤401和步骤402,图2和图4中的步骤103,以及图3中的步骤201,和/或用于执行本文所描述的技术的其它步骤;通信模块32用于支持图像识别装置30与其他网络实体的通信,例如,可以与其他功能模块或网络实体之间进行通信,还可以用于支持图像识别装置30执行图2至图4中的步骤101。图像识别装置30还可以包括存储模块33,用于存储图像识别装置30的程序代码和数据。
其中,处理模块31可以是处理器或控制器,例如可以是中央处理器(英文:central processing unit,CPU),通用处理器,数字信号处理器(英文:digital signal processor,DSP),专用集成电路(英文:application-specific integrated circuit,ASIC),现场可编程门阵列(英文:field programmable gate array,FPGA)或者其他可编程逻辑器件、晶体管逻辑器件、硬件部件或者其任意组合。其可以实现或执行结合本申请公开内容所描述的各种示例性的逻辑方框,模块和电路。所述处理器也可以是实现计算功能的组合,例如包含一个或多个微处理器组合,DSP和微处理器的组合等等。通信模块32可以是收发器、收发电路或通信接口等。存储模块33可以是存储器。
当处理模块31为处理器,通信模块32为通信接口,存储模块33为存储器时,本申请所涉及的图像识别装置30可以为图7所示的图像识别装置40。该图像识别装置40包括:处理器41、通信接口42、存储器43以及总线44。其中,处理器41、通信接口42以及存储器43可以通过总线44相互连接,该总线44可以是外设部件互连标准(英文:peripheral component interconnect,PCI)总线或扩展工业标准结构(英文:extended industry standard  architecture,EISA)总线等。所述总线可以分为地址总线、数据总线、控制总线等。为便于表示,图7中仅用一条粗线表示,但并不表示仅有一根总线或一种类型的总线。
结合本申请公开内容所描述的方法或者算法的步骤可以硬件的方式来实现,也可以是由处理器执行软件指令的方式来实现。本申请还提供一种存储介质,该存储介质可以包括存储器43,用于储存为图像识别装置40所用的计算机软件指令,其包含执行人机混合决策方法所设计的程序代码。具体的,软件指令可以由相应的软件模块组成,软件模块可以被存放于随机存取存储器(英文:random access memory,RAM)、闪存、只读存储器(英文:read only memory,ROM)、可擦除可编程只读存储器(英文:erasable programmable ROM,EPROM)、电可擦可编程只读存储器(英文:electrically EPROM,EEPROM)或者本领域熟知的任何其它形式的存储介质中。一种示例性的存储介质耦合至处理器41,从而使处理器41能够从该存储介质读取信息,且可向该存储介质写入信息。当然,存储介质也可以是处理器41的组成部分。处理器41和存储介质可以位于ASIC中。另外,该ASIC可以位于核心网接口设备中。当然,处理器41和存储介质也可以作为分立组件存在于核心网接口设备中。
本申请还提供一种计算机程序产品,该计算机程序可直接加载到存储器43中,并含有软件代码,该计算机程序经由计算机载入并执行后能够实现上述的障碍物检测方法。
以上所述,仅为本申请的具体实施方式,但本申请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申请揭露的技术范围内,可轻易想到变化或替换,都应涵盖在本申请的保护范围之内。因此,本申请的保护范围应所述以权利要求的保护范围为准。

Claims (14)

  1. 一种图像识别方法,其特征在于,所述方法包括:
    获取样本图像;
    确定在当前场景下出现每个类别图像的先验概率分布;
    根据所述先验概率分布对所述样本图像进行识别,得到识别结果所述识别结果至少包括所述样本图像对应的图像类别置信度。
  2. 根据权利要求1所述的方法,其特征在于,所述根据所述先验概率分布对所述样本图像进行识别,得到识别结果,具体包括:
    根据公式
    Figure PCTCN2016111270-appb-100001
    计算软最大化层输出的识别结果,得到所述样本图像对应的每个类别图像的置信度,其中,y(i)用于表示类别图像的种类,x(i)用于表示第i张样本图像,p(y(i)=n|x(i);θ)用于表示样本图像x(i)第n类别的置信度,p0(n)用于表示第n类别图像在当前场景下出现的先验概率,p0(j)用于表示第j类别图像在当前场景下出现的先验概率,n、j和i均为大于或等于1且小于或等于k的正整数,k为大于或等于1的正整数,θ用于表示图像分类模型的基础参数,∈为影响比重。
  3. 根据权利要求2所述的方法,其特征在于,所述影响比重的取值与图像特征对识别所述样本图像带来影响的大小呈正相关,所述影响比重的取值与所述先验概率对识别所述样本图像带来影响的大小呈负相关。
  4. 根据权利要求1至3中任意一项所述的方法,其特征在于,所述确定在当前场景下出现每个类别图像的先验概率分布,具体包括:
    在指定集合中,确定当前场景下出现每个类别图像的次数占属于当前场景的样本图像总数量的占比,所述指定集合至少包括训练集合中的部分图像或是全部图像;
    将所述占比确定为所述先验概率分布。
  5. 根据权利要求1至3中任意一项所述的方法,其特征在于,所述确定在当前场景下出现每个类别图像的先验概率分布,具体包括:
    确定当前场景所处地理位置对应的场景类别;
    确定与所述场景类别对应的所述先验概率分布。
  6. 一种图像识别装置,其特征在于,所述装置包括:
    获取单元,用于获取样本图像;
    确定单元,用于确定在当前场景下出现每个类别图像的先验概率分布;
    处理单元,用于根据所述确定单元确定的所述先验概率对所述获取单元获取的所述样本图像进行识别,得到识别结果,所述识别结果至少包括所述样本图像对应的图像类别置信度。
  7. 根据权利要求6所述的装置,其特征在于,所述处理单元,具体用于:
    根据公式
    Figure PCTCN2016111270-appb-100002
    计算软最大化层输出的识别结果,得到所述样本图像对应的每个类别图像的置信度,其中,y(i)用于表示类别图像的种类,x(i)用于表示第i张样本图像,p(y(i)=n|x(i);θ)用于表示样本图像x(i)属于第n类别图像的置信度,p0(n)用于表示第n类别图像在当前场景下出现的先验概率,p0(j)用于表示第j类别图像在当前场景下出现的先验概率,n、j和i均为大于或等于1且小于或等于k的正整数,k为大于或等于1的正整数,θ用于表示图像分类模型的基础参数,∈为影响比重。
  8. 根据权利要求7所述的装置,其特征在于,所述影响比重的取值与图像特征对识别所述样本图像带来影响的大小呈正相关,所述影响比重的取值与所述先验概率对识别所述样本图像带来影响的大小呈负相关。
  9. 根据权利要求6至8中任意一项所述的装置,其特征在于,所述确定单元,具体用于:
    在指定集合中,确定当前场景下出现每个类别图像的次数占属于当前场景的样本图像总数量的占比,所述指定集合至少包括训练集合中的部分图像或是全部图像;
    将所述占比确定为所述先验概率分布。
  10. 根据权利要求6至8中任意一项所述的装置,其特征在于,所述确定单元,具体用于:
    确定当前场景所处地理位置对应的场景类别;
    确定与所述场景类别对应的所述先验概率分布。
  11. 一种计算机存储介质,其特征在于,用于储存计算机软件指令,其包含执行权利要求1至5中任意一项所述的方法所设计的程序代码。
  12. 一种计算机程序产品,其特征在于,可直接加载到计算机的内部存储器中,并含有软件代码,计算机程序经由计算机载入并执行后能够实现权利要求1至5中任意一项所述的方法。
  13. 一种计算机设备,其特征在于,所述计算机设备包括:存储器、通信接口和处理器,所述存储器用于存储计算机可执行代码,所述处理器用于执行所述计算机可执行代码控制执行权利要求1至5中任意一项所述的方法,所述通信接口用于所述计算机设备与外部设备的数据传输。
  14. 一种机器人,其特征在于,所述机器人至少包括权利要求13所述的计算机设备。
PCT/CN2016/111270 2016-12-21 2016-12-21 图像识别方法及装置 WO2018112783A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
PCT/CN2016/111270 WO2018112783A1 (zh) 2016-12-21 2016-12-21 图像识别方法及装置
CN201680020414.8A CN107690660B (zh) 2016-12-21 2016-12-21 图像识别方法及装置

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2016/111270 WO2018112783A1 (zh) 2016-12-21 2016-12-21 图像识别方法及装置

Publications (1)

Publication Number Publication Date
WO2018112783A1 true WO2018112783A1 (zh) 2018-06-28

Family

ID=61152324

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2016/111270 WO2018112783A1 (zh) 2016-12-21 2016-12-21 图像识别方法及装置

Country Status (2)

Country Link
CN (1) CN107690660B (zh)
WO (1) WO2018112783A1 (zh)

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110070101A (zh) * 2019-03-12 2019-07-30 平安科技(深圳)有限公司 植物种类的识别方法及装置、存储介质、计算机设备
CN110222718A (zh) * 2019-05-09 2019-09-10 华为技术有限公司 图像处理的方法及装置
CN110705651A (zh) * 2019-10-17 2020-01-17 上海眼控科技股份有限公司 用于对视频识别准确率测试的方法、装置、设备及介质
CN110991458A (zh) * 2019-11-25 2020-04-10 创新奇智(北京)科技有限公司 基于图像特征的人工智能识别结果抽样系统及抽样方法
CN111325709A (zh) * 2019-12-26 2020-06-23 联博智能科技有限公司 无线胶囊内窥镜图像检测系统及检测方法
CN111783812A (zh) * 2019-11-18 2020-10-16 北京沃东天骏信息技术有限公司 违禁图像识别方法、装置和计算机可读存储介质
CN111814846A (zh) * 2020-06-19 2020-10-23 浙江大华技术股份有限公司 属性识别模型的训练方法、识别方法及相关设备
CN112749702A (zh) * 2019-10-30 2021-05-04 中移(苏州)软件技术有限公司 一种图像识别方法、装置、终端及存储介质
CN112800403A (zh) * 2021-01-05 2021-05-14 北京小米松果电子有限公司 生成预测模型以及预测指纹识别异常的方法、装置及介质

Families Citing this family (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108764321B (zh) * 2018-05-21 2019-08-30 Oppo广东移动通信有限公司 图像识别方法和装置、电子设备、存储介质
CN108805198B (zh) * 2018-06-08 2021-08-31 Oppo广东移动通信有限公司 图像处理方法、装置、计算机可读存储介质和电子设备
CN110414572B (zh) * 2019-07-08 2021-09-07 北京三快在线科技有限公司 一种图像识别的方法及装置
CN110334228A (zh) * 2019-07-09 2019-10-15 广西壮族自治区基础地理信息中心 一种基于深度学习的互联网问题地图筛查方法
CN110490225B (zh) * 2019-07-09 2022-06-28 北京迈格威科技有限公司 基于场景的图像分类方法、装置、系统和存储介质
CN110610575B (zh) * 2019-09-20 2021-09-07 北京百度网讯科技有限公司 硬币识别方法及装置、收银机
CN114418036B (zh) * 2022-03-28 2022-06-21 浙江所托瑞安科技集团有限公司 神经网络的性能测试和训练方法、设备和存储介质
CN115171048B (zh) * 2022-07-21 2023-03-17 北京天防安全科技有限公司 基于图像识别的资产分类方法、系统、终端及存储介质

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101178721A (zh) * 2007-10-12 2008-05-14 北京拓尔思信息技术有限公司 一种对论坛中有用帖子信息进行分类并整理的方法
CN101329731A (zh) * 2008-06-06 2008-12-24 南开大学 图像中数学公式的自动识别方法
CN103903008A (zh) * 2014-03-26 2014-07-02 国家电网公司 一种基于图像识别输电线路的雾等级的方法以及系统
CN203825644U (zh) * 2014-03-26 2014-09-10 国家电网公司 一种基于图像识别输电线路的雾等级的系统
CN104881681A (zh) * 2015-05-22 2015-09-02 浙江大学 基于混合图模型的图像序列类别标注方法

Family Cites Families (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7274832B2 (en) * 2003-11-13 2007-09-25 Eastman Kodak Company In-plane rotation invariant object detection in digitized images
US8775341B1 (en) * 2010-10-26 2014-07-08 Michael Lamport Commons Intelligent control with hierarchical stacked neural networks
CN102172849A (zh) * 2010-12-17 2011-09-07 西安交通大学 基于小波包和概率神经网络的刀具破损自适应报警方法
US20160217378A1 (en) * 2013-08-30 2016-07-28 Hewlett Packard Enterprise Development Lp Identifying anomalous behavior of a monitored entity
CN104408435A (zh) * 2014-12-05 2015-03-11 浙江大学 一种基于随机池化卷积神经网络的人脸识别方法
CN104537647B (zh) * 2014-12-12 2017-10-20 中安消技术有限公司 一种目标检测方法及装置
CN105894025A (zh) * 2016-03-30 2016-08-24 中国科学院自动化研究所 基于多任务深度学习的自然图像美感质量评估方法
CN106204587B (zh) * 2016-05-27 2019-01-08 浙江德尚韵兴图像科技有限公司 基于深度卷积神经网络和区域竞争模型的多器官分割方法
CN105930901B (zh) * 2016-07-18 2018-05-25 河海大学 一种基于rbpnn的变压器故障诊断方法

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101178721A (zh) * 2007-10-12 2008-05-14 北京拓尔思信息技术有限公司 一种对论坛中有用帖子信息进行分类并整理的方法
CN101329731A (zh) * 2008-06-06 2008-12-24 南开大学 图像中数学公式的自动识别方法
CN103903008A (zh) * 2014-03-26 2014-07-02 国家电网公司 一种基于图像识别输电线路的雾等级的方法以及系统
CN203825644U (zh) * 2014-03-26 2014-09-10 国家电网公司 一种基于图像识别输电线路的雾等级的系统
CN104881681A (zh) * 2015-05-22 2015-09-02 浙江大学 基于混合图模型的图像序列类别标注方法

Cited By (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110070101A (zh) * 2019-03-12 2019-07-30 平安科技(深圳)有限公司 植物种类的识别方法及装置、存储介质、计算机设备
CN110070101B (zh) * 2019-03-12 2024-05-14 平安科技(深圳)有限公司 植物种类的识别方法及装置、存储介质、计算机设备
CN110222718A (zh) * 2019-05-09 2019-09-10 华为技术有限公司 图像处理的方法及装置
CN110222718B (zh) * 2019-05-09 2023-11-03 华为技术有限公司 图像处理的方法及装置
CN110705651A (zh) * 2019-10-17 2020-01-17 上海眼控科技股份有限公司 用于对视频识别准确率测试的方法、装置、设备及介质
CN112749702B (zh) * 2019-10-30 2023-04-28 中移(苏州)软件技术有限公司 一种图像识别方法、装置、终端及存储介质
CN112749702A (zh) * 2019-10-30 2021-05-04 中移(苏州)软件技术有限公司 一种图像识别方法、装置、终端及存储介质
CN111783812A (zh) * 2019-11-18 2020-10-16 北京沃东天骏信息技术有限公司 违禁图像识别方法、装置和计算机可读存储介质
CN110991458A (zh) * 2019-11-25 2020-04-10 创新奇智(北京)科技有限公司 基于图像特征的人工智能识别结果抽样系统及抽样方法
CN111325709A (zh) * 2019-12-26 2020-06-23 联博智能科技有限公司 无线胶囊内窥镜图像检测系统及检测方法
CN111814846B (zh) * 2020-06-19 2023-08-01 浙江大华技术股份有限公司 属性识别模型的训练方法、识别方法及相关设备
CN111814846A (zh) * 2020-06-19 2020-10-23 浙江大华技术股份有限公司 属性识别模型的训练方法、识别方法及相关设备
CN112800403A (zh) * 2021-01-05 2021-05-14 北京小米松果电子有限公司 生成预测模型以及预测指纹识别异常的方法、装置及介质
CN112800403B (zh) * 2021-01-05 2024-05-03 北京小米松果电子有限公司 生成预测模型以及预测指纹识别异常的方法、装置及介质

Also Published As

Publication number Publication date
CN107690660B (zh) 2021-03-23
CN107690660A (zh) 2018-02-13

Similar Documents

Publication Publication Date Title
WO2018112783A1 (zh) 图像识别方法及装置
CN107633265B (zh) 用于优化信用评估模型的数据处理方法及装置
TWI784941B (zh) 一種多重抽樣模型訓練方法及裝置
WO2021238262A1 (zh) 一种车辆识别方法、装置、设备及存储介质
WO2019214309A1 (zh) 模型测试的方法及装置
WO2019051941A1 (zh) 车型识别方法、装置、设备及计算机可读存储介质
CN110807491A (zh) 车牌图像清晰度模型训练方法、清晰度检测方法及装置
WO2021164232A1 (zh) 用户识别方法、装置、设备及存储介质
CN110909784B (zh) 一种图像识别模型的训练方法、装置及电子设备
CN113608916B (zh) 故障诊断的方法、装置、电子设备及存储介质
CN110135428B (zh) 图像分割处理方法和装置
CN111539456A (zh) 一种目标识别方法及设备
CN114387451A (zh) 异常图像检测模型的训练方法、装置及介质
WO2022121164A1 (zh) 封停敏感词预测方法、装置、计算机设备及存储介质
US11367311B2 (en) Face recognition method and apparatus, server, and storage medium
CN112434717B (zh) 一种模型训练方法及装置
CN116468702A (zh) 黄褐斑评估方法、装置、电子设备及计算机可读存储介质
CN112434547B (zh) 一种用户身份稽核方法和设备
CN114743048A (zh) 检测异常秸秆图片的方法和检测装置
CN112446428B (zh) 一种图像数据处理方法及装置
CN113139679A (zh) 基于神经网络的城市道路救援预警方法、系统及设备
CN114723939B (zh) 基于注意力机制的非极大值抑制方法、系统、设备和介质
CN111414992A (zh) 使用卷积神经网络对图像进行卷积计算的方法和装置
CN114581751B (zh) 图像识别模型的训练方法和图像识别方法、装置
CN116541713B (zh) 基于局部时频特征迁移学习的轴承故障诊断模型训练方法

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: 16924829

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS (EPO FORM 1205A DATED 18.10.2019)

122 Ep: pct application non-entry in european phase

Ref document number: 16924829

Country of ref document: EP

Kind code of ref document: A1