[go: up one dir, main page]

WO2018176195A1 - 一种室内场景的分类方法及装置 - Google Patents

一种室内场景的分类方法及装置 Download PDF

Info

Publication number
WO2018176195A1
WO2018176195A1 PCT/CN2017/078291 CN2017078291W WO2018176195A1 WO 2018176195 A1 WO2018176195 A1 WO 2018176195A1 CN 2017078291 W CN2017078291 W CN 2017078291W WO 2018176195 A1 WO2018176195 A1 WO 2018176195A1
Authority
WO
WIPO (PCT)
Prior art keywords
classification
observation area
scene
picture
training
Prior art date
Application number
PCT/CN2017/078291
Other languages
English (en)
French (fr)
Inventor
张俊宇
黄惠
Original Assignee
中国科学院深圳先进技术研究院
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 中国科学院深圳先进技术研究院 filed Critical 中国科学院深圳先进技术研究院
Priority to PCT/CN2017/078291 priority Critical patent/WO2018176195A1/zh
Priority to US16/495,401 priority patent/US11042777B2/en
Publication of WO2018176195A1 publication Critical patent/WO2018176195A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING OR CALCULATING; COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00Scenes; Scene-specific elements
    • G06V20/50Context or environment of the image
    • G06V20/52Surveillance or monitoring of activities, e.g. for recognising suspicious objects
    • GPHYSICS
    • G06COMPUTING OR CALCULATING; COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/21Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/211Selection of the most significant subset of features
    • G06F18/2115Selection of the most significant subset of features by evaluating different subsets according to an optimisation criterion, e.g. class separability, forward selection or backward elimination
    • GPHYSICS
    • G06COMPUTING OR CALCULATING; COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/24Classification techniques
    • GPHYSICS
    • G06COMPUTING OR CALCULATING; COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00Scenes; Scene-specific elements
    • G06V20/10Terrestrial scenes

Definitions

  • the invention belongs to the technical field of computers, and in particular relates to a method and a device for classifying indoor scenes.
  • Intelligent identification and classification is a key issue in computer vision.
  • hotspots focus on object recognition (a picture contains one or more objects) and face recognition (an image with a face).
  • indoor scene recognition is extremely challenging and is one of the most difficult classification tasks.
  • the difficulty lies in the fact that indoor scenes not only contain a large number of different objects, but also the placement of these objects in space is very different.
  • To accurately classify indoor scenes not only the information of objects in the scene but also the entire scene structure needs to be extracted. Characteristics.
  • the current scene recognition classification methods mainly include spatial pyramid method, high-level semantic information-based methods and convolutional neural network-based methods.
  • the feature representation of the spatial pyramid method relies only on low-level geometric information. Without the extraction of high-level semantic information, the ability to identify scenes is limited. Scene recognition methods based on high-level semantic information are limited. The range of selected objects greatly affects the ability of model classification.
  • the main disadvantage of the method based on convolutional neural network is that the training process needs to consume a lot of resources, and the effect is mainly on the detection and classification of objects.
  • the use is based on The convolutional neural network method can achieve 94% recognition rate when performing object recognition on the computer vision system identification (ImageNet) data set, and use the convolutional neural network based method to perform scenes on the public MIT-67 data set. When the classification is performed, only 69% of the recognition rate can be achieved. The reason is that the identification of the indoor scene does not only depend on the objects in the scene, but also the overall relationship between the connected objects, and the features directly extracted by the convolutional neural network method are not good. Grasp the integration of overall and local information.
  • An object of the present invention is to provide a method and a device for classifying indoor scenes, which aim to solve the problem that the existing scene recognition classification method is not accurate and the classification rate is not good.
  • the present invention provides a method for classifying an indoor scene, the method comprising the steps of:
  • the classification label of the to-be-classified scene picture is obtained according to the classification prediction result.
  • the present invention provides a classification device for an indoor scene, the device comprising:
  • a picture receiving unit configured to receive an input picture of the scene to be classified
  • An area obtaining unit configured to acquire a current local observation area from the picture to be classified according to a preset observation area positioning model
  • a vector acquiring unit configured to process image information of the current local observation area to obtain a feature vector of the picture to be classified
  • condition determining unit configured to acquire, according to the feature vector, a classification prediction result of the to-be-classified scene picture, and determine whether the classification prediction result satisfies a preset scene picture classification condition
  • a repeating execution unit configured to: when the classification prediction result does not satisfy the scene picture classification condition, acquire a next partial observation area from the to-be-classified scene image according to the observation area positioning model, and Setting a local observation area as the current local observation area, and triggering the vector acquisition unit to process image information of the current local observation area;
  • a scene classification unit configured to acquire, according to the classification prediction result, a classification label of the to-be-classified scene picture, when the classification prediction result satisfies the scene picture classification condition.
  • the present invention After receiving the input scene image to be classified, the present invention acquires the current local observation area from the image to be classified according to the preset observation area positioning model, and processes the image information of the current local observation area to obtain the picture of the scene to be classified.
  • the feature vector obtains the classification prediction result of the scene image to be classified according to the feature vector, and determines whether the classification prediction result satisfies the preset scene picture classification condition. When the classification prediction result does not satisfy the scene picture classification condition, the classification area model is to be classified according to the observation area.
  • Embodiment 1 is a flowchart of implementing a method for classifying an indoor scene according to Embodiment 1 of the present invention
  • FIG. 2 is a flowchart showing an implementation of establishing an observation area positioning model in a method for classifying an indoor scene according to Embodiment 2 of the present invention
  • FIG. 3 is a schematic structural diagram of a device for classifying an indoor scene according to Embodiment 3 of the present invention.
  • FIG. 4 is a schematic structural diagram of an apparatus for classifying an indoor scene according to Embodiment 4 of the present invention.
  • Embodiment 1 is a diagrammatic representation of Embodiment 1:
  • FIG. 1 is a flowchart showing an implementation process of a method for classifying an indoor scene according to Embodiment 1 of the present invention. For convenience of description, only parts related to the embodiment of the present invention are shown, which are described in detail as follows:
  • step S101 the input scene picture to be classified is received.
  • step S102 the current local observation area is obtained from the picture to be classified according to the preset observation area positioning model.
  • the picture to be classified is a picture corresponding to the indoor scene to be identified.
  • the observation area positioning model only one local observation area is selected from the scene picture at a time for identification and classification.
  • step S103 the image information of the current local observation area is processed to obtain a feature vector of the scene picture to be classified.
  • the image information of the current local observation area is acquired, after processing the image information of the current local observation area, the image information of the current local observation area is first encoded to obtain a local feature vector. Then, the obtained local feature vector and the previously obtained feature vector are subjected to a fusion operation to obtain a feature vector of the image information of the image to be classified, thereby improving the comprehensiveness of the feature vector and improving the accuracy of the classification of the scene image.
  • step S104 the classification prediction result of the picture to be classified is acquired according to the feature vector.
  • step S105 it is determined whether the classification prediction result satisfies a preset scene picture classification condition.
  • the classification prediction result includes a classification result and a corresponding prediction probability.
  • the plurality of classification results of the scene image and the corresponding prediction probability may be predicted according to the feature vector.
  • the sum of the predicted probabilities of the plurality of classification results is 100%, and it is judged whether there is a classification result corresponding to the preset prediction probability in the plurality of classification results, that is, whether the classification prediction result satisfies the preset classification scene image to be classified.
  • the preset threshold of the prediction probability may be set to 65%, and it is determined whether there is a classification result corresponding to the prediction probability greater than 65% among the plurality of classification results.
  • step S106 when the classification prediction result does not satisfy the scene picture classification condition, the next partial observation area is obtained from the picture to be classified according to the observation area positioning model, and the next partial observation area is set as the current local observation area, and the jump is performed. Go to the step of processing the image information of the current local observation area to obtain the feature vector of the scene picture to be classified.
  • the existing classification prediction result does not satisfy the preset condition for classifying the scene to be classified.
  • the next local observation area is acquired according to the observation area positioning model, and the next partial observation area is set as the current local observation area, and the repetition is repeated.
  • the image information processing is performed and the classification prediction result is obtained until the classification prediction result satisfies the scene picture classification condition.
  • step S107 when the classification prediction result satisfies the scene picture classification condition, the classification label of the scene picture to be classified is obtained according to the classification prediction result.
  • the condition that the classification prediction result has met the preset classification of the scene to be classified is determined, that is, The classification of the picture to be classified can be implemented. Therefore, the classification result of the corresponding prediction probability in the classification prediction result is obtained, and the classification result is set as the classification label of the picture to be classified, thereby improving the classification of the scene picture. accuracy.
  • the input scene image to be classified is received, and the current local observation area is obtained from the image to be classified according to the preset observation area positioning model, thereby reducing the complexity of the image recognition classification of the scene to be classified, and improving the complexity.
  • Embodiment 2 is a diagrammatic representation of Embodiment 1:
  • step S201 the input scene to be trained is received, and the current training local observation area is obtained from the to-be-trained scene picture according to a preset Gaussian distribution.
  • step S202 the classification operation is performed on the training scene picture according to the current training local observation area, and the reward value of the classification operation is calculated.
  • the training scene image is classified according to the feature vector, and the classification label of the scene image to be trained is obtained.
  • the image information of the local observation area is acquired, when the image information of the current training local observation area is processed, the image information of the current training local observation area is first encoded to obtain a local feature vector, and then The obtained local feature vector performs a fusion operation with the previously obtained feature vector to obtain a feature vector of the image information of the image to be classified, thereby improving the comprehensiveness of the feature vector and improving the accuracy of the classification of the scene image.
  • the dimensions of the feature vector can be adjusted during training to optimize the training results.
  • the standard classification label of the image to be trained is obtained, and the classification label of the image to be trained is determined to be correct, and the reward value of the classification is calculated.
  • the feedback value in the calculation formula of the reward value may be appropriately changed during the training process to optimize the speed of the model convergence, thereby optimizing the training model.
  • step S204 when the preset training end condition is not reached, the next training local observation area is obtained from the to-be-trained scene picture according to the Gaussian distribution, and the next training local observation area is set as the current training local observation area. And jumping to the step of classifying the training scene picture according to the current training local observation area and calculating the reward value of the classification operation.
  • the next training local observation area may be sampled from a Gaussian distribution of a given variance.
  • the next training local observation area obtained by sampling is repeatedly identified, and the training scene picture is classified according to the identified information to obtain a classification label, and each classification can be calculated to obtain a corresponding reward value.
  • step S205 when the preset training end condition is reached, the algebraic sum of the reward values of each picture to be trained in the picture to be trained is acquired, to obtain the total reward value of each picture to be trained, according to the total The reward value establishes an observation area location model that maximizes the total reward value.
  • the algebraic sum of the reward values of the scene to be trained is obtained, to obtain the total reward value of the image to be trained, and each of the images to be trained is
  • the picture to be trained has a corresponding total reward value, and an observation area positioning model that can maximize the total reward value is established according to the total reward value, for use in classifying the picture of the scene to be classified. Determine the optimal next local observation area to improve the classification rate and accuracy of the scene recognition classification.
  • Embodiment 3 is a diagrammatic representation of Embodiment 3
  • FIG. 3 is a diagram showing the structure of an indoor scene classification apparatus according to Embodiment 3 of the present invention. For the convenience of description, only parts related to the embodiment of the present invention are shown.
  • the classification device of the indoor scene includes a picture receiving unit 31, an area obtaining unit 32, a vector obtaining unit 33, a condition determining unit 34, and a scene classifying unit 35, wherein:
  • the picture receiving unit 31 is configured to receive the input picture of the scene to be classified.
  • the area obtaining unit 32 is configured to obtain a current local observation area from the picture to be classified according to the preset observation area positioning model.
  • the image information of the local observation area is acquired after the image information of the local observation area is acquired, the image information of the local observation area is first encoded to obtain a local feature vector, and then The obtained local feature vector performs a fusion operation with the previously obtained feature vector to obtain a feature vector of the image information of the image to be classified, thereby improving the comprehensiveness of the feature vector and improving the accuracy of the classification of the scene image.
  • the vector obtaining unit 33 includes:
  • the encoding operation unit 331 is configured to encode image information of the current local observation area to obtain a local feature vector
  • the merging operation unit 332 is configured to perform a merging operation on the local feature vector and the pre-stored feature vector to obtain a feature vector of the scene picture.
  • the scene classification unit 36 is configured to obtain a classification label of the scene image to be classified according to the classification prediction result when the classification prediction result satisfies the scene picture classification condition.
  • the training area obtaining unit 401 is configured to receive the input scene image to be trained, and obtain the current training local observation area from the to-be-trained scene picture according to the preset Gaussian distribution.
  • the feedback value in the calculation formula of the reward value may be appropriately changed during the training process to optimize the speed of the model convergence, thereby optimizing the training model.
  • the regional training unit 402 includes:
  • the loop training unit 403 is configured to: when the preset training end condition is not reached, obtain the next training local observation area from the to-be-trained scene picture according to the Gaussian distribution, and set the next training local observation area as the current training local part.
  • the observation area, and the trigger area training unit 402 performs a classification operation on the training scene picture according to the current training local observation area and calculates a bonus value of the classification operation.
  • the area obtaining unit 406 is configured to obtain a current local observation area from the picture to be classified according to the preset observation area positioning model.
  • the vector obtaining unit 407 is configured to process image information of the current local observation area to obtain a feature vector of the picture to be classified.
  • the condition determining unit 408 is configured to obtain a classification prediction result of the picture to be classified according to the feature vector, and determine whether the classification prediction result satisfies a preset scene picture classification condition.
  • the plurality of classification results of the scene image and the corresponding prediction probability may be predicted according to the feature vector, and the total predicted probability of the multiple classification results is 100%.
  • the condition judging unit judges whether there is a classification result corresponding to the preset prediction probability in the plurality of classification results, that is, whether the classification prediction result satisfies a condition for classifying the preset scene image to be classified.
  • the repetition execution unit 409 is configured to: when the classification prediction result does not satisfy the scene picture classification condition, obtain the next partial observation area from the image to be classified according to the observation area positioning model, and set the next partial observation area as the current local observation area.
  • trigger vector acquisition unit 407 processes the image information of the current local observation area.
  • the scene classification unit 410 is configured to obtain, when the classification prediction result satisfies the scene picture classification condition, the classification label of the scene picture to be classified according to the classification prediction result.
  • each unit of the classification device of the indoor scene may be implemented by a corresponding hardware or software unit, and each unit may be an independent software and hardware unit, or may be integrated into one soft and hardware unit, and is not limited thereto. this invention.
  • each unit may be implemented by a corresponding hardware or software unit, and each unit may be an independent software and hardware unit, or may be integrated into one soft and hardware unit, and is not limited thereto.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Multimedia (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
  • Image Analysis (AREA)

Abstract

本发明适用计算机技术领域,提供了一种室内场景的分类方法及装置,该方法包括:接收输入的待分类场景图片,根据预设的观测区域定位模型从待分类场景图片中获取当前局部观测区域,对当前局部观测区域的图像信息进行处理,以得到待分类场景图片的特征向量,根据特征向量获取待分类场景图片的分类预测结果,判断分类预测结果是否满足预设的场景图片分类条件,当不满足时,根据观测区域定位模型从待分类场景图片中获取下一局部观测区域,并将下一局部观测区域设置为当前局部观测区域,跳转至对当前局部观测区域的图像信息进行处理的步骤,当满足条件时,根据分类预测结果获取待分类场景图片的分类标签,从而提高了场景识别分类的分类速率和准确性。

Description

一种室内场景的分类方法及装置 技术领域
本发明属于计算机技术领域,尤其涉及一种室内场景的分类方法及装置。
背景技术
智能识别和分类是计算机视觉中重点研究的问题。在众多的研究中,热点主要聚焦于物体识别(一张图片包含一个或多个物体)和人脸识别(一张带人脸的图像)。相比于这些研究,室内场景识别极具挑战,是最为困难的分类任务之一。其难点主要在于室内场景不仅包含了大量不同的物体,而且这些物体在空间中的摆放形式千差万别,要对室内场景进行准确地分类,不仅要分析场景中物体的信息,还需提取整个场景结构的特征。
为了提高识别场景的能力,不少学者对此进行了深入研究,提出了很多行之有效的方法。目前的场景识别分类方法主要包括空间金字塔法、基于高层次语义信息的方法和基于卷积神经网络的方法。
这些方法有着明显的缺陷,空间金字塔法的特征表示只依赖于低层次的几何信息,缺少对高层次语义信息的提取,识别场景的能力很受限制,基于高层次语义信息的场景识别方法受限于所选物体的范围,大大地影响了模型分类的能力,基于卷积神经网络的方法主要缺点在于训练过程需要消耗大量的资源,而且主要在物体的检测和分类上效果明显,例如,使用基于卷积神经网络的方法在计算机视觉系统识别(ImageNet)数据集上进行物体识别时,可以达到94%的识别率,而使用基于卷积神经网络的方法在公开的MIT-67数据集上进行场景的分类时,只能达到69%的识别率,原因是室内场景的识别不只依赖于场景中的物体,还需要连接物体之间的整体关系,而卷积神经网络方法直接提取的特征不能较好地把握整体和局部信息的融合。
发明内容
本发明的目的在于提供一种室内场景的分类方法及装置,旨在解决现有的场景识别分类方法的准确性不高、分类速率不佳的问题。
一方面,本发明提供了一种室内场景的分类方法,所述方法包括下述步骤:
接收输入的待分类场景图片;
根据预设的观测区域定位模型从所述待分类场景图片中获取当前局部观测区域;
对所述当前局部观测区域的图像信息进行处理,以得到所述待分类场景图片的特征向量;
根据所述特征向量获取所述待分类场景图片的分类预测结果,判断所述分类预测结果是否满足预设的场景图片分类条件;
当所述分类预测结果不满足所述场景图片分类条件时,根据所述观测区域定位模型从所述待分类场景图片中获取下一局部观测区域,并将所述下一局部观测区域设置为所述当前局部观测区域,跳转至所述对所述当前局部观测区域的图像信息进行处理,以得到所述待分类场景图片的特征向量的步骤;
当所述分类预测结果满足所述场景图片分类条件时,根据所述分类预测结果获取所述待分类场景图片的分类标签。
另一方面,本发明提供了一种室内场景的分类装置,所述装置包括:
图片接收单元,用于接收输入的待分类场景图片;
区域获取单元,用于根据预设的观测区域定位模型从所述待分类场景图片中获取当前局部观测区域;
向量获取单元,用于对所述当前局部观测区域的图像信息进行处理,以得到所述待分类场景图片的特征向量;
条件判断单元,用于根据所述特征向量获取所述待分类场景图片的分类预测结果,判断所述分类预测结果是否满足预设的场景图片分类条件;
重复执行单元,用于当所述分类预测结果不满足所述场景图片分类条件时,根据所述观测区域定位模型从所述待分类场景图片中获取下一局部观测区域,并将所述下一局部观测区域设置为所述当前局部观测区域,并触发所述向量获取单元对所述当前局部观测区域的图像信息进行处理;以及
场景分类单元,用于当所述分类预测结果满足所述场景图片分类条件时,根据所述分类预测结果获取所述待分类场景图片的分类标签。
本发明在接收输入的待分类场景图片后,根据预设的观测区域定位模型从待分类场景图片中获取当前局部观测区域,对当前局部观测区域的图像信息进行处理,以得到待分类场景图片的特征向量,根据特征向量获取待分类场景图片的分类预测结果,判断分类预测结果是否满足预设的场景图片分类条件,当分类预测结果不满足场景图片分类条件时,根据观测区域定位模型从待分类场景图片中获取下一局部观测区域,并将下一局部观测区域设置为当前局部观测区域,跳转至对当前局部观测区域的图像信息进行处理,以得到待分类场景图片的特征向量的步骤,直至分类预测结果满足场景图片分类条件,当分类预测结果满足场景图片分类条件时,根据分类预测结果获取待分类场景图片的分类标签,从而提高了场景识别分类的分类速率和准确性。
附图说明
图1是本发明实施例一提供的室内场景的分类方法的实现流程图;
图2是本发明实施例二提供的室内场景的分类方法中建立观测区域定位模型的实现流程图;
图3是本发明实施例三提供的室内场景的分类装置的结构示意图;以及
图4是本发明实施例四提供的室内场景的分类装置的结构示意图。
具体实施方式
为了使本发明的目的、技术方案及优点更加清楚明白,以下结合附图及实 施例,对本发明进行进一步详细说明。应当理解,此处所描述的具体实施例仅仅用以解释本发明,并不用于限定本发明。
以下结合具体实施例对本发明的具体实现进行详细描述:
实施例一:
图1示出了本发明实施例一提供的室内场景的分类方法的实现流程,为了便于说明,仅示出了与本发明实施例相关的部分,详述如下:
在步骤S101中,接收输入的待分类场景图片。
在步骤S102中,根据预设的观测区域定位模型从待分类场景图片中获取当前局部观测区域。
在本发明实施例中,待分类场景图片为待识别分类的室内场景对应的图片。为了降低识别分类过程中计算的复杂度,提高识别分类的可控性,根据观测区域定位模型每次只从场景图片中选择一个局部观测区域,以进行识别和分类。
在步骤S103中,对当前局部观测区域的图像信息进行处理,以得到待分类场景图片的特征向量。
在本发明实施例中,优选地,在获取到当前局部观测区域的图像信息之后,对当前局部观测区域的图像信息进行处理时,首先对当前局部观测区域的图像信息进行编码,得到局部特征向量,然后对得到的局部特征向量与预先得到的特征向量执行融合操作,得到待分类场景图片图像信息的特征向量,从而提高了特征向量的全面性,进而提高对场景图片分类的准确性。
在步骤S104中,根据特征向量获取待分类场景图片的分类预测结果。
在步骤S105中,判断分类预测结果是否满足预设的场景图片分类条件。
在本发明实施例中,分类预测结果包括分类结果和对应的预测概率,在得到待分类场景图片图像信息的特征向量之后,根据特征向量可以预测得到场景图片的多个分类结果及对应的预测概率,多个分类结果的预测概率总和为100%,判断这些多个分类结果中是否存在对应的预测概率大于预设阈值的分类结果,即判断分类预测结果是否满足预设的对待分类场景图片进行分类的条件。作为 示例地,预测概率的预设阈值可以设置为65%,判断这些多个分类结果中是否存在对应的预测概率大于65%的分类结果。
在步骤S106中,当分类预测结果不满足场景图片分类条件时,根据观测区域定位模型从待分类场景图片中获取下一局部观测区域,并将下一局部观测区域设置为当前局部观测区域,跳转至对当前局部观测区域的图像信息进行处理,以得到待分类场景图片的特征向量的步骤。
在本发明实施例中,当这些多个分类结果中不存在对应的预测概率大于预设阈值的分类结果时,说明现有的分类预测结果不满足预设的对待分类场景图片进行分类的条件,若要实现对待分类场景图片的分类,还需要获取场景图片更多的区域信息,因此,根据观测区域定位模型获取下一个局部观测区域,并将下一局部观测区域设置为当前局部观测区域,重复进行图像信息处理并获取分类预测结果,直至分类预测结果满足场景图片分类条件。
在步骤S107中,当分类预测结果满足场景图片分类条件时,根据分类预测结果获取待分类场景图片的分类标签。
在本发明实施例中,当这些预测得到的多个分类结果中存在对应的预测概率大于预设阈值的分类结果时,说明分类预测结果已经满足预设的对待分类场景图片进行分类的条件,即已经可以实现对待分类场景图片的分类,因此,获取分类预测结果中对应的预测概率大于预设阈值的分类结果,将该分类结果设置为待分类场景图片的分类标签,从而提高了场景图片分类的准确性。
在本发明实施例中,接收输入的待分类场景图片,根据预设的观测区域定位模型从待分类场景图片中获取当前局部观测区域,从而降低了待分类场景图片识别分类的复杂度,提高了识别分类的可控性,对当前局部观测区域的图像信息进行处理,以得到待分类场景图片的特征向量,从而提高场景图片分类的分类速率,根据特征向量获取待分类场景图片的分类预测结果,判断分类预测结果是否满足预设的场景图片分类条件,当分类预测结果不满足场景图片分类条件时,根据观测区域定位模型从待分类场景图片中获取下一局部观测区域, 并将下一局部观测区域设置为当前局部观测区域,重复进行图像信息处理并获取分类预测结果,直至分类预测结果满足场景图片分类条件,当分类预测结果满足场景图片分类条件时,根据分类预测结果获取待分类场景图片的分类标签,从而提高了场景图片分类的准确性。
实施例二:
图2示出了本发明实施例二提供的室内场景的分类方法中建立观测区域定位模型的实现流程,为了便于说明,仅示出了与本发明实施例相关的部分,详述如下:
在步骤S201中,接收输入的待训练场景图片,根据预设的高斯分布从待训练场景图片中获取当前训练用局部观测区域。
在本发明实施例中,待训练场景图片为室内场景的局部观测区域定位训练过程中输入的需要分类的场景图片。为了降低识别分类过程中计算的复杂度,提高识别分类的可控性,根据观测区域定位模型每次从场景图片中选择一个局部观测区域,以进行识别和分类。优选地,在训练过程中可以根据实际情况调整局部观测区域的大小,以优化训练结果。
在步骤S202中,根据当前训练用局部观测区域对待训练场景图片进行分类操作并计算分类操作的奖励值。
在本发明实施例中,通过对当前训练用局部观测区域的图像信息的处理,得到待训练场景图片的特征向量之后,根据特征向量对待训练场景图片进行分类,得到待训练场景图片的分类标签。优选地,在获取到局部观测区域的图像信息之后,在对当前训练用局部观测区域的图像信息进行处理时,首先对当前训练用局部观测区域的图像信息进行编码,得到局部特征向量,然后对得到的局部特征向量与预先得到的特征向量执行融合操作,得到待分类场景图片图像信息的特征向量,从而提高了特征向量的全面性,进而提高对场景图片分类的准确性。优选地,在训练过程中可以调节特征向量的维度,以优化训练结果。
在本发明实施例中,在每次得到待训练场景图片的分类标签之后,获取待 训练场景图片的标准分类标签,判断得到的待训练场景图片的分类标签是否正确,并计算分类的奖励值。优选地,在计算分类的奖励值时,根据分类奖励值的计算公式
Figure PCTCN2017078291-appb-000001
计算分类的奖励值,其中,rt为第t次分类的奖励值,t为分类次数,y为训练得到的分类标签,losst为第t次分类的分类误差,y=maxylogp(y)表示得到的分类标签是正确的,从而避免重复观察同样区域,避免观测噪声太多的区域。优选地,在训练过程中可适当改变奖励值的计算公式中的反馈值,以优化模型收敛的速度,从而优化训练模型。
在步骤S203中,判断是否达到预设的训练结束条件。
在步骤S204中,当未达到预设的训练结束条件时,根据高斯分布从待训练场景图片中获取下一训练用局部观测区域,将下一训练用局部观测区域设置为当前训练用局部观测区域,跳转至根据当前训练用局部观测区域对待训练场景图片进行分类操作并计算分类操作的奖励值的步骤。
在本发明实施例中,在训练的过程中,下一训练用局部观测区域可以从一个给定方差的高斯分布中采样得到。重复对采样得到的下一训练用局部观测区域进行识别,根据识别到的信息对待训练场景图片进行分类,得到分类标签,每一次分类都可以通过计算得到对应的奖励值。
在步骤S205中,当达到预设的训练结束条件时,获取所有待训练场景图片中每张待训练场景图片的奖励值的代数和,以得到每张待训练场景图片的总奖励值,根据总奖励值建立总奖励值最大化的观测区域定位模型。
在本发明实施例中,若达到预设的训练结束条件,则获取待训练场景图片的奖励值的代数和,以得到这张待训练场景图片的总奖励值,所有待训练场景图片中每张待训练场景图片都有对应的总奖励值,根据总奖励值建立可以使总奖励值最大化的观测区域定位模型,以用于在对待分类场景图片分类的过程中 确定最优的下一个局部观测区域,从而提高场景识别分类的分类速率和准确性。
本领域普通技术人员可以理解实现上述实施例方法中的全部或部分步骤是可以通过程序来指令相关的硬件来完成,所述的程序可以存储于一计算机可读取存储介质中,所述的存储介质,如ROM/RAM、磁盘、光盘等。
实施例三:
图3示出了本发明实施例三提供的室内场景的分类装置的结构,为了便于说明,仅示出了与本发明实施例相关的部分。
在本发明实施例中,室内场景的分类装置包括图片接收单元31、区域获取单元32、向量获取单元33、条件判断单元34以及场景分类单元35,其中:
图片接收单元31,用于接收输入的待分类场景图片。
区域获取单元32,用于根据预设的观测区域定位模型从待分类场景图片中获取当前局部观测区域。
向量获取单元33,用于对当前局部观测区域的图像信息进行处理,以得到待分类场景图片的特征向量。
在本发明实施例中,优选地,在获取到局部观测区域的图像信息之后,对局部观测区域的图像信息进行处理时,首先对局部观测区域的图像信息进行编码,得到局部特征向量,然后对得到的局部特征向量与预先得到的特征向量执行融合操作,得到待分类场景图片图像信息的特征向量,从而提高了特征向量的全面性,进而提高对场景图片分类的准确性。
因此,优选地,该向量获取单元33包括:
编码操作单元331,用于对当前局部观测区域的图像信息进行编码,得到局部特征向量;以及
融合操作单元332,用于对局部特征向量与预先存储的特征向量执行融合操作,得到场景图片的特征向量。
条件判断单元34,用于根据特征向量获取待分类场景图片的分类预测结果,判断分类预测结果是否满足预设的场景图片分类条件。
重复执行单元35,用于当分类预测结果不满足场景图片分类条件时,根据观测区域定位模型从待分类场景图片中获取下一局部观测区域,并将下一局部观测区域设置为当前局部观测区域,并触发向量获取单元33对当前局部观测区域的图像信息进行处理。
场景分类单元36,用于当分类预测结果满足场景图片分类条件时,根据分类预测结果获取待分类场景图片的分类标签。
在本发明实施例中,当这些预测得到的多个分类结果中存在对应的预测概率大于预设阈值的分类结果时,说明分类预测结果已经满足预设的对待分类场景图片进行分类的条件,即已经可以实现对待分类场景图片的分类,因此,场景分类单元获取分类预测结果中对应的预测概率大于预设阈值的分类结果,将该分类结果设置为待分类场景图片的分类标签,从而提高了场景图片分类的准确性。
在本发明实施例中,室内场景的分类装置的各单元可由相应的硬件或软件单元实现,各单元可以为独立的软、硬件单元,也可以集成为一个软、硬件单元,在此不用以限制本发明。各单元的具体实施方式可参考前述实施例一的描述,在此不再赘述。
实施例四:
图4示出了本发明实施例四提供的室内场景的分类装置的结构,为了便于说明,仅示出了与本发明实施例相关的部分。
在本发明实施例中,室内场景的分类装置包括训练区域获取单元401、区域训练单元402、循环训练单元403、定位模型建立单元404、图片接收单元405、区域获取单元406、向量获取单元407、条件判断单元408以及场景分类单元409,其中:
训练区域获取单元401,用于接收输入的待训练场景图片,根据预设的高斯分布从待训练场景图片中获取当前训练用局部观测区域。
区域训练单元402,用于根据当前训练用局部观测区域对待训练场景图片 进行分类操作并计算分类操作的奖励值。
在本发明实施例中,在每次得到待训练场景图片的分类标签之后,获取待训练场景图片的标准分类标签,判断得到的待训练场景图片的分类标签是否正确,奖励值计算单元根据判断结果计算分类的奖励值。优选地,在计算分类的奖励值时,根据分类奖励值的计算公式
Figure PCTCN2017078291-appb-000002
计算分类的奖励值,其中,rt为第t次分类的奖励值,t为分类次数,y为训练得到的分类标签,losst为第t次分类的分类误差,y=maxylogp(y)表示得到的分类标签是正确的,从而避免重复观察同样区域,避免观测噪声太多的区域。优选地,在训练过程中可适当改变奖励值的计算公式中的反馈值,以优化模型收敛的速度,从而优化训练模型。
因此,优选地,该区域训练单元402包括:
训练分类单元4021,用于对当前训练用局部观测区域的图像信息进行处理,得到待训练场景图片的当前特征向量,根据当前特征向量对待训练场景图片进行分类,得到待训练场景图片的分类标签;以及
奖励值计算单元4022,用于获取待训练场景图片的标准分类标签,将得到的分类标签与标准分类标签进行比较,判断得到的分类标签是否正确,根据判断结果计算分类的奖励值。
优选地,该奖励值计算单元4022包括:
计算子单元,用于计算分类误差,获取分类次数,通过分类奖励值的计算
Figure PCTCN2017078291-appb-000003
计算分类的奖励值,其中,rt为第t次分类的奖励值,t为分类次数,y为训练得到的分类标签,losst为第t次分类的分类误差,y=maxylogp(y)表示得到的分类 标签是正确的。
循环训练单元403,用于当未达到预设的训练结束条件时,根据高斯分布从待训练场景图片中获取下一训练用局部观测区域,将下一训练用局部观测区域设置为当前训练用局部观测区域,并触发区域训练单元402根据当前训练用局部观测区域对待训练场景图片进行分类操作并计算分类操作的奖励值。
定位模型建立单元404,用于当达到预设的训练结束条件时,获取所有待训练场景图片中每张待训练场景图片的奖励值的代数和,以得到每张待训练场景图片的总奖励值,根据总奖励值建立总奖励值最大化的观测区域定位模型。
图片接收单元405,用于接收输入的待分类场景图片。
区域获取单元406,用于根据预设的观测区域定位模型从待分类场景图片中获取当前局部观测区域。
向量获取单元407,用于对当前局部观测区域的图像信息进行处理,以得到待分类场景图片的特征向量。
条件判断单元408,用于根据特征向量获取待分类场景图片的分类预测结果,判断分类预测结果是否满足预设的场景图片分类条件。
在本发明实施例中,得到待分类场景图片图像信息的特征向量之后,根据特征向量可以预测得到场景图片的多个分类结果及对应的预测概率,多个分类结果的预测概率总和为100%,条件判断单元判断这些多个分类结果中是否存在对应的预测概率大于预设阈值的分类结果,即判断分类预测结果是否满足预设的对待分类场景图片进行分类的条件。
重复执行单元409,用于当分类预测结果不满足场景图片分类条件时,根据观测区域定位模型从待分类场景图片中获取下一局部观测区域,并将下一局部观测区域设置为当前局部观测区域,并触发向量获取单元407对当前局部观测区域的图像信息进行处理。
场景分类单元410,用于当分类预测结果满足场景图片分类条件时,根据分类预测结果获取待分类场景图片的分类标签。
在本发明实施例中,室内场景的分类装置的各单元可由相应的硬件或软件单元实现,各单元可以为独立的软、硬件单元,也可以集成为一个软、硬件单元,在此不用以限制本发明。各单元的具体实施方式可参考前述实施例的描述,在此不再赘述。
以上所述仅为本发明的较佳实施例而已,并不用以限制本发明,凡在本发明的精神和原则之内所作的任何修改、等同替换和改进等,均应包含在本发明的保护范围之内。

Claims (10)

  1. 一种室内场景的分类方法,其特征在于,所述方法包括下述步骤:
    接收输入的待分类场景图片;
    根据预设的观测区域定位模型从所述待分类场景图片中获取当前局部观测区域;
    对所述当前局部观测区域的图像信息进行处理,以得到所述待分类场景图片的特征向量;
    根据所述特征向量获取所述待分类场景图片的分类预测结果,判断所述分类预测结果是否满足预设的场景图片分类条件;
    当所述分类预测结果不满足所述场景图片分类条件时,根据所述观测区域定位模型从所述待分类场景图片中获取下一局部观测区域,并将所述下一局部观测区域设置为所述当前局部观测区域,跳转至所述对所述当前局部观测区域的图像信息进行处理,以得到所述待分类场景图片的特征向量的步骤;
    当所述分类预测结果满足所述场景图片分类条件时,根据所述分类预测结果获取所述待分类场景图片的分类标签。
  2. 如权利要求1所述的方法,其特征在于,接收输入的待分类场景图片的步骤之前,所述方法还包括:
    接收输入的待训练场景图片,根据预设的高斯分布从所述待训练场景图片中获取当前训练用局部观测区域;
    根据所述当前训练用局部观测区域对所述待训练场景图片进行分类操作并计算所述分类操作的奖励值;
    当未达到预设的训练结束条件时,根据所述高斯分布从所述待训练场景图片中获取下一训练用局部观测区域,将所述下一训练用局部观测区域设置为当前训练用局部观测区域,跳转至根据所述当前训练用局部观测区域对所述待训练场景图片进行分类操作并计算所述分类操作的奖励值的步骤;
    当达到预设的训练结束条件时,获取所有待训练场景图片中每张待训练场 景图片的所述奖励值的代数和,以得到每张待训练场景图片的总奖励值,根据所述总奖励值建立总奖励值最大化的观测区域定位模型。
  3. 如权利要求2所述的方法,其特征在于,根据所述当前训练用局部观测区域对所述待训练场景图片进行分类操作并计算所述分类操作的奖励值的步骤,包括:
    对所述当前训练用局部观测区域的图像信息进行处理,得到所述待训练场景图片的当前特征向量,根据所述当前特征向量对所述待训练场景图片进行分类,得到所述待训练场景图片的分类标签;
    获取所述待训练场景图片的标准分类标签,将所述得到的分类标签与所述标准分类标签进行比较,判断所述得到的分类标签是否正确,根据所述判断结果计算所述分类的奖励值。
  4. 如权利要求3所述的方法,其特征在于,根据所述判断结果计算所述分类的奖励值的步骤,包括:
    计算分类误差,获取分类次数,通过预设的分类奖励值的计算公式
    Figure PCTCN2017078291-appb-100001
    计算所述分类的奖励值,所述rt为第t次分类的奖励值,所述t为分类次数,所述y为所述训练得到的分类标签,所述losst为第t次分类的分类误差,所述y=maxylog p(y)表示得到的分类标签是正确的。
  5. 如权利要求1所述的方法,其特征在于,对所述当前局部观测区域的图像信息进行处理,以得到所述待分类场景图片的特征向量的步骤,包括:
    对所述当前局部观测区域的图像信息进行编码,得到局部特征向量;
    对所述局部特征向量与预先存储的特征向量执行融合操作,得到所述场景图片的特征向量。
  6. 一种室内场景的分类装置,其特征在于,所述装置包括:
    图片接收单元,用于接收输入的待分类场景图片;
    区域获取单元,用于根据预设的观测区域定位模型从所述待分类场景图片中获取当前局部观测区域;
    向量获取单元,用于对所述当前局部观测区域的图像信息进行处理,以得到所述待分类场景图片的特征向量;
    条件判断单元,用于根据所述特征向量获取所述待分类场景图片的分类预测结果,判断所述分类预测结果是否满足预设的场景图片分类条件;
    重复执行单元,用于当所述分类预测结果不满足所述场景图片分类条件时,根据所述观测区域定位模型从所述待分类场景图片中获取下一局部观测区域,并将所述下一局部观测区域设置为所述当前局部观测区域,并触发所述向量获取单元对所述当前局部观测区域的图像信息进行处理;以及
    场景分类单元,用于当所述分类预测结果满足所述场景图片分类条件时,根据所述分类预测结果获取所述待分类场景图片的分类标签。
  7. 如权利要求6所述的装置,其特征在于,所述装置还包括:
    训练区域获取单元,用于接收输入的待训练场景图片,根据预设的高斯分布从所述待训练场景图片中获取当前训练用局部观测区域;
    区域训练单元,用于根据所述当前训练用局部观测区域对所述待训练场景图片进行分类操作并计算所述分类操作的奖励值;
    循环训练单元,用于当未达到预设的训练结束条件时,根据所述高斯分布从所述待训练场景图片中获取下一训练用局部观测区域,将所述下一训练用局部观测区域设置为当前训练用局部观测区域,并触发所述区域训练单元根据所述当前训练用局部观测区域对所述待训练场景图片进行分类操作并计算所述分类操作的奖励值;以及
    定位模型建立单元,用于当达到预设的训练结束条件时,获取所有待训练场景图片中每张待训练场景图片的所述奖励值的代数和,以得到每张待训练场景图片的总奖励值,根据所述总奖励值建立总奖励值最大化的观测区域定位模 型。
  8. 如权利要求7所述的装置,其特征在于,所述区域训练单元包括:
    训练分类单元,用于对所述当前训练用局部观测区域的图像信息进行处理,得到所述待训练场景图片的当前特征向量,根据所述当前特征向量对所述待训练场景图片进行分类,得到所述待训练场景图片的分类标签;以及
    奖励值计算单元,用于获取所述待训练场景图片的标准分类标签,将所述得到的分类标签与所述标准分类标签进行比较,判断所述得到的分类标签是否正确,根据所述判断结果计算所述分类的奖励值。
  9. 如权利要求8所述的装置,其特征在于,所述奖励值计算单元包括:
    计算子单元,用于计算分类误差,获取分类次数,通过预设的分类奖励值的计算公式
    Figure PCTCN2017078291-appb-100002
    计算所述分类的奖励值,所述rt为第t次分类的奖励值,所述t为分类次数,所述y为所述训练得到的分类标签,所述losst为第t次分类的分类误差,所述y=maxylog p(y)表示得到的分类标签是正确的。
  10. 如权利要求6所述的装置,其特征在于,所述向量获取单元包括:
    编码操作单元,用于对所述当前局部观测区域的图像信息进行编码,得到局部特征向量;以及
    融合操作单元,用于对所述局部特征向量与预先存储的特征向量执行融合操作,得到所述场景图片的特征向量。
PCT/CN2017/078291 2017-03-27 2017-03-27 一种室内场景的分类方法及装置 WO2018176195A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
PCT/CN2017/078291 WO2018176195A1 (zh) 2017-03-27 2017-03-27 一种室内场景的分类方法及装置
US16/495,401 US11042777B2 (en) 2017-03-27 2017-03-27 Classification method and classification device of indoor scene

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2017/078291 WO2018176195A1 (zh) 2017-03-27 2017-03-27 一种室内场景的分类方法及装置

Publications (1)

Publication Number Publication Date
WO2018176195A1 true WO2018176195A1 (zh) 2018-10-04

Family

ID=63673984

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2017/078291 WO2018176195A1 (zh) 2017-03-27 2017-03-27 一种室内场景的分类方法及装置

Country Status (2)

Country Link
US (1) US11042777B2 (zh)
WO (1) WO2018176195A1 (zh)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108898082A (zh) * 2018-06-19 2018-11-27 Oppo广东移动通信有限公司 图片处理方法、图片处理装置及终端设备
CN111753386A (zh) * 2019-03-11 2020-10-09 北京嘀嘀无限科技发展有限公司 一种数据处理方法及装置
CN112261719A (zh) * 2020-07-24 2021-01-22 大连理智科技有限公司 一种slam技术结合深度学习的区域定位方法

Families Citing this family (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108875821A (zh) * 2018-06-08 2018-11-23 Oppo广东移动通信有限公司 分类模型的训练方法和装置、移动终端、可读存储介质
EP3660750B1 (en) * 2018-11-30 2022-01-05 Secondmind Limited Method and system for classification of data
CN111625674A (zh) * 2020-06-01 2020-09-04 联想(北京)有限公司 一种图片处理方法及装置
CN114612706B (zh) * 2020-12-04 2024-10-15 腾讯科技(深圳)有限公司 图像处理方法、装置、电子设备及可读存储介质
CN113743459B (zh) * 2021-07-29 2024-04-02 深圳云天励飞技术股份有限公司 目标检测方法、装置、电子设备及存储介质
CN113643136B (zh) * 2021-09-01 2024-06-18 京东科技信息技术有限公司 信息处理方法、系统和装置
CN113850023B (zh) * 2021-09-30 2022-11-29 深圳市瑞云科技有限公司 基于场景文件参数预估硬件配置的模型建立方法、装置
CN113762426A (zh) * 2021-11-09 2021-12-07 紫东信息科技(苏州)有限公司 一种胃镜视野丢失帧检测方法及系统
CN114781548B (zh) * 2022-05-18 2024-07-12 平安科技(深圳)有限公司 图像场景分类方法、装置、设备及存储介质
CN115035146A (zh) * 2022-05-18 2022-09-09 咪咕文化科技有限公司 图片拍摄方法、设备、存储介质及装置
CN116681997B (zh) * 2023-06-13 2024-05-17 北京数美时代科技有限公司 一种不良场景图像的分类方法、系统、介质及设备
CN118447443A (zh) * 2024-04-30 2024-08-06 浙江能珈信息技术有限公司 基于深度学习的大数据智能分析处理系统及方法

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101389004A (zh) * 2007-09-13 2009-03-18 中国科学院自动化研究所 一种基于在线学习的运动目标分类方法
CN102722520A (zh) * 2012-03-30 2012-10-10 浙江大学 一种基于支持向量机的图片重要性分类方法
CN104268546A (zh) * 2014-05-28 2015-01-07 苏州大学 一种基于主题模型的动态场景分类方法
US20160086015A1 (en) * 2007-01-09 2016-03-24 Si Corporation Method and system for automated face detection and recognition
CN105809146A (zh) * 2016-03-28 2016-07-27 北京奇艺世纪科技有限公司 一种图像场景识别方法和装置

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2012012943A1 (en) * 2010-07-28 2012-02-02 Shenzhen Institute Of Advanced Technology Chinese Academy Of Sciences Method for reconstruction of urban scenes
KR102096398B1 (ko) * 2013-07-03 2020-04-03 삼성전자주식회사 자율 이동 로봇의 위치 인식 방법
US10452924B2 (en) * 2018-01-10 2019-10-22 Trax Technology Solutions Pte Ltd. Withholding alerts due to temporary shelf occlusion

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20160086015A1 (en) * 2007-01-09 2016-03-24 Si Corporation Method and system for automated face detection and recognition
CN101389004A (zh) * 2007-09-13 2009-03-18 中国科学院自动化研究所 一种基于在线学习的运动目标分类方法
CN102722520A (zh) * 2012-03-30 2012-10-10 浙江大学 一种基于支持向量机的图片重要性分类方法
CN104268546A (zh) * 2014-05-28 2015-01-07 苏州大学 一种基于主题模型的动态场景分类方法
CN105809146A (zh) * 2016-03-28 2016-07-27 北京奇艺世纪科技有限公司 一种图像场景识别方法和装置

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108898082A (zh) * 2018-06-19 2018-11-27 Oppo广东移动通信有限公司 图片处理方法、图片处理装置及终端设备
CN108898082B (zh) * 2018-06-19 2020-07-03 Oppo广东移动通信有限公司 图片处理方法、图片处理装置及终端设备
CN111753386A (zh) * 2019-03-11 2020-10-09 北京嘀嘀无限科技发展有限公司 一种数据处理方法及装置
CN111753386B (zh) * 2019-03-11 2024-03-26 北京嘀嘀无限科技发展有限公司 一种数据处理方法及装置
CN112261719A (zh) * 2020-07-24 2021-01-22 大连理智科技有限公司 一种slam技术结合深度学习的区域定位方法
CN112261719B (zh) * 2020-07-24 2022-02-11 大连理智科技有限公司 一种slam技术结合深度学习的区域定位方法

Also Published As

Publication number Publication date
US20200019816A1 (en) 2020-01-16
US11042777B2 (en) 2021-06-22

Similar Documents

Publication Publication Date Title
WO2018176195A1 (zh) 一种室内场景的分类方法及装置
CN107169503B (zh) 一种室内场景的分类方法及装置
KR102077260B1 (ko) 확룔 모델에 기반한 신뢰도를 이용하여 얼굴을 인식하는 방법 및 장치
EP3792818A1 (en) Video processing method and device, and storage medium
US9053358B2 (en) Learning device for generating a classifier for detection of a target
WO2020253127A1 (zh) 脸部特征提取模型训练方法、脸部特征提取方法、装置、设备及存储介质
CN103093212B (zh) 基于人脸检测和跟踪截取人脸图像的方法和装置
CN104036236B (zh) 一种基于多参数指数加权的人脸性别识别方法
CN108090406B (zh) 人脸识别方法及系统
US20120093362A1 (en) Device and method for detecting specific object in sequence of images and video camera device
CN109190446A (zh) 基于三元组聚焦损失函数的行人再识别方法
CN108230291B (zh) 物体识别系统训练方法、物体识别方法、装置和电子设备
US11126827B2 (en) Method and system for image identification
CN111401374A (zh) 基于多任务的模型训练方法、字符识别方法及装置
CN112016353B (zh) 一种基于视频的人脸图像进行身份识别方法及装置
WO2021031954A1 (zh) 对象数量确定方法、装置、存储介质与电子设备
WO2022041830A1 (zh) 行人重识别方法和装置
CN111626371A (zh) 一种图像分类方法、装置、设备及可读存储介质
CN115715385A (zh) 用于预测体育运动中的队形的系统和方法
CN113327212B (zh) 人脸驱动、模型的训练方法、装置、电子设备及存储介质
CN110717407A (zh) 基于唇语密码的人脸识别方法、装置及存储介质
US20190114470A1 (en) Method and System for Face Recognition Based on Online Learning
CN114332071A (zh) 一种基于前景信息增强的视频异常检测方法
CN111241930A (zh) 一种用于人脸识别的方法及系统
CN118172546B (zh) 模型生成方法、检测方法、装置、电子设备、介质和产品

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: 17903809

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: 17903809

Country of ref document: EP

Kind code of ref document: A1

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

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 03/02/2020)

122 Ep: pct application non-entry in european phase

Ref document number: 17903809

Country of ref document: EP

Kind code of ref document: A1