WO2014194620A1 - 图像特征提取、训练、检测方法及模块、装置、系统 - Google Patents

图像特征提取、训练、检测方法及模块、装置、系统 Download PDF

Info

Publication number
WO2014194620A1
WO2014194620A1 PCT/CN2013/088563 CN2013088563W WO2014194620A1 WO 2014194620 A1 WO2014194620 A1 WO 2014194620A1 CN 2013088563 W CN2013088563 W CN 2013088563W WO 2014194620 A1 WO2014194620 A1 WO 2014194620A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
gradient
scanning
scan
size
Prior art date
Application number
PCT/CN2013/088563
Other languages
English (en)
French (fr)
Inventor
魏代猛
赵勇
黎国梁
程如中
李晶晶
全冬兵
梁浩
魏江月
Original Assignee
北京大学深圳研究生院
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 北京大学深圳研究生院 filed Critical 北京大学深圳研究生院
Publication of WO2014194620A1 publication Critical patent/WO2014194620A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00Arrangements for image or video recognition or understanding
    • G06V10/40Extraction of image or video features
    • G06V10/50Extraction of image or video features by performing operations within image blocks; by using histograms, e.g. histogram of oriented gradients [HoG]; by summing image-intensity values; Projection analysis
    • G06V10/507Summing image-intensity values; Histogram projection analysis

Definitions

  • the present invention relates to the field of image processing, and in particular, to an image feature extraction, training, detection method, module, device, and system.
  • Object detection is one of the core issues of computer vision.
  • the statistical-based object detection method mainly uses machine learning to train a classifier from a series of training data, and then uses the classifier to identify the input window.
  • feature extraction There are two key points in the object detection method, one is feature extraction, and the other is classifier design.
  • the purpose of feature extraction is to reduce the dimensionality of the data and obtain features that reflect the properties of the object, thus facilitating classification. Good features should be characterized by strong discriminating power, simple calculation, strong robustness and simple form.
  • the classifier design belongs to the category of machine learning, and its purpose is to obtain a classifier with low computational complexity and good generalization.
  • the gradient histogram feature is a feature descriptor used for object detection. It constructs features by calculating and counting gradient histograms of local regions of the image. Gradient histogram features combined with classifiers have been widely used in image recognition, especially in pedestrian detection.
  • the existing gradient histogram feature extraction method is roughly as follows: First, the grayscale image is normalized to adjust the contrast of the image, reduce the influence of local shadow and illumination changes, and suppress noise interference; Calculating the gradient of each pixel in the image (including the gradient size and gradient direction) to form a gradient image to capture the contour information to further weaken the interference of the illumination; secondly, to divide the cell unit into the gradient image, for example, the cell size can be 6 pixels* 6 pixel size; further, the gradient histogram of each cell unit is counted to form a descriptor for each cell unit; further, several cell units are grouped, for example, each block contains 9 cell units (two In the direction of 3 cell units), the gradient histogram cascade of all cell units in a block obtains the
  • an image feature extraction method includes:
  • the first gradient scan window is used to scan the original gradient image by using the first scan window to obtain an N-dimensional vector corresponding to each first scan region in the original gradient image for characterizing the gradient size and direction information of the first scan region.
  • N takes a positive integer; and performs the following process to obtain N directions images: mapping the nth dimension vector corresponding to the xth first scan area to the xth mapping unit of the nth direction image, where n ⁇ (1, 2, ..., N), x ⁇ (1, 2, ..., X), where X is the number of first scanning regions;
  • each direction is processed as follows: the direction image is divided by a pre-set size cell, each cell contains a plurality of mapping units; and the second scanning window is used to divide the divided direction image by the second preset step size Performing scanning to obtain an M dimension vector corresponding to each second scanning area in the directional image for characterizing the gradient information of the second scanning area, where M is a positive integer; combining M dimensions of each second scanning area of each direction image a vector that obtains a gradient histogram feature of the directional image;
  • a training method comprising:
  • the main gradient direction histogram feature of the image to be processed is obtained from the original gradient image, wherein the first gradient window is used to scan the original gradient image by using the first scanning window to obtain corresponding to each first scanning region in the original gradient image.
  • Performing the following process to obtain N directions images mapping the nth dimension vector corresponding to the xth first scan area to the xth mapping unit of the nth direction image, where n ⁇ (1, 2, . N), x ⁇ (1, 2, ..., X), X is the number of first scanning areas; for each of the direction images, the following processing is performed: dividing the direction image by a cell of a predetermined size, each The cell includes a plurality of mapping units; the second scanning window is used to scan the divided direction image in a second preset step to obtain a gradient corresponding to each second scanning area in the direction image for characterizing the second scanning area.
  • the M dimension vector of the size information wherein M takes a positive integer; combines the M dimension vectors of the second scan regions of each direction image to obtain a gradient histogram feature of the direction image; combines the gradient histogram features of the images in each direction to obtain the original
  • the additional gradient direction histogram features corresponding to the gradient image are trained using additional gradient direction histogram features and main gradient direction histogram features.
  • a detection method comprising:
  • the classifier obtained by the training method provided by the second aspect of the present invention detects whether an object and a position of the object are detected in the image to be detected.
  • an additional image feature extraction module including:
  • an image generating module configured to scan the original gradient image by using the first scan window in a first preset step, and the first scan area corresponding to the obtained original gradient image is used to represent the gradient of the first scan area
  • the N-dimensional vector of the direction information is processed as follows to obtain an N-direction image: the n-th dimension vector corresponding to the x-th first scan area is mapped to the x-th mapping unit of the n-th direction image, where N is taken a positive integer, n ⁇ (1, 2, ..., N), x ⁇ (1, 2, ..., X), where X is the number of first scanning regions;
  • An additional feature forming module is configured to process each direction image by dividing a direction image into cells of a preset size, each cell comprising a plurality of mapping units; and adopting a second scanning window to adopt a second preset step Longly scanning the divided direction image to obtain an M dimension vector corresponding to each second scanning area in the direction image for characterizing the gradient information of the second scanning area, where M is a positive integer; combining each direction image The M-dimensional vector of the second scanning area obtains the directional image feature; and combines the image features of each direction to obtain an additional gradient direction histogram feature corresponding to the original gradient image.
  • a training apparatus includes:
  • a sample collection module for collecting an image set of objects and a background image set
  • a feature extraction module configured to extract features of each image to be processed in the object image set and the background image set
  • a training module for training with features to obtain a classifier for distinguishing between objects and backgrounds
  • the feature extraction module includes:
  • a main feature extraction module configured to obtain an original gradient image of the image to be processed, and obtain a main gradient direction histogram feature from the original gradient image, wherein the first scan window is used to adopt a first preset step size to the original
  • the gradient image is scanned to obtain an N-dimensional vector corresponding to each first scanning area in the original gradient image for characterizing the gradient size and direction information of the first scanning area, where N takes a positive integer;
  • the additional image feature extraction module trains the trigger training module using additional gradient direction histogram features and main gradient direction histogram features.
  • a detecting apparatus comprising:
  • An image input module for obtaining an image to be detected
  • a detecting module configured to detect, by using a training method according to the second aspect of the present invention, whether the object and the location of the object are detected in the image to be detected.
  • a detection system comprising the training device according to the fifth aspect of the invention, and the detecting device according to the sixth aspect of the invention.
  • FIG. 1 is a flowchart of a training method according to Embodiment 1 of the present invention.
  • step 102 is a specific flowchart of step 102 in the training method according to Embodiment 1 of the present invention
  • FIG. 3 is a schematic diagram of a first scan area 301 in an original gradient image according to Embodiment 1 of the present invention
  • FIG. 4 is a schematic diagram of a first scan area 301 corresponding to an N dimension vector according to Embodiment 1 of the present invention.
  • FIG. 5 is a schematic diagram of mapping a first scan area to a direction image according to Embodiment 1 of the present invention
  • FIG. 6 is a schematic diagram of a second scan area corresponding to an M dimension vector according to Embodiment 1 of the present invention.
  • FIG. 7 is a schematic structural diagram of a training device according to Embodiment 1 of the present invention.
  • FIG. 8 is a schematic structural diagram of an additional image feature extraction module in a training device according to Embodiment 1 of the present invention.
  • FIG. 9 is a schematic structural diagram of a detecting apparatus according to Embodiment 1 of the present invention.
  • the object detecting method of this embodiment mainly uses a classifier to detect an object. To obtain the classifier used for the detection, it is necessary to train the training set using the training method shown in Figure 1. Referring to FIG. 1, the training method of this embodiment mainly includes:
  • Step 101 Collect an object image set and a background image set.
  • Step 102 Extract features of each image to be processed in the object image set and the background image set;
  • Step 103 Perform training using the extracted features to obtain a classifier for distinguishing objects from the background.
  • the classifier may be a linear classifier or a nonlinear classifier, such as a radial basis function kernel support vector machine classifier, and linear. Support vector machine classifiers, etc.; on the other hand, the classifier can also be a single classifier or a cascade structure classifier.
  • the step 102 specifically includes the process shown in FIG. 2:
  • Step 201 Obtain an original gradient image of the image to be processed, specifically, treat the image to be processed as a three-dimensional gray image, and normalize the image to be processed by using a gamma correction method to obtain a normalized original gradient image;
  • Step 202 Obtain a main gradient histogram feature from the original gradient image. Specifically, calculate a gradient for each pixel in the original gradient image, including a gradient size and a gradient direction, and then divide the original gradient image into cell units, and then count each a gradient histogram of cell units forming a descriptor for each cell unit. Further, several cell units are composed of blocks, for example, each block contains several cell units, and cell units may appear in two or more In the block, the gradient histogram cascading of all the cell units in one block obtains the gradient histogram descriptor of the block; finally, the gradient histogram descriptors of all the blocks in the original gradient image are cascaded to obtain the whole original The gradient histogram of the gradient image is characterized.
  • the gradient histogram feature is called the main gradient histogram feature only to distinguish the subsequent additional gradient histogram features.
  • the main gradient histogram features the corresponding gradient size and direction information. ;
  • Step 203 using an image feature extraction method as described below, to obtain an additional gradient histogram feature to perform the training of step 103 using the additional gradient histogram feature and the main gradient histogram feature, specifically, when a single classifier is required to be trained,
  • the additional gradient histogram feature and the main gradient histogram feature can be simultaneously input for training to obtain a single classifier; when the training needs to obtain the cascade structure classifier, the first level sub-classifier uses the additional gradient histogram feature, the second level The sub-classifier uses the main gradient histogram feature to cascade the first-level sub-classifier with the second-level sub-classifier to obtain a cascade structure classifier.
  • the above image feature extraction methods mainly include:
  • the first gradient scan window is used to scan the original gradient image in a first preset step, and the first scan region corresponding to the first scan region is used to represent the gradient and direction of the first scan region.
  • the N-dimensional vector of the information which is also the gradient histogram descriptor of the block proposed in step 202.
  • the original gradient image size is 64 pixels * 128 pixels
  • the first scan window size is 4 pixels * 4 pixels
  • the first preset step size L1 is 2 pixels
  • N takes a positive integer 9 so, horizontally
  • the number of scanning first scanning windows is 31, and the number of vertical scanning first scanning windows is 63, and a total of 1953 first scanning areas 301 are obtained by scanning, and each of the first scanning areas 301 corresponds to one N-dimensional vector.
  • mapping the nth dimension vector corresponding to the xth first scan area to the xth mapping unit of the nth direction image where n ⁇ (1, 2, . N), x ⁇ (1, 2, ..., X), where X is the number of first scanning regions.
  • the three first scanning areas 301 respectively correspond to three N-dimensional vectors A1, A2, and A3.
  • different first scanning areas 301 are distinguished by a center point, so the first in FIG.
  • the three points in the scan area 301 represent three different first scan areas 301.
  • FIG. 1 mapping the nth dimension vector corresponding to the xth first scan area to the xth mapping unit of the nth direction image, where n ⁇ (1, 2, . N), x ⁇ (1, 2, ..., X), where X is the number of first scanning regions.
  • the three first scanning areas 301 respectively correspond to three N-dimensional vectors A1, A2, and A3.
  • different first scanning areas 301 are distinguished by a center point, so the first in FIG.
  • the first dimension vector 501 corresponding to the N-dimensional vector A1 of the first first scan area is mapped to the first mapping unit 503 of the first direction image 502, respectively, and the first first scan is performed.
  • the second dimension vector 504 corresponding to the area A1 is mapped into the first mapping unit 506 of the second direction image 505, and the third dimension vector 507 corresponding to the first first scanning area A1 is mapped to the third direction image 508.
  • the first mapping unit 509 is sequentially performed; the first dimension vector 510 corresponding to the second first scanning area A2 is mapped to the second mapping unit 511 of the first direction image 502, and the second first scanning area is The second dimension vector 512 corresponding to A2 is mapped to the second mapping unit 513 of the second direction image 505, and the third dimension vector 514 corresponding to the second first scanning area A2 is mapped to the third direction image 508.
  • the second mapping unit 515 is sequentially performed; the third first scanning area A3 is similarly processed; finally, all the N-dimensional vectors corresponding to the first scanning area are split into corresponding direction images, for a total of 9 binary image, A direction of the image size is 31 * 63 mapping unit mapping unit;
  • the image of each direction is processed as follows: the direction image is divided into cells of a preset size, each cell contains a plurality of mapping units, wherein the cell size is 4 mapping units * 4 mapping units, then, To ensure complete division, each direction image needs to be filled to a size of 32 mapping units * 64 mapping units; the second scanning window is used to scan the divided direction image in a second preset step size to obtain each direction in the direction image.
  • M takes a positive integer 4; combines the M-dimensional vector of each second scanning area of each direction image to obtain a directional image feature; as shown in FIG.
  • each second scanning area 601 corresponds to one M-dimensional.
  • the degree vector for example, the three second scan areas 601 respectively correspond to three M dimension vectors B1, B2 and B3, wherein the M dimension vector is calculated by the bilinear interpolation algorithm, and the M dimension vector level of each direction image is obtained.
  • the direction image feature can be obtained by combining, and the corresponding image size information is only included in the direction image feature;
  • the image features in each direction are combined to obtain additional gradient histogram features corresponding to the original gradient image.
  • all the direction image features are cascaded to obtain additional gradient histogram features.
  • the training apparatus of the present embodiment may include the structure as shown in FIG. 7:
  • the sample collection module 701 is configured to collect an object image set and a background image set
  • a feature extraction module 702 configured to extract features of each image to be processed in the object image set and the background image set;
  • the training module 703 is configured to perform training by using the features extracted by the feature extraction module 702 to obtain a classifier for distinguishing objects from the background.
  • the feature extraction unit 702 includes:
  • a main feature extraction module configured to obtain an original gradient image of the image to be processed, and obtain a main gradient histogram feature from the original gradient image, wherein the main feature extraction module needs to adopt the first scan window to the first preset step size to the original
  • the gradient image is scanned to obtain an N-dimensional vector corresponding to each first scanning area in the original gradient image for characterizing the gradient size and direction information of the first scanning area, where N takes a positive integer;
  • An additional image feature extraction module is used to trigger the training module 703 to perform training using additional gradient histogram features and main gradient histogram features.
  • the additional image feature extraction module is mainly used to obtain additional gradient histogram features, which mainly include the structure shown in FIG. 8:
  • the image generation module 801 is configured to: according to the N-dimensional vector obtained by the main feature extraction module, perform the following processing to obtain N directional images: map the nth dimension vector corresponding to the xth first scan region to the nth directional image In the xth mapping unit, where n ⁇ (1, 2, ..., N), x ⁇ (1, 2, ..., X), where X is the number of first scanning regions;
  • An additional feature forming module 802 is configured to process each direction image by dividing a direction image into cells of a preset size, each cell including a plurality of mapping units, and adopting a second scanning window to be a second preset
  • the step size scans the divided direction image to obtain an M dimension vector corresponding to each second scan area in the direction image for characterizing the gradient information of the second scan area, where M is a positive integer; combining each direction image
  • the M-dimension vector of each second scanning area obtains a directional image feature; and combines the image features of each direction to obtain an additional gradient histogram feature corresponding to the original gradient image.
  • the detecting device of this embodiment mainly includes the structure as shown in FIG. 9:
  • An image input module 901 configured to obtain an image to be detected
  • the detecting module 902 is configured to detect, by using the above-mentioned training classifier, whether the object and the location of the object are detected in the image to be detected. Specifically, the detecting module 902 needs to pre-process the detected image, and extracts the main gradient histogram feature and the additional gradient histogram feature of the pre-processed image to be detected according to the content as described in the feature extraction module 702, and These features are input to the classifier obtained by the above training, so that it is possible to detect whether there is an object and the position of the object in the image to be detected.
  • the above-described training device and detecting device constitute the detecting system for detecting whether or not an object exists in the image and the position of the object in the present embodiment.
  • the difference between this embodiment and the first embodiment mainly lies in: the size of the original image and the original gradient image, the size of the first scanning window, the value of the first preset step, the value of N, the size of the cell, and the second
  • the size of the scanning window, the value of the second preset step, and the value of M may also be other values as appropriate.
  • the size of the original image and the original gradient image is not limited, and may be 256 pixels * 128 pixels or 64 pixels * 32 pixels, etc.; the first scanning window is rectangular, and the size may be 8 pixels * 8 pixels or 16 pixels * 16 pixels, etc.
  • the first preset step size is determined according to the size of the first scanning window, and the value may be 4 pixels. 8 pixels or the like, preferably, the first preset step size is 1/2 of the size of the first scan window, that is, when the first scan window is 4 pixels * 4 pixels, the first preset step L1 is 2 Pixel, or when the first scan window is 8 pixels * 8 pixels, the first preset step L1 is 4 pixels; the N dimension vector is to divide the 180° average into N direction intervals, for example, N is 3, the direction The interval is 0-60°, 61-120°, and 121-180°, so the value of N is required to be a positive integer.
  • N is 9; the size of the cell is the size of the first scan window; The second scanning window is square.
  • the size is not larger than the size of the first scanning window, for example, when the size of the first scanning window is 8 When the pixel is 8 pixels, the size of the second scanning window may be 4 pixels*4 pixels; the second preset step is similar to the first preset step, and the value is determined according to the size of the second scanning window.
  • the second preset step size is 1/2 of the size of the second scan window, that is, when the second scan window is 4 mapping units *4 mapping units, the second preset step size is 2 mapping units; M is a degree vector obtained by scanning the divided direction image in a second preset step according to the second scan window, and the value is required to be a positive integer.
  • the second preset step is Two scan window sizes of 1/2, the cell is divided into 4 dimensional vectors, preferably 4 .

Abstract

一种图像特征提取、训练、检测方法及模块、装置、系统,首先获得原始图像的原始梯度图像,并对原始梯度图像进行第一次扫描,得到每个第一扫描区域对应的N维度向量,然后将N维度向量的分量分别映射到不同方向图像中,形成N个方向图像(101);其次得到方向图像特征,并组合方向图像特征形成原始梯度图像对应的附加梯度方向直方图特征(102);在进行训练时,使用原始图像对应的主梯度方向直方图特征以及附加梯度方向直方图特征获得分类器(103)。这样,在主梯度方向直方图特征的基础上,结合了保留有原始图像中的边缘梯度信息附加梯度方向直方图特征参与训练,从而增强了训练所用特征中的边缘梯度信息,进而在物体检测时,提高了物体检测的检测率。

Description

图像特征提取、训练、检测方法及模块、装置、系统 技术领域
本发明涉及图像处理领域,具体涉及一种图像特征提取、训练、检测方法及模块、装置、系统。
背景技术
物体检测是计算机视觉的核心问题之一。基于统计学的物体检测方法主要是通过机器学习,从一系列训练数据中训练得到一个分类器,然后利用分类器对输入窗口进行识别。物体检测方法关键的有两点,其一是特征提取,其二是分类器设计。特征提取的目的是降低数据的维数,得到能反映物体属性的特征,从而方便分类。好的特征应当具有区分能力强、计算简单、鲁棒性强和形式简单等特点。分类器设计属于机器学习范畴,其目的是得到一个计算复杂度较低且推广性较好的分类器。
梯度直方图特征是一种用来进行物体检测的特征描述子,它通过计算和统计图像局部区域的梯度直方图来构成特征。梯度直方图特征结合分类器已经被广泛应用于图像识别中,尤其在行人检测中获得了极大的成功。现有的梯度直方图特征提取方法大致如下述:首先,将灰度图像进行归一化,以调节图像的对比度,降低图像局部的阴影和光照变化所造成的影响,同时抑制噪音干扰;然后,计算图像中每个像素的梯度(包括梯度大小和梯度方向)形成梯度图像,以捕获轮廓信息,进一步弱化光照的干扰;其次,对梯度图像划分细胞单元,例如,细胞单元大小可为6像素*6像素大小;再者,统计每个细胞单元的梯度直方图,形成每个细胞单元的描述子;进一步地,将几个细胞单元组成块,例如,每个块包含9个细胞单元(两个方向上均为3个细胞单元),一个块内的所有细胞单元的梯度直方图级联便得到该块的梯度直方图描述子;最后,将整个图像内所有块的梯度直方图描述子级联,即可得到整个图像的梯度直方图特征了。
但是,现有的梯度直方图特征提取方法的检测率仍然较低,限制了其在物体检测中的推广应用。
发明内容
依据本发明的第一方面提供一种图像特征提取方法,包括:
采用第一扫描窗以第一预设定步长对原始梯度图像进行扫描,得到原始梯度图像中各第一扫描区域对应的、用于表征第一扫描区域梯度大小及方向信息的N维度向量,其中,N取正整数;并进行如下处理,得到N个方向图像:将第x个第一扫描区域对应的第n维度向量映射到第n个方向图像的第x映射单位中,其中,n∈(1,2,……,N),x∈(1,2,……,X),X为第一扫描区域的数量;
对每个方向图像进行如下处理:以预设定大小的单元格划分方向图像,每个单元格包含多个映射单位;采用第二扫描窗以第二预设定步长对划分后的方向图像进行扫描,得到方向图像中各第二扫描区域对应的、用于表征第二扫描区域梯度大小信息的M维度向量,其中,M取正整数;组合每个方向图像各第二扫描区域的M维度向量,得到方向图像的梯度直方图特征;
组合各方向图像的梯度直方图特征,得到原始梯度图像对应的附加梯度方向直方图特征。
依据本发明的第二方面提供一种训练方法,包括:
采集物体图像集及背景图像集;
对物体图像集及背景图像集中各待处理图像提取特征;
使用特征进行训练,得到用于区分物体及背景的分类器,
对物体图像集及背景图像集中各待处理图像提取特征具体包括:
获得待处理图像的原始梯度图像;
由原始梯度图像得到待处理图像的主梯度方向直方图特征,其中,采用第一扫描窗以第一预设定步长对原始梯度图像进行扫描,得到原始梯度图像中各第一扫描区域对应的、用于表征第一扫描区域梯度大小及方向信息的N维度向量,其中,N取正整数;
进行如下处理,得到N个方向图像:将第x个第一扫描区域对应的第n维度向量映射到第n个方向图像的第x映射单位中,其中,n∈(1,2,……,N),x∈(1,2,……,X),X为第一扫描区域的数量;对每个所述方向图像进行如下处理:以预设定大小的单元格划分方向图像,每个单元格包含多个映射单位;采用第二扫描窗以第二预设定步长对划分后的方向图像进行扫描,得到方向图像中各第二扫描区域对应的、用于表征第二扫描区域梯度大小信息的M维度向量,其中,M取正整数;组合每个方向图像各第二扫描区域的M维度向量,得到方向图像的梯度直方图特征;组合各方向图像的梯度直方图特征,得到原始梯度图像对应的附加梯度方向直方图特征,以使用附加梯度方向直方图特征及主梯度方向直方图特征进行训练。
依据本发明的第三方面提供一种检测方法,包括:
获得待检测图像;
采用如本发明第二方面提供的训练方法所得分类器对待检测图像中是否存在物体及物体所在位置进行检测。
依据本发明的第四方面提供一种附加图像特征提取模块,包括:
图像生成模块,用于基于采用第一扫描窗以第一预设定步长对原始梯度图像进行扫描,得到的原始梯度图像中各第一扫描区域对应的、用于表征第一扫描区域梯度大小及方向信息的N维度向量,进行如下处理,得到N个方向图像:将第x个第一扫描区域对应的第n维度向量映射到第n个方向图像的第x映射单位中,其中,N取正整数,n∈(1,2,……,N),x∈(1,2,……,X),X为第一扫描区域的数量;以及,
附加特征形成模块,用于对每个方向图像进行如下处理:以预设定大小的单元格划分方向图像,每个单元格包含多个映射单位;采用第二扫描窗以第二预设定步长对划分后的方向图像进行扫描,得到方向图像中各第二扫描区域对应的、用于表征第二扫描区域梯度大小信息的M维度向量,其中,M取正整数;组合每个方向图像各第二扫描区域的M维度向量,得到方向图像特征;并组合各方向图像特征,得到原始梯度图像对应的附加梯度方向直方图特征。
依据本发明的第五方面提供一种训练装置,包括:
样本采集模块,用于采集物体图像集及背景图像集;
特征提取模块,用于对物体图像集及背景图像集中各待处理图像提取特征;以及,
训练模块,用于使用特征进行训练,得到用于区分物体及背景的分类器,
其中,特征提取模块包括:
主特征提取模块,用于获得所述待处理图像的原始梯度图像,并由所述原始梯度图像得到主梯度方向直方图特征,其中,采用第一扫描窗以第一预设定步长对原始梯度图像进行扫描,得到原始梯度图像中各第一扫描区域对应的、用于表征第一扫描区域梯度大小及方向信息的N维度向量,其中,N取正整数;以及,
如上述的附加图像特征提取模块,以触发训练模块使用附加梯度方向直方图特征及主梯度方向直方图特征进行训练。
依据本发明的第六方面提供一种检测装置,包括:
图像输入模块,用于获得待检测图像;以及,
检测模块,用于采用如本发明第二方面提供的训练方法所得分类器对待检测图像中是否存在物体及物体所在位置进行检测。
依据本发明的第七方面提供一种检测系统,包括如本发明第五方面提供的训练装置,以及如本发明第六方面提供的检测装置。
附图说明
图1是本发明实施例一的训练方法的流程图;
图2是本发明实施例一的训练方法中步骤102的具体流程图;
图3是本发明实施例一的第一扫描区域301在原始梯度图像中的示意图;
图4是本发明实施例一的第一扫描区域301对应N维度向量的示意图;
图5是本发明实施例一的第一扫描区域映射到方向图像的示意图;
图6是本发明实施例一的第二扫描区域对应M维度向量的示意图;
图7是本发明实施例一的训练装置的结构示意图;
图8是本发明实施例一的训练装置中的附加图像特征提取模块的结构示意图;
图9是本发明实施例一的检测装置的结构示意图。
具体实施方式
实施例1:
本实施例的物体检测方法主要可采用分类器对物体进行检测。而要获得检测所用分类器,需要采用如图1所示的训练方法对训练集进行训练得到。请参考图1,本实施例的训练方法主要包括:
步骤101,采集物体图像集及背景图像集;
步骤102,对物体图像集及背景图像集中各待处理图像提取特征;
步骤103,使用所提取特征进行训练,得到用于区分物体及背景的分类器,具体地,分类器可为线性分类器或非线性分类器,如径向基函数核支持向量机分类器、线性支持向量机分类器等;从另一方面来讲,分类器还可以是单一分类器或级联结构分类器。
其中,步骤102具体包括如图2所示的流程:
步骤201,获得待处理图像的原始梯度图像,具体地,将待处理图像看作一个三维灰度图像,采用伽马校正法对待处理图像进行归一化处理,得到归一化的原始梯度图像;
步骤202,由所述原始梯度图像得到主梯度直方图特征,具体地,对原始梯度图像中每个像素计算梯度,包括梯度大小和梯度方向,然后将原始梯度图像划分成细胞单元,再统计每个细胞单元的梯度直方图,形成每个细胞单元的描述子,进一步地,将几个细胞单元组成块,例如,每个块包含若干个细胞单元,细胞单元可能会出现在两个或多个块中,一个块内的所有细胞单元的梯度直方图级联便得到该块的梯度直方图描述子;最后,将原始梯度图像内所有块的梯度直方图描述子级联,即可得到整个原始梯度图像的梯度直方图特征了,该梯度直方图特征即所称主梯度直方图特征,仅是为了区别后续的附加梯度直方图特征,主梯度直方图特征中包含了相应的梯度大小及方向信息;
步骤203,采用如下述的图像特征提取方法,得到附加梯度直方图特征,以使用附加梯度直方图特征及主梯度直方图特征进行步骤103的训练,具体地,当需要训练得到单一分类器时,可将附加梯度直方图特征及主梯度直方图特征同时输入进行训练,得到单一分类器;当需要训练得到级联结构分类器时,第一级子分类器使用附加梯度直方图特征,第二级子分类器使用主梯度直方图特征,将第一级子分类器与第二级子分类器级联,得到级联结构分类器。
上述图像特征提取方法主要包括:
首先,采用第一扫描窗以第一预设定步长对原始梯度图像进行扫描,得到所述原始梯度图像中各第一扫描区域对应的、用于表征所述第一扫描区域梯度大小及方向信息的N维度向量,该N维度向量也就是步骤202中所提的块的梯度直方图描述子。如图3所示,原始梯度图像大小为64像素*128像素,第一扫描窗大小为4像素*4像素,第一预设定步长L1为2像素,N取正整数9,因此,横向扫描第一扫描窗个数为31个,纵向扫描第一扫描窗个数为63个,扫描所得共1953个第一扫描区域301,每个第一扫描区域301均对应一个N维度向量。进行如下处理,得到N个方向图像:将第x个第一扫描区域对应的第n维度向量映射到第n个方向图像的第x映射单位中,其中,n∈(1,2,……,N),x∈(1,2,……,X),X为第一扫描区域的数量。如图4所示,三个第一扫描区域301分别对应三个N维度向量A1、A2及A3,为方便说明,不同的第一扫描区域301以中心点来显示区分,因此图4中第一扫描区域301中的3个点代表了3个不同的第一扫描区域301。如图5所示,分别将第一个第一扫描区域的N维度向量A1对应的第一维度向量501映射到第一个方向图像502的第一映射单位503中,将第一个第一扫描区域A1对应的第二维度向量504映射到第二个方向图像505的第一映射单位506中,将第一个第一扫描区域A1对应的第三维度向量507映射到第三个方向图像508的第一映射单位509中,依次进行;将第二个第一扫描区域A2对应的第一维度向量510映射到第一个方向图像502的第二映射单位511中,将第二个第一扫描区域A2对应的第二维度向量512映射到第二个方向图像505的第二映射单位513中,将第二个第一扫描区域A2对应的第三维度向量514映射到第三个方向图像508的第二映射单位515中,依次进行;对第三个第一扫描区域A3也同样类似处理;最终,所有第一扫描区域对应的N维度向量均拆分到对应的方向图像中,一共为9个二进制图像,每个方向图像的大小为31映射单位*63映射单位;
其次,对每个方向图像进行如下处理:以预设定大小的单元格划分方向图像,每个单元格包含多个映射单位,其中,单元格大小为4映射单位*4映射单位,那么,为了保证完整划分,每个方向图像需要补齐为32映射单位*64映射单位的大小;采用第二扫描窗以第二预设定步长对划分后的方向图像进行扫描,得到方向图像中各第二扫描区域对应的、用于表征第二扫描区域梯度大小信息的M维度向量,其中,第二扫描窗大小为8映射单位*8映射单位,第二预设定步长L2为4映射单位,M取正整数4;组合每个方向图像各第二扫描区域的M维度向量,得到方向图像特征;如图6所示,对于大小为32映射单位*64映射单位的方向图像而言,横向第二扫描区域601为7个,纵向第二扫描区域601为15个,因此扫描单个方向图像共105个第二扫描区域601,每个第二扫描区域601均对应一个M维度向量,例如,三个第二扫描区域601分别对应三个M维度向量B1、B2及B3,其中,M维度向量通过双线性插值算法计算得到的,将每个方向图像的M维度向量级联,即可得到方向图像特征,方向图像特征中仅包含相应的梯度大小信息;
最后,组合各方向图像特征,得到原始梯度图像对应的附加梯度直方图特征,具体是将所有方向图像特征级联,即可得到附加梯度直方图特征。
相应地,本实施例的训练装置可包括如图7所示的结构:
样本采集模块701,用于采集物体图像集及背景图像集;
特征提取模块702,用于对物体图像集及背景图像集中各待处理图像提取特征;以及,
训练模块703,用于使用特征提取模块702所提取的特征进行训练,得到用于区分物体及背景的分类器,
其中,特征提取单元702包括:
主特征提取模块,用于获得待处理图像的原始梯度图像,并由原始梯度图像得到主梯度直方图特征,其中,主特征提取模块需要采用第一扫描窗以第一预设定步长对原始梯度图像进行扫描,得到原始梯度图像中各第一扫描区域对应的、用于表征所述第一扫描区域梯度大小及方向信息的N维度向量,其中,N取正整数;以及,
附加图像特征提取模块,以触发训练模块703使用附加梯度直方图特征及主梯度直方图特征进行训练。
附加图像特征提取模块主要是用于获得附加梯度直方图特征,其主要包括如图8所示的结构:
图像生成模块801,用于基于主特征提取模块所获得的N维度向量,进行如下处理,得到N个方向图像:将第x个第一扫描区域对应的第n维度向量映射到第n个方向图像的第x映射单位中,其中,n∈(1,2,……,N),x∈(1,2,……,X),X为第一扫描区域的数量;以及,
附加特征形成模块802,用于对每个方向图像进行如下处理:以预设定大小的单元格划分方向图像,每个单元格包含多个映射单位;采用第二扫描窗以第二预设定步长对划分后的方向图像进行扫描,得到方向图像中各第二扫描区域对应的、用于表征第二扫描区域梯度大小信息的M维度向量,其中,M取正整数;组合每个方向图像各第二扫描区域的M维度向量,得到方向图像特征;并组合各方向图像特征,得到原始梯度图像对应的附加梯度直方图特征。
相应地,本实施例的检测装置主要包括如图9所示的结构:
图像输入模块901,用于获得待检测图像;以及,
检测模块902,用于采用上述训练所得分类器对待检测图像中是否存在物体及物体所在位置进行检测。具体地,检测模块902需要先对待检测图像进行预处理,按照如上述特征提取模块702中描述的内容,提取预处理后的待检测图像的主梯度直方图特征及附加梯度直方图特征,并将这些特征输入上述训练所得分类器,从而可对待检测图像中是否存在物体及物体所在位置进行检测。
这样,上述训练装置及检测装置构成了本实施例的对图像中是否存在物体以及物体所在位置进行检测的检测系统。
实施例2:
本实施例与实施例一的区别主要在于:原始图像及原始梯度图像的大小、第一扫描窗大小、第一预设定步长的取值、N的取值、单元格的大小、第二扫描窗的大小、第二预设定步长的取值以及M的取值,还可以视情况采用其他数值。例如,原始图像及原始梯度图像的大小不作限制,可为256像素*128像素或64像素*32像素等;第一扫描窗为矩形,大小可以为8像素*8像素或16像素*16像素等,当第一扫描窗的尺寸越小,扫描得出的单个原始图像的第一扫描窗越多;第一预设定步长根据第一扫描窗的大小而定,取值可以为4像素、8像素等,优选情况下,第一预设定步长为第一扫描窗尺寸的1/2,即当第一扫描窗为4像素*4像素时,第一预设定步长L1为2像素,或当第一扫描窗为8像素*8像素时,第一预设定步长L1为4像素;N维度向量是为了将180°平均划分为N个方向区间,譬如N为3,方向区间依次为0-60°、61-120°、121-180°,因此N的取值要求为正整数,优选情况下,N取值为9;单元格的大小即第一扫描窗的大小;第二扫描窗为正方形,优选情况下,大小不大于第一扫描窗的大小,譬如当第一扫描窗大小为8像素*8像素时,第二扫描窗的大小可以是4像素*4像素;第二预设定步长跟第一预设定步长相似,其取值根据第二扫描窗的大小而定,优选情况下,第二预设定步长为第二扫描窗尺寸的1/2,即当第二扫描窗为4映射单位*4映射单位时,第二预设定步长为2映射单位;M是根据第二扫描窗以第二预设定步长对划分后的方向图像进行扫描得出的度向量,取值要求为正整数,在优选情况下,第二预设定步长为第二扫描窗尺寸的1/2,则单元格被划分成4维度向量,即优选为4。
在法国国立计算机及自动化研究院数据库上,相比于现有的基于梯度直方图和线性支持向量机分类器的行人检测方法,本申请的图像特征提取、训练、检测方法及模块、装置、系统若采用直方图交叉核支持向量机作为分类器,所能达到的检测率能提高4.5%,若采用线性支持向量机分类器,所能达到的检测率能提高2.5%。
以上应用了具体个例对本发明进行阐述,只是用于帮助理解本发明并不用以限制本发明。对于本领域的一般技术人员,依据本发明的思想,可以对上述具体实施方式进行变化。

Claims (10)

  1. 一种图像特征提取方法,其特征在于,包括:
    用第一扫描窗以第一预设定步长对原始梯度图像进行扫描,得到所述原始梯度图像中各第一扫描区域对应的、用于表征所述第一扫描区域梯度大小及方向信息的N维度向量,其中,N取正整数;并进行如下处理,得到N个方向图像:将第x个第一扫描区域对应的第n维度向量映射到第n个方向图像的第x映射单位中,其中,n∈(1,2,……,N),x∈(1,2,……,X),X为第一扫描区域的数量;
    对每个所述方向图像进行如下处理:以预设定大小的单元格划分所述方向图像,每个所述单元格包含多个映射单位;采用第二扫描窗以第二预设定步长对所述划分后的方向图像进行扫描,得到所述方向图像中各第二扫描区域对应的、用于表征所述第二扫描区域梯度大小信息的M维度向量,其中,M取正整数;组合每个所述方向图像各所述第二扫描区域的M维度向量,得到方向图像特征;
    组合各所述方向图像特征,得到所述原始梯度图像对应的附加梯度方向直方图特征。
  2. 如权利要求1所述的图像特征提取方法,其特征在于,所述M维度向量通过双线性插值算法计算得到。
  3. 如权利要求1或2所述的图像特征提取方法,其特征在于,所述第一扫描窗大小为4像素*4像素;所述第一预设定步长为2像素; N的取值为9;所述单元格大小为4映射单位*4映射单位;所述第二扫描窗大小为8映射单位*8映射单位;所述第二预设定步长为4映射单位;M的取值为4。
  4. 一种训练方法,包括:
    采集物体图像集及背景图像集;
    对所述物体图像集及背景图像集中各待处理图像提取特征;
    使用所述特征进行训练,得到用于区分物体及背景的分类器,
    其特征在于,对所述物体图像集及背景图像集中各待处理图像提取特征具体包括:
    获得所述待处理图像的原始梯度图像;
    由所述原始梯度图像得到待处理图像的主梯度方向直方图特征,其中,采用第一扫描窗以第一预设定步长对所述原始梯度图像进行扫描,得到所述原始梯度图像中各第一扫描区域对应的、用于表征所述第一扫描区域梯度大小及方向信息的N维度向量,其中,N取正整数;
    进行如下处理,得到N个方向图像:将第x个第一扫描区域对应的第n维度向量映射到第n个方向图像的第x映射单位中,其中,n∈(1,2,……,N),x∈(1,2,……,X),X为第一扫描区域的数量;对每个所述方向图像进行如下处理:以预设定大小的单元格划分所述方向图像,每个所述单元格包含多个映射单位;采用第二扫描窗以第二预设定步长对所述划分后的方向图像进行扫描,得到所述方向图像中各第二扫描区域对应的、用于表征所述第二扫描区域梯度大小信息的M维度向量,其中,M取正整数;组合每个所述方向图像各所述第二扫描区域的M维度向量,得到方向图像特征;组合各所述方向图像特征,得到所述原始梯度图像对应的附加梯度方向直方图特征,以使用所述附加梯度方向直方图特征及主梯度方向直方图特征进行训练。
  5. 如权利要求4所述的训练方法,其特征在于,所述分类器为线性分类器或非线性分类器。
  6. 一种检测方法,其特征在于,包括:
    获得待检测图像;
    采用如权利要求4或5所述的训练方法所得分类器对待检测图像中是否存在物体及物体所在位置进行检测。
  7. 一种附加图像特征提取模块,其特征在于,包括:
    图像生成模块,用于基于采用第一扫描窗以第一预设定步长对原始梯度图像进行扫描,得到的所述原始梯度图像中各第一扫描区域对应的、用于表征所述第一扫描区域梯度大小及方向信息的N维度向量,进行如下处理,得到N个方向图像:将第x个第一扫描区域对应的第n维度向量映射到第n个方向图像的第x映射单位中,其中,N取正整数,n∈(1,2,……,N),x∈(1,2,……,X),X为第一扫描区域的数量;以及,
    附加特征形成模块,用于对每个所述方向图像进行如下处理:以预设定大小的单元格划分所述方向图像,每个所述单元格包含多个映射单位;采用第二扫描窗以第二预设定步长对所述划分后的方向图像进行扫描,得到所述方向图像中各第二扫描区域对应的、用于表征所述第二扫描区域梯度大小信息的M维度向量,其中,M取正整数;组合每个所述方向图像各所述第二扫描区域的M维度向量,得到方向图像特征;并组合各所述方向图像特征,得到所述原始梯度图像对应的附加梯度方向直方图特征。
  8. 一种训练装置,包括:
    样本采集模块,用于采集物体图像集及背景图像集;
    特征提取模块,用于对所述物体图像集及背景图像集中各待处理图像提取特征;以及,
    训练模块,用于使用所述特征进行训练,得到用于区分物体及背景的分类器,
    其特征在于,所述特征提取单元包括:
    主特征提取模块,用于获得所述待处理图像的原始梯度图像,并由所述原始梯度图像得到主梯度方向直方图特征,其中,采用第一扫描窗以第一预设定步长对所述原始梯度图像进行扫描,得到所述原始梯度图像中各第一扫描区域对应的、用于表征所述第一扫描区域梯度大小及方向信息的N维度向量,其中,N取正整数;以及,
    如权利要求7所述的附加图像特征提取模块,以触发所述训练模块使用附加梯度方向直方图特征及主梯度方向直方图特征进行训练。
  9. 一种检测装置,其特征在于,包括:
    图像输入模块,用于获得待检测图像;以及,
    检测模块,用于采用如权利要求4或5所述的训练方法所得分类器对待检测图像中是否存在物体及物体所在位置进行检测。
  10. 一种检测系统,其特征在于,包括如权利要求8所述的训练装置,以及如权利要求9所述的检测装置。
PCT/CN2013/088563 2013-06-04 2013-12-05 图像特征提取、训练、检测方法及模块、装置、系统 WO2014194620A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201310218938.6A CN103345631B (zh) 2013-06-04 2013-06-04 图像特征提取、训练、检测方法及模块、装置、系统
CN201310218938.6 2013-06-04

Publications (1)

Publication Number Publication Date
WO2014194620A1 true WO2014194620A1 (zh) 2014-12-11

Family

ID=49280426

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2013/088563 WO2014194620A1 (zh) 2013-06-04 2013-12-05 图像特征提取、训练、检测方法及模块、装置、系统

Country Status (2)

Country Link
CN (1) CN103345631B (zh)
WO (1) WO2014194620A1 (zh)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109344768A (zh) * 2018-09-29 2019-02-15 南京理工大学 基于巡检机器人的指针式断路器识别方法
CN110751084A (zh) * 2019-10-17 2020-02-04 中国计量大学 一种基于行扫描统计向量的步态特征表示和提取方法
CN111428778A (zh) * 2020-03-20 2020-07-17 华侨大学 一种胎儿颜面部标准切面自动分类法
CN113505815A (zh) * 2021-06-15 2021-10-15 杭州电子科技大学 一种基于多特征融合的蚊虫识别方法

Families Citing this family (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103345631B (zh) * 2013-06-04 2016-12-28 北京大学深圳研究生院 图像特征提取、训练、检测方法及模块、装置、系统
CN104751163B (zh) 2013-12-27 2018-06-19 同方威视技术股份有限公司 对货物进行自动分类识别的透视检查系统和方法
CN105809182B (zh) * 2014-12-31 2019-12-13 中国科学院深圳先进技术研究院 一种图像分类的方法及装置
CN104463180B (zh) * 2014-12-31 2018-01-12 深圳市华星光电技术有限公司 识别导电粒子压痕的方法
CN106033550B (zh) * 2015-03-16 2019-09-06 北京大学 目标跟踪方法及装置
CN107316131A (zh) * 2017-06-09 2017-11-03 国网重庆市电力公司电力科学研究院 一种基于图像识别的电能计量装置安装工艺质量检测系统
CN112560935B (zh) * 2020-12-11 2022-04-01 上海集成电路装备材料产业创新中心有限公司 一种提升缺陷检测性能的方法
CN116434220B (zh) * 2023-04-24 2024-02-27 济南大学 基于描述符和AdaBoost算法的三维物体分类方法及系统

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101894262A (zh) * 2009-05-20 2010-11-24 索尼株式会社 对图像进行分类的方法和设备
JP2011096081A (ja) * 2009-10-30 2011-05-12 Fujitsu Ltd 画像処理方法及び装置並びにプログラム
CN102169544A (zh) * 2011-04-18 2011-08-31 苏州市慧视通讯科技有限公司 基于多特征融合的人脸遮挡检测方法
CN103345631A (zh) * 2013-06-04 2013-10-09 北京大学深圳研究生院 图像特征提取、训练、检测方法及模块、装置、系统

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101655910A (zh) * 2008-08-21 2010-02-24 索尼(中国)有限公司 训练系统、训练方法及检测方法
EP2583216B1 (en) * 2010-06-17 2017-09-13 Valeo Schalter und Sensoren GmbH Method of predetermined road object recognition in images and vehicle driving assistance device
CN102479329A (zh) * 2010-11-22 2012-05-30 索尼公司 分类器生成装置和方法,检测图像中的对象的装置和方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101894262A (zh) * 2009-05-20 2010-11-24 索尼株式会社 对图像进行分类的方法和设备
JP2011096081A (ja) * 2009-10-30 2011-05-12 Fujitsu Ltd 画像処理方法及び装置並びにプログラム
CN102169544A (zh) * 2011-04-18 2011-08-31 苏州市慧视通讯科技有限公司 基于多特征融合的人脸遮挡检测方法
CN103345631A (zh) * 2013-06-04 2013-10-09 北京大学深圳研究生院 图像特征提取、训练、检测方法及模块、装置、系统

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109344768A (zh) * 2018-09-29 2019-02-15 南京理工大学 基于巡检机器人的指针式断路器识别方法
CN110751084A (zh) * 2019-10-17 2020-02-04 中国计量大学 一种基于行扫描统计向量的步态特征表示和提取方法
CN110751084B (zh) * 2019-10-17 2023-04-18 中国计量大学 一种基于行扫描统计向量的步态特征表示和提取方法
CN111428778A (zh) * 2020-03-20 2020-07-17 华侨大学 一种胎儿颜面部标准切面自动分类法
CN111428778B (zh) * 2020-03-20 2023-04-07 华侨大学 一种胎儿颜面部标准切面自动分类法
CN113505815A (zh) * 2021-06-15 2021-10-15 杭州电子科技大学 一种基于多特征融合的蚊虫识别方法

Also Published As

Publication number Publication date
CN103345631B (zh) 2016-12-28
CN103345631A (zh) 2013-10-09

Similar Documents

Publication Publication Date Title
WO2014194620A1 (zh) 图像特征提取、训练、检测方法及模块、装置、系统
WO2020186886A1 (zh) 一种人脸识别模型的生成方法及设备
CN101980242B (zh) 人脸面相判别方法、系统及公共安全系统
WO2018174623A1 (ko) 가상 3차원 심층 신경망을 이용하는 영상 분석 장치 및 방법
Tang et al. Robust image hash function using local color features
WO2022011977A1 (zh) 一种网络异常检测方法、系统、终端以及存储介质
CN108614999B (zh) 基于深度学习的眼睛睁闭状态检测方法
WO2014069822A1 (en) Apparatus and method for face recognition
WO2023082784A1 (zh) 一种基于局部特征注意力的行人重识别方法和装置
WO2021080233A1 (ko) 이미지 복원 방법 및 장치
CN106056083B (zh) 一种信息处理方法及终端
US20200279166A1 (en) Information processing device
EP3756160A1 (en) System and method for fast object detection
WO2019104627A1 (zh) 一种图像的灰阶分类方法、装置以及可读存储介质
WO2023068795A1 (ko) 이미지 분석을 이용한 메타버스 생성 장치 및 방법
CN107742094A (zh) 提高人证比对结果的图像处理方法
CN112804253A (zh) 一种网络流量分类检测方法、系统及存储介质
CN110032946A (zh) 一种基于机器视觉的铝/铝泡罩包装药片识别与定位方法
WO2015196878A1 (zh) 一种电视虚拟触控方法及系统
CN108898098A (zh) 基于监控平台的早期视频烟雾检测方法
Laroca et al. A first look at dataset bias in license plate recognition
US11481919B2 (en) Information processing device
Lian Pedestrian detection using quaternion histograms of oriented gradients
Zhao et al. Hand detection using multi-resolution HOG features
WO2018113206A1 (zh) 一种图像处理方法及终端

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: 13886442

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: 13886442

Country of ref document: EP

Kind code of ref document: A1