WO2012151755A1 - 商标检测与识别的方法 - Google Patents

商标检测与识别的方法 Download PDF

Info

Publication number
WO2012151755A1
WO2012151755A1 PCT/CN2011/073985 CN2011073985W WO2012151755A1 WO 2012151755 A1 WO2012151755 A1 WO 2012151755A1 CN 2011073985 W CN2011073985 W CN 2011073985W WO 2012151755 A1 WO2012151755 A1 WO 2012151755A1
Authority
WO
WIPO (PCT)
Prior art keywords
feature
trademark
point
image
features
Prior art date
Application number
PCT/CN2011/073985
Other languages
English (en)
French (fr)
Inventor
卢汉清
王金桥
傅建龙
Original Assignee
中国科学院自动化研究所
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 中国科学院自动化研究所 filed Critical 中国科学院自动化研究所
Priority to PCT/CN2011/073985 priority Critical patent/WO2012151755A1/zh
Publication of WO2012151755A1 publication Critical patent/WO2012151755A1/zh

Links

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/50Information retrieval; Database structures therefor; File system structures therefor of still image data
    • G06F16/53Querying
    • G06F16/532Query formulation, e.g. graphical querying
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/50Information retrieval; Database structures therefor; File system structures therefor of still image data
    • G06F16/58Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
    • G06F16/583Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content
    • G06F16/5838Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content using colour

Definitions

  • the present invention relates to multimedia content analysis and retrieval.
  • Traditional image search engines such as Google, Yahoo, Bing, etc., according to the relevance of the related text information of the network image and the query keywords, are sorted to present the search results to the user.
  • traditional image search engines implement a text-to-image retrieval process.
  • the recognition of trademarks is an image-to-text recognition process, which belongs to the domain of image understanding. If the user takes a photo of a brand's trademark after using the mobile phone, the system can automatically detect and identify the trademark, return the trademark name information to the user and retrieve it on the Internet or in a local database. Such as: the latest styles, prices, businesses, similar products and other complete information, will greatly promote the marketing of goods.
  • the automatic detection and identification of trademarks is the basis for constructing an e-commerce product recommendation system.
  • this method of identifying trademarks based on image similarity can provide an effective means for the relevant departments to prevent the registration of similar trademarks, provide effective trademark copyright protection, and shorten the certification period of trademarks.
  • a method for detecting and identifying a trademark includes the steps of:
  • the extracted point features, contour features and regional features are respectively clustered to obtain a visual codebook;
  • the hierarchical segmentation algorithm using mean moving is used to segment the query image containing the trademark; the recognition results corresponding to all the segmented region queries are sorted according to the score, and the final recognition result is obtained.
  • the invention solves the problem that the recognition rate of the same trademark is reduced due to different backgrounds. Since each sub-region is relatively complete, the content of the expression is relatively semantically clear, including possible trademark images, which reduces the difference from the database image and improves the recognition rate.
  • the final average accuracy of the present invention is up to 98%, and the recognition rate is increased by 7 90% and 17%, respectively, compared with the basic method and the current international advanced method.
  • Figure 1 is a flow chart of the algorithm of the present invention
  • Figure 2 is a different type of trademark image
  • 3 is a schematic diagram of multi-view modeling based on feature points
  • Figure 4 is a statistical analysis of the distribution of the nearest contour sampling points of the letters in the log polar coordinate system;
  • Figure 5 is a schematic diagram of the structure of the inverted document;
  • Figure 6 is a schematic diagram of the query process
  • Figure 7 is a trademark database
  • Figure 8 is a schematic diagram of the segmentation, detection and identification of the trademark image
  • Figure 9 is a schematic diagram of comparison of several algorithms.
  • the invention analyzes the image content visually by extracting a combination of low-level features with spatially high correlation in the image, and proposes an automatic detection and recognition algorithm for the trademark.
  • the patent consists of three parts: (1) multi-view modeling based on feature points; (2) feature clustering and inversion index establishment; (3) automatic recognition of trademark images by region search and weak geometry restriction. Algorithm flow chart See Figure 1.
  • the "point feature” is quite rich, saying that such trademarks are "point type.” Like Windows, Pepsi and Bouigues, although the three trademarks are all circular in appearance, the internal colors are extremely rich, and the grayscale histogram features of the region have strong discriminating power, which is called “regional type”.
  • point features such as scale-independent feature points SI.FT, fast robust feature points SURF or affine-invariant scale-independent feature points ASIFT, have been shown to perform best in feature matching. , but only one feature is not enough to model the entire trademark.
  • the contour shape of the trademark and the color information of the area can overcome more kinds of image changes, which is a good complement to the "point feature”. Therefore, the present invention simultaneously extracts three kinds of spatially highly correlated low-level features based on feature points to perform multi-view modeling on the trademark image.
  • FIG. 1 The specific implementation of the present invention will be described in detail below with reference to FIG.
  • a 128-dimensional SIFT feature point is extracted for each trademark image.
  • the Canny operator is used to extract the edges of the trademark image, and each closed edge is called a contour.
  • h(k) # ⁇ pj: pj e bin(k) ⁇
  • represents a sample point on the nearest contour, which is the ⁇ th interval in the log polar coordinate system, which is the statistical distribution of the sample points.
  • Histogram. "# " is an operator that counts the number of elements in a collection. Take the feature point "a” in the letter “A” as an example, see Figure 4.
  • the nearest contour is defined as: The smallest of the contours of all contours surrounding the SIFT feature points.
  • the sampling interval of the most recent upsampled point is set to 10.
  • H ⁇ ® W contour contour I point e contour® W area area I point e area
  • £ feature points on the three kinds of highly relevant spatial area represents the above-mentioned points
  • the contour region and the histogram , w point, w outline, ⁇ area are the fusion weights of the three types of features
  • ® represents the feature fusion.
  • the present invention uses hierarchical hierarchical K-means to cluster and obtain visual codebooks. At the beginning of clustering, there are 10 features for each class. The visual codebook and weak geometric limit scores are then encoded into the inverted index document to improve retrieval and recognition speed. See Figure 5 for the structure of the inverted index document. (3) The specific calculation of the weak geometric limit score and the process of reordering based on geometric information will be described in detail.
  • the present invention proposes a region search method to solve the problem that the recognition rate of the same trademark is reduced due to different backgrounds.
  • the Mean-Shift-based hierarchical segmentation method is used to segment the query images that may contain the trademark.
  • the segmentation process roughly divides the image into 5-20 sub-regions, each of which is relatively complete, and the content of the expression is relatively semantically explicit, including possible trademark images, which reduces the difference from the database image.
  • the method in (1) is applied to each sub-area to perform multi-view modeling based on feature points, and clusters are formed according to the method described in (2) to form a "word", that is, a visual codebook, and in the established inverted row. Query on the index. Each sub-area will get the most similar recognition result after querying. See Figure 6 for the schematic diagram of the query process.
  • the present invention proposes a weak geometric constraint method to reorder the top 20 scores with the highest scores.
  • the specific implementation of the weak geometry limitation is described in detail below.
  • each contour in the database image its internal "point feature" is projected to the X and Y coordinate directions, respectively.
  • the Y coordinates are labeled from 1 to n in order from bottom to top, which is called natural order.
  • the actual order of its internal "point features" in the X and Y coordinate directions can be obtained for each contour in each query sub-region as described above.
  • SIFT feature descriptors and Euclidean distance metrics can get mutual Matching feature point pairs.
  • the two contours with the largest number of SIFT feature point matches are called matching contours.
  • contour c in the query sub-area and the contour in the database image are a pair of matching contours, and their weaknesses are: And, +1 is the two adjacent "point features" in the sub-area of a certain query, and the coordinate values of the projection in the X direction are smaller than ;; 0 (p and 0(p) are the same as p q , i and p q , i+ , the coordinates of the matching point. If the OO ) projection has a larger coordinate value in the X direction than , indicating that the actual order is inconsistent with the natural order, giving it a penalty score of (-1).
  • the weak geometric limit scores of all n matching contours in the two images are summed and added to the search score of the first stage. Reorder the top 20 most similar recognition results by new scores.
  • the best recognition result obtained by each sub-area query is counted.
  • the scores of the same recognition result obtained by different sub-areas are accumulated, and the recognition results corresponding to all the divided sub-areas are sorted according to the score, and the highest one is the final recognition result.
  • Figure 7 shows the complete process of segmentation, detection and identification of the trademark image of the query.
  • the first column is the original query image
  • the red frame indicates the area where the target trademark is located
  • the second column is the result of the segmentation
  • the third column is the result of the recognition.
  • the accuracy rate is defined as (TP+TN) / (P+N).
  • TP true positive
  • TN true negative
  • P positive sample
  • N negative sample.
  • Table 1 shows the accuracy values for several typical trademarks and the average accuracy values for all 62 trademarks on the database.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Databases & Information Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Library & Information Science (AREA)
  • Mathematical Physics (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
  • Image Analysis (AREA)

Description

商标检测与识别的方法 技术领域
本发明涉及多媒体内容分析与检索。
背景技术
随着网络技术的发展和数字媒体技术的普及, 图像作为信息传递的 最重要载体,己经深入到日常生活的各个方面。商标是最常见的一类图像, 它是商业标志的简称, 是用户了解企业品牌的最直接手段。随着社会经济 的发展, 商标已成为商品、 服务质量、 商家信誉和实力的体现, 同时也是 企业知识产权的重要组成部分。如何对商标图像进行有效的组织管理, 构 建高效的商品推荐系统, 让用户快速、准确地查找到该品牌的更多相关信 息;同时也为有关部门提供有效的商标版权保护已成为目前亟需解决的一 大难题。
传统的图像搜索引擎如 Google, Yahoo, Bing等, 根据网络图像的相 关文本信息与査询关键词的相关程度,经过排序,将检索结果呈现给用户。 通过文字之间的简单匹配,传统的图像搜索引擎实现了由文字到图像的检 索过程。而商标的识别则是由图像到文字的识别过程, 属于图像理解的范 畴。如果当用户使用手机对某种品牌的商标进行拍照后, 系统能够自动检 测、识别该商标, 把商标的名称信息返回给用户并在互联网或本地数据库 中检索。 诸如: 最新款式、 价格、 商家、 同类商品等完整的相关信息, 将 会对商品的营销起到极大的促进作用。 因此, 商标的自动检测与识别是构 建电子商务商品推荐系统的基础。 同时, 这种根据图像相似度对商标进行 识别的方法, 可以为有关部门防止相似商标的注册、提供有效的商标版权 保护以及缩短商标的认证周期提供有效的手段。
发明内容
本发明的目的是提供一种基于特征点的多视角建模和区域搜索的商标 检测和识别方法。
为实现上述目的, 一种商标检测与识别的方法, 包括步骤:
提取商标图像的点特征、 轮廓特征和区域特征;
对提取的点特征、 轮廓特征和区域特征分别类聚得到视觉码本; 采用均值移动的层次分割算法对含有商标的査询图像进行分割; 将所有分割的区域查询对应的识别结果按分数高低排序,得到最终识 别结果。
本发明解决了同一个商标由于背景不同而带来的识别率下降的问题。 由于每个子区域相对完整, 表达的内容在语义上相对明确, 包含可能的商 标图像, 减小了与数据库图像的差异, 提高了识别率。本发明最终的平均 准确率可达 98%, 识别率与基本方法和现阶段国际先进方法相比分别提高 7 90%和 17%。 附图说明
图 1是本发明的算法流程图;
图 2是不同类型的商标图像;
图 3是基于特征点的多视角建模示意图;
图 4是对数极坐标系下的字母 , 的最近轮廓采样点分布情况统计; 图 5是倒排文档结构示意图;
图 6是査询过程示意图;
图 7是商标数据库;
图 8商标图像的分割、 检测和识别示意图;
图 9是几种算法的比较示意图。
具体实施方式
本发明通过提取图像中具有空间上高度相关的低层特征组合, 从视觉 上分析图像内容, 提出了一种商标的自动检测和识别算法。专利包括 3个 部分: (1)基于特征点的多视角建模; (2)特征聚类和倒排索引的建立; (3) 采用区域搜索和弱几何限制的方法对商标图像进行自动识别。算法流程图 参见图 1。
(1)基于特征点的多视角建模
本专利在对大量商标图像进行分析整理的基础上,按照能最好地描述 该商标的局部特征, 包括 "点特征"、 "轮廓特征"和 "区域特征", 把商 标分组命名为 "点型"、 "轮廓型"和"区域型"。参见图 2,像 Adidas、 Canon、 Hp这类商标, 商标图像的主体部分是英文字母, 图像信息简单, 一般情况 下无法提取到足够的 "点特征"对其建模。 但是, 一些研究表明 "轮廓特 征"却对字母、数字、文字等有很好的识别效果, 因此,称这类商标是 "轮 廓型"。又如 Kappa、 B丽和麦当劳, "点特征"比较丰富,称这类商标是 "点 型"。 再如 Windows, Pepsi和 Bouigues, 尽管 3个商标的外观都是圆形, 但 内部色彩极其丰富, 区域的灰度直方图特征具有较强的辨别力, 称这类商 标是"区域型 "。在实际应用中,尽管"点特征",如尺度无关的特征点 SI.FT、 快速鲁棒特征点 SURF或仿射不变的尺度无关的特征点 ASIFT, 已经被证 明在特征匹配方面性能最佳, 但是仅用一种特征不足以对整幅商标建模。 而商标的轮廓形状和区域的颜色信息可以克服更多种类的图像变化, 对 "点特征"是很好的补充。 因此, 本发明同时提取基于特征点的三种空间 上高度相关的低层特征对商标图像进行多视角建模。 下面结合附图 3对本 发明的具体实现方式详细介绍。
首先, 提取每幅商标图像 128维的 SIFT特征点。 采用 Canny算子提 取商标图像的边缘, 每一个封闭的边缘称为轮廓。
其次, 对每一个特征点 , 提取其最近轮廓上采样点的分布情况。 分 布情况通过对数极坐标系下统计得到的直方图表达, 见下式:
h(k) = # {pj: pj e bin(k)} 其中, ·代表最近轮廓上的一个采样点, 是对数极坐标系下的第 κ 个区间, 是统计得到的采样点分布情况的直方图。 "# "是计算集合中 元素个数的运算符。 以字母 "A" 中的特征点 "a" 为例, 参见图 4。 以 目标点 "a"为坐标系的圆心, 这个坐标系将目标点的周围邻域分为角度 上 m个等级、半径方向上 n个等级, 即分割成 m*n个子区间。应用中, m = 12, n : 5。 然后分别统计每个子区域内像素数目, 并将其数值化为 m*n 的矩阵。 最近轮廓定义为: 包围该 SIFT特征点的所有轮廓中面积最小的 轮廓。 最近轮廓上采样点的采样间隔设为 10。
最后,提取最近轮廓中商标图像的灰度直方图以表达不同颜色的分布 情况。 灰度直方图的维数设定为 36。 至此, 基于特征点的三种空间上高 相关的低层特征提取完成并按下式进行融合-
H = ^^点 ® W轮廓 轮廓 I点 e轮廓 ® W区域 区域 I点 e区域 其中, 代表基于特征点的多视角建模的直方图, ½、 轮廓 |点£轮廓、 ¾区域 |点£区域代表上面提到的三种空间上高度相关的点、轮廓和区域的特征 直方图, w点、 w轮廓、 ^区域是三类特征的融合权重, ®代表特征融合。
(2)特征聚类和倒排索引的建立 针对以上提取的三种特征, 本发明采用分层次的均值聚类 (hierarchical K- means)方法分别聚类得到视觉码本。 聚类开始时, 设定 每类初始有 10个特征。 然后, 将视觉码本和弱几何限制分数编码进倒排 索引文档,以提高检索和识别速度。 倒排索引文档的结构示意图参见图 5。 (3)将详细介绍弱几何限制分数的具体计算和基于几何信息重新排序的过 程。
(3)采用区域搜索和弱几何限制的方法对商标图像进行自动识别 本发明提出区域搜索方法, 以解决同一个商标由于背景不同而带来的 识别率下降的问题。
首先, 采用基于均值移动 (Mean- Shift)的层次分割方法对可能含有商 标的査询图像进行分割。 分割过程大致把图像分成 5- 20个子区域, 每个 区域相对完整,表达的内容在语义上相对明确, 包含可能的商标图像, 减 小了与数据库图像的差异。
其次, 对每个子区域应用(1)中的方法进行基于特征点的多视角建模, 按 (2)中描述的方法聚类形成"词", 即视觉码本, 并在已建立的倒排索引 上进行查询。每一个子区域经过査询都会得到一个最相似的识别结果, 查 询过程示意图参见图 6。
继而, 本发明提出一种弱几何限制方法对前 20个分数最高的识别结 果进行重新排序。 以下详细介绍弱几何限制的具体实现。
对于数据库图像中的每一个轮廓, 将其内部的 "点特征" 向 X和 Y坐 标方向分别做投影。 按照 X坐标从左到右的顺序, Y坐标从下到上的顺序 把特征点标记为 1至 n, 称为自然顺序。 在检索阶段, 对于每一个查询子 区域中的每一个轮廓按上述方法可以得到其内部 "点特征"在 X和 Y坐标 方向上的实际顺序。 采用 SIFT特征描述子和欧式距离度量可以得到相互 匹配的特征点对。 含有 SIFT特征点匹配数目最多的两个轮廓称为匹配的 轮廓。假设查询子区域中的轮廓 c和数据库图像中的轮廓 是一对匹配的 轮廓, 它们的弱几 算:
Figure imgf000007_0001
和 ,+1是某一个査询的子区域中相邻的两个 "点特征", 投影在 X 方向上的坐标值小于; ,; 0(p 和 0(p )是与 pq,i和 pq,i+、匹配点的坐标。 如果 OO )投影在 X方向上的坐标值大于
Figure imgf000007_0002
,表明实际顺序与自然顺 序不一致, 赋予其(-1 )的惩罚分。 把两幅图像中所有的 n个匹配轮廓的 弱几何限制分数累计求和, 并与第一阶段的检索分数相加。按照新的分数 对前 20个最相似的识别结果重新排序。
最后, 统计每一个子区域査询得到的最佳识别结果。 将不同子区域査 询得到的相同的识别结果的分数累加,把所有分割的子区域查询对应的识 别结果按分数高低进行排序, 最高的一项为最终识别结果。
为了更好地评估本发明提出的方法, 我们收集了 62个品牌的商标作 为数据库图像, 参见图 7。 每个品牌有 10至 15个尺度不一, 角度变化的 商标样本。 图 8显示了本方法对查询的商标图像进行分割、检测和识别的 完整过程。 第一列是原始查询图像, 红框示意了目标商标所在的区域; 第 二列是分割的结果; 第三列是识别的结果。
我们采用准确率和识别率两个指标测试方法的性能。 准确率定义为 (TP+TN) / (P+N)。 其中, TP代表真阳性、 TN代表真阴性、 P代表正样本、 N代表负样本。表 1给出了几个典型商标的准确率数值和数据库上全体 62 种商标的平均准确率数值。
表 1. 准确率
商标名称 准确率
AMD 0-996
CANON 0.985
DHL 0 94
GEELY 0-994
INTEL 0.982
平均 0.98 由于本专利实现的方法对于错误检测的拒绝率相当高, 即 TN 的值很 大, 因此平均准确率可达 98%。 为了进一步评估方法性能, 我们定义识别 率为 TP/P, 并将本专利的方法与基本方法和现阶段国际先进方法进行了 比较, 参见图 9。 我们用于比较的基本方法是图像检索领域普遍采用的提 取 SIFT特征对图像建模,整个图像进行查询的方法;国际先进方法是 2009 年微软亚洲研究院 (MSRA)提出的 Bundling Features方法。 实验表明, 本 专利提出的方法与基本方法和现阶段国际先进方法相比, 平均提高了 90% 和 17%。

Claims

权 利 要 求
1.一种商标检测与识别的方法, 包括步骤:
提取商标图像的点特征、 轮廓特征和区域特征;
对提取的点特征、 轮廓特征和区域特征分别类聚得到视觉码本; 采用均值移动的层次分割算法对含有商标的査询图像进行分割; 将所有分割的区域查询对应的识别结果按分数高低排序,得到最终识 别结果。
2.根据权利要求 1所述的方法, 其特征在于所述轮廓特征包括形状上 下文。
3.根据权利要求 1所述的方法, 其特征在于所述区域特征包括灰度直 方图。
4.根据权利要求 1所述的方法, 其特征在于所述点特征包括尺度无关 的特征点 SIFT、 快速鲁棒特征点 SURF或仿射不变的尺度无关的特征点 ASIFT。
5.根据权利要求 1所述的方法, 其特征在于所述提取商标图像的点特 征、 轮廓特征和区域特征包括:
提取每幅商标图像 128维的 SIFT特征点;
对每个特征点, 提取其最近轮廓上采样点的分布情况, 所述分布情况 通过对数极坐标统计得到直方图;
提取最近轮廓中商标图像的灰度直方图, 以表达不同颜色的分布情 况。
6.根据权利要求 5所述的方法,其特征在于采用 Canny算法提取商标图 像的边缘。
7.根据权利要求 5所述的方法, 其特征在于最近轮廓上采样点的采样 间隔设为 10。
8.根据权利要求 5所述的方法, 其特征在于所属灰度直方图的维数设 为 36。
9.根据权利要求 1所述的方法, 其特征在于所述分别类聚包括: 设定每类初始有 10个特征;
将视觉码本和弱几何限制分数编码进到拍索引文档。
10. 根据权利要求 1所述的方法, 其特征在于所述分割包括- 将商标图像分成 5- 20个子区域;
对每个子区域进行基于特征点的多视角建模。
11.根据权利要求 9所述的方法, 其特征在于所述弱几何限制包括- 对于数据库图像中的每一个轮廓, 将其内部的 "点特征"向 X和 Y坐标 方向分别做投影;
按照 X坐标从左到右的顺序, Y坐标从下到上的顺序把特征点标记为 1 对于每一个査询子区域中的每一个轮廓按上述方法可以得到其内部 "点特征"在 X和 Y坐标方向上的实际顺序;
采用 SIFT特征描述子和欧式距离度量得到相互匹配的特征点对。
12. 根据权利要求 11所述的方法, 其特征在于所述弱几何限制分数 Μ^, 按下式计
Figure imgf000010_0001
和 /^+1是某一个査询的子区域中相邻的两个 "点特征", 投影在
X方向上的坐标值小于; ; 0(Pj和 是与 和 匹配点的坐标。
PCT/CN2011/073985 2011-05-12 2011-05-12 商标检测与识别的方法 WO2012151755A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/CN2011/073985 WO2012151755A1 (zh) 2011-05-12 2011-05-12 商标检测与识别的方法

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2011/073985 WO2012151755A1 (zh) 2011-05-12 2011-05-12 商标检测与识别的方法

Publications (1)

Publication Number Publication Date
WO2012151755A1 true WO2012151755A1 (zh) 2012-11-15

Family

ID=47138661

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2011/073985 WO2012151755A1 (zh) 2011-05-12 2011-05-12 商标检测与识别的方法

Country Status (1)

Country Link
WO (1) WO2012151755A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109857897A (zh) * 2019-02-14 2019-06-07 厦门一品威客网络科技股份有限公司 一种商标图像检索方法、装置、计算机设备及存储介质

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20030108237A1 (en) * 2001-12-06 2003-06-12 Nec Usa, Inc. Method of image segmentation for object-based image retrieval
CN101714254A (zh) * 2009-11-16 2010-05-26 哈尔滨工业大学 联合多尺度sift和区域不变矩特征的配准控制点提取方法
CN101763429A (zh) * 2010-01-14 2010-06-30 中山大学 一种基于颜色和形状特征的图像检索方法
CN101763440A (zh) * 2010-03-26 2010-06-30 上海交通大学 对搜索图像的过滤方法
US20100195914A1 (en) * 2009-02-02 2010-08-05 Michael Isard Scalable near duplicate image search with geometric constraints
CN101866352A (zh) * 2010-05-28 2010-10-20 广东工业大学 一种基于图像内容分析的外观设计专利检索方法

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20030108237A1 (en) * 2001-12-06 2003-06-12 Nec Usa, Inc. Method of image segmentation for object-based image retrieval
US20100195914A1 (en) * 2009-02-02 2010-08-05 Michael Isard Scalable near duplicate image search with geometric constraints
CN101714254A (zh) * 2009-11-16 2010-05-26 哈尔滨工业大学 联合多尺度sift和区域不变矩特征的配准控制点提取方法
CN101763429A (zh) * 2010-01-14 2010-06-30 中山大学 一种基于颜色和形状特征的图像检索方法
CN101763440A (zh) * 2010-03-26 2010-06-30 上海交通大学 对搜索图像的过滤方法
CN101866352A (zh) * 2010-05-28 2010-10-20 广东工业大学 一种基于图像内容分析的外观设计专利检索方法

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109857897A (zh) * 2019-02-14 2019-06-07 厦门一品威客网络科技股份有限公司 一种商标图像检索方法、装置、计算机设备及存储介质
CN109857897B (zh) * 2019-02-14 2021-06-29 厦门一品威客网络科技股份有限公司 一种商标图像检索方法、装置、计算机设备及存储介质

Similar Documents

Publication Publication Date Title
Antonacopoulos et al. ICDAR2005 page segmentation competition
US9042659B2 (en) Method and system for fast and robust identification of specific product images
CN109614508B (zh) 一种基于深度学习的服装图像搜索方法
CN104376105B (zh) 一种社会媒体中图像低层视觉特征与文本描述信息的特征融合系统及方法
CN102176208B (zh) 基于三维空时特征的鲁棒视频指纹方法
CN102385592B (zh) 图像概念的检测方法和装置
Dong et al. An adult image detection algorithm based on Bag-of-Visual-Words and text information
CN104317946A (zh) 一种基于多张关键图的图像内容检索方法
CN105760875B (zh) 基于随机森林算法的判别二进制图像特征相似实现方法
CN103399863B (zh) 基于边缘方向差特征袋的图像检索方法
CN104965928B (zh) 一种基于形状匹配的汉字图像检索方法
CN108664968B (zh) 一种基于文本选取模型的无监督文本定位方法
CN113705310A (zh) 特征学习的方法、目标物体的识别方法和对应装置
Le et al. Improving logo spotting and matching for document categorization by a post-filter based on homography
CN107357834A (zh) 一种基于视觉显著性融合的图像检索方法
CN106066887A (zh) 一种广告序列图像快速检索和分析方法
CN107423294A (zh) 一种社群图像检索方法及系统
WO2012151755A1 (zh) 商标检测与识别的方法
Tian et al. Research on image classification based on a combination of text and visual features
Xu et al. Application of image content feature retrieval based on deep learning in sports public industry
Sun et al. A novel region-based approach to visual concept modeling using web images
Zhong et al. Fast and robust text detection in MOOCs videos
Thollard et al. Content-based re-ranking of text-based image search results
Ho et al. A scene text-based image retrieval system
Kalaiarasi et al. Visual content based clustering of near duplicate web search images

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: 11864985

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: 11864985

Country of ref document: EP

Kind code of ref document: A1