1 Deep Learning for Logo Detection A Survey Sujuan Hou Member IEEE Jiacheng Li Weiqing Min Senior Member IEEE Qiang Hou

2025-04-28 0 0 1.68MB 13 页 10玖币

侵权投诉

Deep Learning for Logo Detection: A Survey

Sujuan Hou, Member, IEEE, Jiacheng Li, Weiqing Min, Senior Member, IEEE, Qiang Hou,

Yanna Zhao, Member, IEEE, Yuanjie Zheng, Member, IEEE, and Shuqiang Jiang, Senior Member, IEEE

Abstract—When logos are increasingly created, logo detection

has gradually become a research hotspot across many domains

and tasks. Recent advances in this area are dominated by

deep learning-based solutions, where many datasets, learning

strategies, network architectures, etc. have been employed. This

paper reviews the advance in applying deep learning techniques

to logo detection. Firstly, we discuss a comprehensive account of

public datasets designed to facilitate performance evaluation of

logo detection algorithms, which tend to be more diverse, more

challenging, and more reﬂective of real life. Next, we perform an

in-depth analysis of the existing logo detection strategies and the

strengths and weaknesses of each learning strategy. Subsequently,

we summarize the applications of logo detection in various

ﬁelds, from intelligent transportation and brand monitoring to

potential challenges and present the future directions for the

development of logo detection to complete this survey.

Index Terms—logo detection, computer vision, deep learning,

datasets

I. INTRODUCTION

Logos usually consist of texts, shapes, images, or their com-

bination. Logo detection beneﬁts a wide range of applications

in different areas, such as intelligent transportation [1, 2],

social media monitoring [3], and infringement detection [4, 5].

Meanwhile, some competitions have emerged, such as Robust

Logo Detection Grand Challenge [5–8] and Few-shot Logo

Detection [9].

The main task of logo detection is to determine the loca-

tion of a speciﬁc logo in images/videos and identify them.

Although it may be regarded as a particular object task, logo

detection in real-world images can be pretty challenging since

numerous brands may have highly diverse contexts, varied

scales, changes in illumination, size, resolution, and even non-

rigid deformation (as shown in Fig. 1).

Many previous works on logo detection employ hand-

crafted features (like SIFT [10]) to represent logos and use sta-

tistical classiﬁers for classiﬁcation. Such methods suffer from

complex image preprocessing pipelines and poor robustness

when dealing with a much larger number of logos. Recent

years have witnessed the rousing success of deep learning

since ImageNet Large Scale Visual Recognition Challenge

This work was supported by the National Nature Science Foundation of

China (No.62072289, 61972378, U193620), CAAI-Huawei MindSpore Open

Fund. S. Hou, J. Li, Q. Hou, Y. Zhao and Y. Zheng are School of Information

Science and Engineering, Shandong Normal University, Shandong, 250358,

China. Email: sujuanhou@sdnu.edu.cn, 2021317140@stu.sdnu.edu.cn,

2019309052@stu.sdnu.edu.cn, yannazhao@sdnu.edu.cn, and zhengyuan-

jie@gmail.com. W. Min and S. Jiang are with the Key Laboratory of

Intelligent Information Processing, Institute of Computing Technology,

Chinese Academy of Sciences, Beijing, 100190, China, and also with

University of Chinese Academy of Sciences, Beijing, 100049, China. Email:

minweiqing@ict.ac.cn, and sqjiang@ict.ac.cn.

(a) dim light (b) image rotation (c) small-scale logos

(d) multi-scale logos (e) non-rigid deformation (f) glare reﬂection

Fig. 1: Examples of images in adverse conditions.

(ILSVRC) [11]. Deep learning-based solutions with expres-

sive feature representation capability offer better robustness,

accuracy, and speed and thus attract increasing attention.

There are more than 100 papers about logo detection from

1993 to 2022, and a concise milestone of logo detection is

shown in Fig. 2. We can see that many deep learning-based

logo detection strategies have been proposed since 2015. This

survey mainly concentrates on deep learning-based solutions

specially developed for logo detection.

Even though deep learning has dominated the logo re-

search community, a comprehensive and in-depth survey on

deep learning-based solutions is lacking. In this survey, we

mainly focus on the advances in recent deep learning for

logo detection. We provide in-depth analysis and discussion on

existing studies in various aspects, covering datasets, pipelined

used, task types, detection strategies, loss functions, their

contributions and limitations. We also try to analyze potential

research challenges and future research directions for logo

detection. We hope our work could provide a novel perspective

to promote the understanding of deep learning-based logo

detection, foster research on open challenges, and speed up

the development of the logo research ﬁeld.

The rest of this paper is organized as follows. In Section II,

we investigate the public logo detection datasets. In Section III,

we review and organize the currently available work on logo

detection. In Section IV, we introduce the applications of logo

detection in real-world scenarios. In Section V, we discuss its

challenges and prominent future research directions. Finally,

we summarize the whole text in Section VI.

II. LOGO DATASETS

Deep learning has brought great success to object detection

in recent years, where datasets play a crucial role. Datasets

arXiv:2210.04399v1 [cs.CV] 10 Oct 2022

Fig. 2: A concise milestone for logo detection.

(a) BelgaLogos (b) FoodLogoDet-1500 (c) QMUL-OpenLogo (d) LogoDet-3K

Fig. 3: Logo images sampled from different datasets.

are not only a common basis for comparing and measuring

the performance of algorithms but also an essential factor in

supporting advanced object detection algorithms. This section

provides an overview of the logo datasets for detection. In

addition, we also give a brief summary of existing logo

classiﬁcation datasets.

In recent years, many logo datasets intended for detection

have been created to solve the problem of large realistic

datasets with accurate ground truth. The use of these datasets

enables qualitative as well as quantitative comparisons and

allows benchmarking of different algorithms. We conducted

statistics on existing datasets commonly used for logo detec-

tion and classiﬁed them into three types based on their scale:

small-scale, medium-scale, and large-scale. Table I provides

statistics on the available logo datasets, and Fig. 3 gives some

illustrative examples from these datasets.

The small-scale datasets include BelgaLogos [12],

FlickrLogos-32 [13], etc. As the ﬁrst benchmark dataset

proposed for logo detection, BelgaLogos [12] consists of

10,000 images of natural scenes, with 37 different logos

and 2,695 instances of logos labeled with bounding boxes.

FlickrLogos-32 [13], one of the most popular small-scale

datasets for logo detection, comprises 32 different classes

with 70 images in each class. The images in this dataset

are mainly captured from the real world, and many contain

occlusions, appearance changes, and lighting changes, making

detecting this dataset very challenging.

Datasets with medium-scale include Logo-Net [14], QMUL-

OpenLogo [15], FoodLogoDet-1500 [16], etc. Logo-Net [14]

is built for detecting logos and identifying brands from real-

world product images. It consists of two sub-datasets, namely

Logo-18 and Logo-160. QMUL-OpenLogo [15] is an open

benchmark for logo detection, constructed by aggregating

seven existing datasets and building an open protocol for

logo detection evaluation. The QMUL-OpenLogo dataset has

a highly imbalanced distribution and signiﬁcant variation in

scale, which is critical to verify the performance of the

detection algorithm. FoodLogoDet-1500 [16] is the ﬁrst high-

quality public dataset of food logos with uneven distribution

among different food logo classes, which poses a challenge to

the small sample food logo detection algorithms. The dataset

is composed of 1,500 food logo classes with 99,768 images.

There are also some large-scale datasets, such as LogoDet-

3K [17] and PL8K [18]. LogoDet-3K [17] divides all logos

into nine super-classes based on the needs of daily life and

the main positioning of common enterprises, namely Cloth-

ing, Food, Transportation, Electronics, Necessities, Leisure,

Medicine, Sports, and Others. The dataset consists of 3,000

logo classes, 158,652 images, and 194,261 logo objects. The

number of logo classes contained in different super-classes

TABLE I: Statistics of existing logo detection datasets.

#Scale #Datasets #Logos #Brands #Images #Objects #Public #Year

Small

BelgaLogos [12] 37 37 10,000 2,695 Yes 2009

FlickrLogos-27 [19] 27 27 1,080 4,671 Yes 2011

FlickrLogos-32 [13] 32 32 2,240 5,644 Yes 2011

MICC-Logos [10] 13 720 - No 2013

Logo-18 [14] 18 10 8,460 16,043 No 2015

Logos-32plus [20] 32 32 7,830 12,302 No 2017

Top-Logo-10 [21] 10 10 700 - No 2017

Video SportsLogo [22] 20 20 2,000 - No 2017

VLD 1.0 [23] 66 66 25,189 - No 2019

SportLogo [24] 31 31 2,836 - Yes 2020

VLD-45 [25] 45 45 45,000 - No 2020

Medium

Logo-160 [14] 160 100 73,414 130,608 No 2015

Logos-in-the-Wild [26] 871 871 11,054 32,850 Yes 2017

QMUL-OpenLogo [15] 352 352 27,083 - Yes 2018

PL2K [27] 2,000 2,000 295,814 - No 2019

FoodLogoDet-1500 [16] 1,500 - 99,768 145,400 Yes 2021

Large

Open Brands [28] 1,216 559 1,437,812 3,113,828 No 2020

LogoDet-3K [17] 3,000 2,864 158,652 194,261 Yes 2020

PL8K [18] 7,888 7,888 3,017,146 - No 2022

varies greatly. For example, the Food, Clothes, and Necessities

contain more images and objects than other super-classes.

The imbalanced distribution across different logo classes of

LogoDet-3K poses a challenge to effectively detecting logos

with few samples. PL8K [18] is a large logo detection dataset

constructed semi-automatically. The dataset consists of 7,888

logo brands and 3,017,146 images, and at least 20 images per

class.

In addition, there are also some datasets built for logo

classiﬁcation, e.g. WebLogo-2M [29] and Logo 2K+ [30].

Weblogo-2M [29] is obtained by automatic web data ac-

quisition and processing. The dataset excludes images with

small widths and/or heights and duplicate images. Compared

with other datasets, the Weblogo-2M presents three unique

properties inherent to large-scale data exploration for learning

scalable logo models: (1) Weak Annotation. (2) Noisy (False

Positives). (3) Class Imbalance. Logo-2K+ [30] is a large-scale

high-quality logo dataset. It covers a variety of logo classes

from the real world, and different types of logo images have

various logo appearances, scales, and backgrounds since they

are collected from different websites. The dataset has high

spatial coverage of categories including Food, Clothes, Insti-

tutions, Accessories, Transportation, Electronics, Necessities,

Cosmetics, Leisure, and Medical. The number of images is

imbalanced among different categories. For instance, Food has

769 logo classes, while Medical has only 48.

As with common object detection, mAP [31] is the most

commonly used evaluation metric to measure logo detectors.

III. LOGO DETECTION

In logo detection, logo classiﬁcation is also an essential part.

Therefore, we brieﬂy summarize logo classiﬁcation before

introducing logo detection.

A. Logo Classiﬁcation

As one of the most critical tasks in computer vision, logo

classiﬁcation aims to recognize the logo name corresponding

to the input image. According to different feature extraction

strategies, existing classiﬁcation methods are divided into two

categories: traditional machine learning-based methods and

deep learning-based methods. Some representative methods

will be described brieﬂy in the following.

1) Traditional Machine Learning-based Methods: The tra-

ditional logo classiﬁcation methods extract features through

manual features, such as SIFT and HOG, and then classify

them by a classiﬁer. Support Vector Machine (SVM) [32]

is a classiﬁer that performs binary data classiﬁcation in a

supervised learning manner. In traditional logo classiﬁcation,

SVM also shows excellent performance [33–36]. Carvalho

et al. [35] proposed a self-learning and automatic detection

method that performs detection without any prior data. The

scheme automatically identiﬁes the candidate regions of the

localization logo, uses the HOG features extracted from the

localization to train the object detector and some sub-detectors

and uses the SVM to recognize the logo image.

K-Nearest Neighbor (KNN) is a supervised learning algo-

rithm commonly used in classiﬁcation [37, 38]. Gopinathan et

al. [38] proposed a vehicle logo recognition system in 2018.

They used Euclidean distance on the initial training dataset

to summarize the pixel intensity distribution by applying each

channel’s mean and standard deviation and used the K-means

algorithm to cluster the color histogram features to group

different logos. Then HOG and KNN algorithms extract logo

features and classify logos.

2) Deep Learning-based Methods: In recent years, with

the continuous development of deep learning, deep learning

based solutions have also been successfully applied to logo

classiﬁcation [39–46]. Karimi et al. [43] used the DCNN

logo recognition algorithm, which conducted a pre-trained

model for feature extraction and then used SVM for logo

classiﬁcation. They also used transfer learning to improve

existing pre-trained models for logo recognition. Finally, the

ﬁne-tuned deep model is applied to the parallel structure to

obtain a more efﬁcient deep model for logo recognition.

The latest trend in logo classiﬁcation is to design efﬁcient

networks with limited resources [18, 30, 47, 48]. Wang et

al. [30] proposed a Discriminative Region Navigation and

Augmentation Network (DRNA-Net), which is capable of

discovering more informative regions and expanding these

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

1DeepLearningforLogoDetection:ASurveySujuanHou,Member,IEEE,JiachengLi,WeiqingMin,SeniorMember,IEEE,QiangHou,YannaZhao,Member,IEEE,YuanjieZheng,Member,IEEE,andShuqiangJiang,SeniorMember,IEEEAbstractWhenlogosareincreasinglycreated,logodetectionhasgraduallybecomearesearchhotspotacrossmanydomainsandtas...

收起<<

1 Deep Learning for Logo Detection A Survey Sujuan Hou Member IEEE Jiacheng Li Weiqing Min Senior Member IEEE Qiang Hou.pdf

共13页,预览3页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

1 Deep Learning for Logo Detection A Survey Sujuan Hou Member IEEE Jiacheng Li Weiqing Min Senior Member IEEE Qiang Hou

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: