Javascript must be enabled for the correct page display

Logo Detection and Recognition in Images of Complex Scenes

Xuan, Win Leong (2018) Logo Detection and Recognition in Images of Complex Scenes. Master's Thesis / Essay, Computing Science.

[img] Text
Restricted to Registered users only

Download (52MB)


Logo recognition is becoming an important marketing indicator and analysis tool for companies to track their position in the current market. A company’s logo is a visual representation that identifies a company towards customers. A well-designed logo offers more than identification, it also provides a sense of a company’s character and values making them unique for each company. In recent years, the use of deep learning has increased for computer vision problems such as object and facial recognition. The use of such methods can also be applied to the logo recognition task. In this thesis, we propose a brand logo detection system using the Faster Region-based Convolutional Neural Network to localize and recognize logos in images of complex scenes. We apply transfer learning by extracting knowledge from pre-trained models to train a new model for the logo recognition task. A novel approach is also proposed where we train one convolutional neural network for each logo class and later aggregating these networks into a model. The FlickrLogos-32 dataset is used to train and test these models. We were able to achieve promising results in localizing and recognizing logos in complex scenes where logos are translated, rotated, skewed and so forth. A mean average precision of 0.77 was achieved by using pre-trained VGG19 for 32 different logo brands. Our novel approach achieved a less promising precision of 0.56 where we propose further improvements to the issues that were discovered.

Item Type: Thesis (Master's Thesis / Essay)
Supervisor name: Azzopardi, G. and Karastoyanova, D.
Degree programme: Computing Science
Thesis type: Master's Thesis / Essay
Language: English
Date Deposited: 27 Aug 2018
Last Modified: 19 Aug 2019 10:48

Actions (login required)

View Item View Item