Authors:
Bing Liu
1
;
Bing Li
2
;
Weiming Hu
2
and
Jinfeng Yang
1
Affiliations:
1
Civil Aviation University of China, China
;
2
Chinese Academy of Sciences, China
Keyword(s):
logo detection, Few-example, Three-stage, Model refinement.
Related
Ontology
Subjects/Areas/Topics:
Artificial Intelligence
;
Computational Intelligence
;
Evolutionary Computing
;
Knowledge Discovery and Information Retrieval
;
Knowledge-Based Systems
;
Machine Learning
;
Soft Computing
;
Symbolic Systems
Abstract:
Logo detection is a laborious but strong practicality task that has a variety of technology applications. Since the fundamental of state-of-the-art detectors, large-scale annotated datasets, is cost-consuming, few-example logo detection is imperative and thought-provoking. In this paper, a three-stage Few-example Logo Detection Refined System (FLDRS) is proposed to detect logo with a few annotated samples. Specifically, the proposed detector is first initialized using large-scale generic target detection dataset with annotations, such as ImageNet, then further updated with large amount of synthetic logo images, and finally refined with a few annotated real examples. To make synthetic data more closer to real scene, a copy-paste-blend strategy is also presented in our model which not only characterizes many kinds of possible logo transformations but also takes the environment attribute of the logo type into consideration. The superior performance in FlickLogo-32 dataset demonstrates t
he efficiency of the proposed FLDRS.
(More)