Author(s):
1. HOANH NGUYEN:
Faculty of Electrical Engineering Technology, Industrial University of Ho Chi Minh City,Ho Chi Minh,
Abstract:
This paper presents a deep learning-based vehicle logo recognition framework based on vehicle regions and single shot framework with multi-scale feature fusion. To exactly locate front and rear regions of a vehicle, a vehicle region detection network based on region proposal network is designed. The vehicle region detection network uses ResNet-50 as the base network to generate base convolution feature maps. Atrous convolution is used after the base network to enlarge the receptive field to incorporate larger context information without increasing the number of parameters or the amount of computation. Each of detected vehicle region is fed into a vehicle logo recognition network based on single shot framework. The vehicle logo recognition network adopts Darknet-53 as the base network to generate proposals from each vehicle region. To enhance the performance of the proposed framework on small vehicle logo recognition, this paper designs a multi-scale feature fusion subnet which generates high-level semantic feature maps from base convolution feature maps. The tasks of detection predictions and object classification are performed on the multi-scale high-level semantic feature maps. Furthermore, a collected dataset based on a large public vehicle dataset is used to evaluate the proposed method. Experimental results on the collected dataset show the effectiveness of the proposed method on vehicle logo recognition. The proposed method provided a promising solution to vehicle logo recognition applications.
Page(s):
3327-3337
DOI:
DOI not available
Published:
Journal: Journal of Theoretical and Applied Information Technology, Volume: 98, Issue: 16, Year: 2020
Keywords:
deep learning
,
Single Shot Framework
,
Atrous Convolution
,
Vehicle Logo Recognition
,
MultiScale Feature Fusion
References:
References are not available for this document.
Citations
Citations are not available for this document.