Paper Info
Title
Detecting Small Objects in Urban Settings Using SlimNet Model
Abstract
The automatic extraction of small objects such as roadside milestones, small traffic signs, and other urban furniture remains a technical challenge. This study focuses on methods of deep learning to detect small urban elements in mobile mapping system (MMS) images. Based on images obtained by an MMS in urban areas, we create an urban element detection (UED) data set containing several kinds of small objects found in a city. A simple feature extraction convolution neural network (CNN) called SlimNet is proposed and combined with an optimized faster R-CNN framework. The resulting deep learning method can automatically extract small objects commonly found in cities, including manhole covers, milestones, and license plates. Experiments on the UED data set show that SlimNet has the highest accuracy compared with other popular networks, including VGG, MobileNet, ResNet, and YOLOv3. The SlimNet model can achieve a mean average precision (AP) that is up to 12.3% higher than that of the lowest ResNet-152 network and can accelerate both training and detection owing to its relative simplicity. Moreover, <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$k$ </tex-math></inline-formula> -means clustering is used to choose the dimensions of the anchor box for detection. We ran <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$k$ </tex-math></inline-formula> -means clustering for different numbers of clusters, and the results show that at least four clusters are needed for detection using a small data set such as the UED. We also propose a method to use templates of different scales for anchors to further improve small object detection; this approach improved the AP by 3%–4% in our experiments.
Year
DOI
Venue
2019
10.1109/TGRS.2019.2921111
IEEE Transactions on Geoscience and Remote Sensing
Keywords
Field
DocType
Feature extraction,Licenses,Urban areas,Deep learning,Three-dimensional displays,Object detection,Roads
Cluster (physics),Object detection,Computer vision,Small data,Pattern recognition,Convolutional neural network,Feature extraction,Artificial intelligence,Deep learning,Cluster analysis,Mobile mapping,Mathematics
Journal
Volume
Issue
ISSN
57
11
0196-2892
Citations 
PageRank 
References 
1
0.35
0
Authors
6
Name
Order
Citations
PageRank
Zheng Yang110.35
Yaolin Liu29725.42
Lirong Liu311.02
Xinming Tang410.35
Junfeng Xie5268.40
Xiaoming Gao6135.08