Open AccessSNIPER Based Multi-Target and Multi-Scale Aerial Image Processing Method
Author(s) -
Yaocheng Li,
Weidong Zhang,
Yingming Cai,
Zhe Li,
Xiuchen Jiang
Publication year - 2020
Publication title -
journal of physics. conference series
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.21
H-Index - 85
eISSN - 1742-6596
pISSN - 1742-6588
DOI - 10.1088/1742-6596/1659/1/012003
Subject(s) - computer science , object detection , computer vision , artificial intelligence , precision and recall , scheme (mathematics) , scale (ratio) , real time computing , pattern recognition (psychology) , mathematical analysis , physics , mathematics , quantum mechanics
The maintenance of power transmission system appeals automatic inspection system to replace manual inspection. Inspection based on Unmanned Aerial Vehicle (UAV) generates a huge number of photos to be processed. These aerial photos have several challenging features: high density, small, scale variation. In order to ameliorate the performance on detection of important objects as insulators and pin bolts, an efficient multi-scale training algorithm called SNIPER is introduced. SNIPER enhanced Faster RCNN was trained on a powerline dataset for object detection. SNIPER provides a good average precision and average recall on large objects like insulators and medium objects with adequate annotated instances like pin bolts and dampers. However, SNIPER fails to locate missing pin or displaced pin, possibly due to their similarity to pin bolts. Future development of a SNIPER-based cascaded detection scheme could help detect defected small objects.