A Cascaded Spatial Transformer Network for Oriented Equipment Detection in Thermal Images

Equipment detection is a fundamental step towards automatic diagnosis in electric power industry. In this paper, we propose a novel approach to detect equipment parts, which are often oriented in thermal images due to the use of hand-held thermographic cameras. The proposed approach integrates a cascaded spatial transformer network (STN) and a proposal-driven detection network together. The cascaded STN introduces a rotation transform globally so that the rigid body constraint between parts is implicitly considered. Moreover, the cascaded scheme is also robust to large variation in orientation. The designed STN and the detection network are first trained separately to get good initialization weights. Further, these two networks are fine-tuned in an end-to-end manner. Experiments demonstrate that both the cascaded scheme and the end-to-end fine-tuning highly boost the detection performance.