Infrared and Visible Images Registration with Adaptable Local-Global Feature Integration for Rail Inspection

Abstract

Active thermography provides infrared images that contain sub-surface defect information, while visible images only reveal surface information. Mapping in- frared information to visible images o ers more comprehensive visualization for decision-making in rail inspection. However, the common information for regis- tration is limited due to di erent modalities in both local and global level. For example, rail track which has low temperature contrast reveals rich details in visible images, but turns blurry in the infrared counterparts. This paper pro- poses a registration algorithm called Edge-Guided Speeded-Up-Robust-Features (EG-SURF) to address this issue. Rather than sequentially integrating local and global information in matching stage which su ered from buckets e ect, this al- gorithm adaptively integrates local and global information into a descriptor to gather more common information before matching. This adaptability consists of two facets, an adaptable weighting factor between local and global information, and an adaptable main direction accuracy. The local information is extracted using SURF while the global information is represented by shape context from edges. Meanwhile, in shape context generation process, edges are weighted ac- cording to local scale and decomposed into bins using a vector decomposition manner to provide more accurate descriptor. The proposed algorithm is qual- itatively and quantitatively validated using eddy current pulsed thermography scene in the experiments. In comparison with other algorithms, better perfor- mance has been achieved.