RFIENet: RGB-thermal feature interactive enhancement network for semantic segmentation of insulator in backlight scenes

Segmentation of insulators from complex images is a critical step for automatic fault inspection. However, it remains challenging in backlight scene parsing due to the visible imaging limitations. Thermal images are robust to harsh lighting conditions. In this paper, an RGB-thermal feature interactive enhancement network (RFIENet) is proposed to achieve backlight scene parsing of the insulator. First, two parallel single-modal feature encoders are constructed to extract the features of RGB or thermal images individually. Subsequently, the RGB and thermal modal features are integrated by two fusion modules separately, i.e. the multi-modal feature fusion module (MFFM) and the global feature fusion module (GFFM), which are used for shallow and deep features, respectively. Finally, the interactive enhancement decoder is developed to restore the resolution of the feature maps. The experimental results demonstrate the superiority of the proposed RFIENet.