Fender segmentation in unmanned aerial vehicle images based on densely connected receptive field block

Abstract

Periodic evaluation of port structures including outer walls and berthing structures with fenders is paramount to ensure their safety. However, most fenders are inaccessible on land and inspectors utilize floating boats for inspections. Therefore, this study aimed to develop a fender segmentation system incorporating a vision sensor with deep learning segmentation approach. The semantic segmentation model in an encoder–decoder framework was densely connected through a receptive field block convolution module each with two different dilation rates and classified various types and sizes of fenders efficiently. The images of various types of fenders were assembled and augmented to train the proposed network, and its performance was compared to conventional segmentation models. The trained network was finally applied to unmanned aerial vehicle images and the results showed that fenders were successfully segmented even in the case of the images that include changes in shapes or colors.

Keywords

• Unmanned aerial vehicle
• Vision sensor
• Deep learning
• Fender segmentation
• Port structure