ABSTRACT
This research aims to provide a cropping-and-stitching segmentation-based method to find corrosions on steel bridge from a great number of images. In this study, a data set is generated by manually labeling damages pixel by pixel, on the images of steel with different levels of corrosion. The data set is used to train a Fully Convolutional Network (FCN) model for detecting location of corrosion on images. Then a new data set with cropped images is built by cropping the images from previous data set and then used to train a new FCN model. By the new model, the location of damage can be shown in a stitched image to visualize the damage and its distribution. Two FCN models are compared with each other to find the one with better performance. Besides, one 4k image taken by UAV is tested by the FCN model with better performance.
