Crack identification inside on-site steel box girder based on fusion convolutional neural network

摘要

In this paper we propose a novel fusion convolutional neural network to identify the local fatigue cracks in steel box girder of cable-stayed bridge. Unlike conventional CNN's chain-like structure, the proposed network fully exploits multiscale and multilevel information of input images by combining all the meaningful convolutional features together. Raw images with high resolution of 3624x4928 are decomposed into three kinds of sub-image sets with lower resolution of 64x64, background, handwriting and crack, respectively. Multi-functional layers are stacked including convolution. ReLU. softmaxResults show that the test error drops to 4% after only 50 epochs and it is more effective compared with other deep learning networks when handling large image datasets.