Surface crack detection based on image stitching and transfer learning with pretrained convolutional neural network

During the operating lifecycle of civil structures, cracks will occur inevitably, which may pose great threat to the safety of the structures without timely maintenance. Digital image processing techniques have great potential in automatically detecting cracks, which can replace the labor-intensive and highly subjective traditional manual on-site inspections. Therefore, this paper presents a crack detection technology based on a convolutional neural network, GoogLeNet Inception V3. Firstly, a crack image dataset is acquired and constructed, which includes 2682 images with cracks and 983 images without crack at a resolution of 256 × 256 pixels. Then, based on a transfer learning method, the pretrained GoogLeNet Inception V3 model is retrained by the crack dataset for better identifying the crack images. The accuracy of the final trained model on the test set can reach 0.985. Moreover, image stitching based on Oriented FAST and Rotated BRIEF feature matching algorithm is realized, in order to overcome the limitation of camera field of view. Compared with the traditional image processing technology, the method adopted in this work can automatically study the characteristics of the object from the dataset, which can adapt to the complex real environment. Due to the transfer learning method, the crack detection can be achieved based on the existing well-trained models after being retrained by a small dataset.