The application of deep learning to the building facade crack detection has the advantage of high efficiency and superior objectivity.However,the diversity of building facade makes it time-consuming to simply use a massive sample set to train the model and difficult to reach a satisfying recognition.In order to improve the precision,recall and training efficiency,a fine-tuning training method based on transfer learning was proposed.After using a large sample set from other buildings for pre-training,fine-tuning training with a small amount of sample from the building under detection was conducted,and then the fine-tuned model was used to detect other parts of the building.Through the comparison of the test results and training time-consumption,the superiority of applying fine-tuning convolutional neural network to building facade crack detection was proved,the influence of frozen layers on fine-tuning model was studied.
关键词
深度学习/微调卷积神经网络/建筑外立面/裂缝检测/训练样本
Key words
deep learning/fine-tuned convolutional neural network/building facade/crack detection/training sample
维普
万方数据