Vision Image Monitoring on Transportation Infrastructures: A Lightweight Transfer Learning Approach

Abstract

Vision monitoring of distress has emerged as a new trend in intelligent transportation infrastructure systems, including roads and bridges. Recently, transfer learning methods and lightweight networks have been used to realize efficient distress identification without large amount of human-labor work. This paper proposed an engineering approach that integrated transfer learning with lightweight models to classify and detect concrete bridge distresses. Two datasets named Distress Dataset of Asphalt Pavement (DDAP) with 2500 asphalt pavement distresses images and Distress Dataset of Concrete Bridge (DDCB) with 906 concrete bridge distresses images were used. The lightweight models MobileNet and MobileNet-SSD were employed to conduct 6 comparative experiments for exploring the model performance with different transfer learning modes (Mode I and II) and procedures (one-step and two-step procedure) in classification and detection tasks. Based on the comparison of the results, the optimum model for two tasks was respectively recognized. For classification task, the model directly transferred specific parameters of partial convolution layers from ImageNet-based model achieved the highest accuracy of 97.8%. For detection task, the two-step transfer learning model using an intermediate transfer learning step trained by DDAP reached a mean average precision of 87.16%. The proposed approach has the application potential for practical road inspection work in intelligent transportation infrastructure maintenance.