A Two-Stream Context-Aware ConvNet for Pavement Distress Detection

Convolutional neural networks (ConvNets) are widely used for pavement distress analysis tasks in which features are typically extracted from a smaller image (e.g. 224 × 224) that has been cropped from an orthophoto. This paper introduces a ConvNet-based method to classify partitioned segments of orthophotos for pavement distress by incorporating two image input streams, one of which provides contextual information of the segment's surroundings. The resulting two-stream ConvNet is evaluated to classify segments of the orthophoto with a better performance for pavement distress than single-stream ConvNets, in terms of precision and recall.