Deep learning for road segmentation and classification

Tim Smole (2019) Deep learning for road segmentation and classification. MSc thesis.

Preview

PDF Download (39Mb)

Abstract

One of the problems road holders are facing is maintaining a record of road's surface quality. They acquire a vast amount of image data and then assess the surface quality by manually inspecting those images, which is time consuming and often inconsistent. In this work we show how to tackle a similar problem of automatic recognition of road surface type. To solve this problem we use the artificial neural network for classification tasks based on ResNet-50 architecture. To boost it's performance we use the information of the road's position in the input image which is obtained with U-Net neural network for semantic segmentation. In case of segmentation we show how to emphasise pixels located near road's edges and focus the network's attention during training to the parts where errors are most frequent. We also consider coarsely annotated images and show how we can use unlabelled pixels assigning them lower weights during the training process. We compare two attention mechanisms for neural networks used for classification tasks. The first mechanism masks input images with zero values where segmentation network detects background. The second mechanism is based on extending the input image with an output of U-Net. We show that by using the second approach F1 score evaluated on the test dataset improves from 0.947 to 0.971.

Actions (login required)