| Abstract [eng] |
Road recognition in aerial images is an important area of research, because having access to up-to-date road network data is essential in GIS applications and GPS navigation. Since manual road mapping is an expensive and tedious process, automated solutions are actively being studied and developed. Since road recognition is a specialized image classification task, in this paper we use a state of the art image classification tool - convolutional neural networks. In addition, we propose custom post-processing algorithms to futher improve the achieved results. We conduct 20 wide-scale experiments related to convolutional neural networks. Each experiment attempts to adjust a single parameter in the network. The goal of these experiments is to design a convolutional neural network, whose architecture is best-suited for the task of recognizing roads in aerial images. After such architecture is achieved, the output of the network undergoes additional post-processing, such as cleaning of false positives and center pixel detection. Most of the post-processing algorithms used are custom-designed in this paper. We also make use of some well-known graph theory algorithms for additional post-processing. The final results of the recognition are then put through an evaluation algorithm, which is specifically tailored to measure the success of road reacognition tasks. We test our method with two datasets - one containing images of urban roads, the other - of rural roads. We achieve a high accuracy result when testing with the rural data set. We achieve a less stellar result when testing with the urban data set - the most prevalent problem being the inability to properly detect closed loops in the road. The reason for this, is that urban images contain a lot more noise and occlusions, such as buildings, bgridges and cars. All of these factors impact the performance of our algorithms in urban environments. Nevertheless, based on our results, we conclude that convolutional neural networks are indeed a suitable tool for road recognition tasks, especially in rural areas. We also conclude, that the problems in urban areas can be alleviated if another algorithm could detect loops in the road network and close them. To that end, we recommend looking into fully-convolutional networks. |
