Deep Learning Road Extraction via Multi-Task Autoencoder

The patent divides the road extraction task into multiple independent sub-tasks (road surface extraction, road center line extraction, road edge extraction, road angle estimation, and object detection). Each sub-task is handled by a dedicated decoder within the autoencoder framework, allowing specialized optimization for each component while maintaining overall system coherence. This segmentation enables higher accuracy without excessive complexity by focusing computational resources where needed.

The patent implements a multi-functional deep learning model that performs multiple tasks simultaneously using a unified architecture. The autoencoder framework with shared encoding layers and multiple decoders enables the system to extract roads, detect objects, and analyze various road characteristics all at once. This multi-functionality improves extraction accuracy by leveraging complementary information from different task outputs while avoiding the need for separate models.

The patent performs preliminary actions by pre-training the deep learning model on diverse satellite image datasets before actual road extraction tasks. The model learns general features and patterns during pre-training, which then transfer to specific extraction tasks. This preliminary learning enables the system to achieve high accuracy on map generation tasks without requiring extensive real-time computation, as the heavy lifting is done during the pre-training phase.

The patent maintains continuous useful action through the autoencoder framework that processes images through multiple decoding branches simultaneously. The encoding-decoding process continues seamlessly through different task outputs, with each decoder contributing to the overall extraction accuracy. This continuous processing approach ensures that all road-related information is extracted and refined in a unified manner, improving accuracy without significant time penalties.