A Lightweight Road Segmentation Method Based on Improved U-Net Model

In recent years, the rapid advancement of autonomous driving technology has brought significant attention and research to road detection, a crucial component of this field. Road detection technology not only greatly enhances vehicle navigation capabilities and driving safety but also plays an essential role in intelligent transportation systems and advanced driver assistance systems. To improve the accuracy of road detection and enhance the deployability of models, this paper proposes enhancements to the U-Net model, resulting in a lightweight and precise network model. By halving the number of channels in the U-Net model and replacing standard convolutions with depthwise separable convolutions, the model's parameters and computational load are reduced, enabling efficient operation on edge devices with limited computational resources. Additionally, the model's ability to focus on and interpret key information, thereby enhancing its performance and accuracy in image segmentation tasks. Experimental results demonstrate that these modifications reduce the model's parameter size from 15.52M to 0.76M, a reduction of 95%, while also achieving improvements in performance metrics, with the Intersection over Union (IoU) increasing from 0.9417 to 0.9435 and the Dice coefficient from 0.9695 to 0.9709. The proposed method significantly reduces the parameter size while slightly enhancing segmentation performance, providing a valuable technical reference for lightweight road segmentation.