RASEEN: Road Segmentation Enhancement Network with Dual Attention Mechanism and Deep Supervision

초록

This paper presents RASEEN (Road Segmentation Enhancement Network), a novel deep learning model for semantic road segmentation built upon the U-Net encoder-decoder framework. To address the persistent challenges of blurry boundaries, shadow interference, and low-contrast road surfaces, RASEEN integrates three key components: a dual channel-spatial attention module, an adaptive residual convolutional block with skip connections and SiLU activation, and a multi-layer deep supervision mechanism applied at each decoder stage. Training is guided by a composite loss function combining Binary Cross-Entropy, Dice, Focal, and Lov & aacute;sz losses, complemented by 5-fold cross-validation and test-time augmentation. Evaluated on the Indian Driving Dataset (IDD) comprising 6,993 image-mask pairs, RASEEN achieves a Dice coefficient of 0.941 and an Intersection over Union (IoU) of 0.883, surpassing DeepLabV3+ (Dice: 0.889, IoU: 0.800) and DANet (Dice: 0.912, IoU: 0.839) by 8.3 and 4.4 percentage points in IoU, respectively. Remarkably, these accuracy gains are achieved with an inference latency of only 38 ms per image, which is faster than both DANet (56 ms) and HRNet (52 ms). Ablation studies confirm that the attention module alone contributes a 2.6% IoU improvement, while the composite loss function contributes a further 4.3% gain. These results demonstrate that RASEEN offers a strong balance between accuracy and computational efficiency, making it well-suited for real-time applications such as autonomous driving and intelligent traffic monitoring.

키워드