Road Scene Segmentation Algorithm Based on Strip Pooling and Densely Connected Pyramid

Semantic segmentation is vital in applications like autonomous driving, medical imaging, and scene understanding. Although DeepLabV3+ achieves strong segmentation results, it relies heavily on the Xception backbone, leading to high computational cost and poor performance on small details, limiting its use in real-time and resource-constrained scenarios. To address this, we replace Xception with the lightweight MobileNetV2 to reduce model complexity and improve adaptability. We introduce the SP-DenseASPP module to expand the receptive field and capture fine details, add a Pyramid Spatial Attention (PSA) module for adaptive multi-scale feature weighting, and apply the efficient, parameter-free SimAM attention to lowlevel features for better discrimination. Experiments on the Cityscapes dataset show our model achieves 74.58% mIoU, outperforming DeepLabV3+ by 1.18%, while reducing parameters by 31.84 million. Visualizations confirm improved segmentation, especially for small objects like poles and traffic signs.