Year: 2025
Type: conference-paper
Abstract: Urban landscape semantic segmentation based on deep learning provides effective technical support for urban construction, such as autonomous driving and traffic control. However, the complexity and variability of elements in urban landscapes pose challenges to the accuracy of semantic segmentation. This paper proposes an improved DeepLabV3 model: the feature extraction network uses ResNet50 with small strides, ensuring computational efficiency while enhancing the granularity of feature extraction; the dilation rates of the spatial pyramid pooling module are appropriately increased to enable the model to capture richer contextual information and compensate for potential missed details during feature extraction; a Convolutional Block Attention Module (CBAM) is embedded between the feature extraction network and the spatial pyramid pooling module, enhancing the model's ability to extract key features in both channel and spatial dimensions. Experimental results show that the proposed algorithm outperforms the existing DeepLabV3 algorithm, with a mean Intersection over Union (mIoU) improvement of 1.5% on the Cityscapes dataset and reaching 81.6% on the GTA5 urban street scene dataset.
Authors: Ying Chen, Bo Li, Zhi feng Wen
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Awards: G3110625756, G3214297619
OpenAlex ID: W7127039423
