Research on Semantic Segmentation Algorithms for Street Scenes

Abstract

Some advanced semantic segmentation models often feature deep network structures and a large number of parameters, requiring significant memory and computational resources. This makes them difficult to run in real-time on devices with limited computational capacity. The model may fail to complete the semantic segmentation task within the required timeframe, while the complex feature extraction process further increases computational demands and processing time, reducing real-time performance. This, in turn, slows system response speed and negatively impacts practical applications. To address the trade-off between segmentation accuracy and real-time performance, this paper proposes an up-sampling module-the Scale-Aware Depth wise Separable Convolution Attention Module (SADAM). This module enhances the model’s ability to focus on key regions, improving feature extraction efficiency while reducing model complexity and boosting inference speed without compromising segmentation accuracy. Additionally, a semantic-assisted optimization branch is introduced to incorporate more feature information, enhancing the model’s representation capacity and adaptability. Furthermore, the loss function is optimized to improve segmentation accuracy and completeness.