A real-time semantic segmentation method for small objects using attention mechanism

Semantic segmentation is an important problem in the field of computer vision, and its goal is to assign a semantic label to each pixel in an image. The effect of the traditional model on the segmentation of large objects such as vehicles and buildings is already very good, but the segmentation effect on small objects such as signal lights and traffic signs is not ideal. This is because the size of small objects is too small to capture their detailed information. General semantic segmentation models still have limitations in small object segmentation. Aiming at the difficulty of extracting small target detail information features, STDCNet (Fan et al., 2021) is improved, and a real-time semantic segmentation network STDCNet based on dual attention (DA-STDCNet) is proposed with a dual attention mechanism, which enhances the The model’s ability to extract spatial detail information and feature representation of global contextual semantic information finally achieves accurate capture and segmentation of various small targets in the image. Our model achieves 76.7% mIOU at an inference speed of 98.6 FPS on the Cityscape test set. Copyright © 2024 Inderscience Enterprises Ltd.