Tongxin xuebao (Feb 2025)
Research on lightweight real-time image segmentation methods based on deep learning
Abstract
In response to the computational and storage burdens caused by the increasing model complexity in deep learning applications, especially in image segmentation tasks where algorithmic complexity, insufficient real-time responsiveness, and high memory usage were prevalent, a lightweight and efficient segmentation network architecture——multi-scale superposition fusion network (MSFNet) was proposed. MSFNet featured a dual-branch multi-scale boundary fusion module, which effectively enhanced segmentation accuracy by integrating feature information and boundary details from different scales. At the same time, it significantly reduced the model parameter count. Experimental results show that MSFNet outperforms other models on three public datasets, with a model size of only 0.6×106 parameters. On the RTX 3070 GPU, it processes 800×800 pixels images in just 12 ms, significantly improving the execution efficiency and resource utilization of segmentation tasks. Therefore, this model is particularly well-suited for deployment on resource-constrained edge or mobile devices, providing a favorable technical foundation for real-time image segmentation applications.
