ULAF-Net: Ultra lightweight attention fusion network for real-time semantic segmentation

Real-time semantic segmentation, laying the foundation of mobile robots and autonomous driving, has attracted much attention in recent years. Currently, most deep models suffer high computational costs due to their complex architectures, making them impractical on resource-limited devices. Some lightweight models have been designed by reducing model complexity at the expense of segmentation accuracy. We propose an ultra-lightweight network, called ULAF-Net, to achieve a balance between segmentation accuracy, model complexity, and inference speed. This network abandons the straightforward concatenation of simple multi-scale fusion methods. First, we employ a parameter-free attention mechanism to process two large-scale feature maps, followed by the initial fusion. Furthermore, we consider the characteristics of varying scales and utilize lightweight spatial and channel attention modules to perform secondary processing on the fused large-scale feature map and the small-scale feature map, respectively, further highlighting important features. Finally, we combine both of them. In addition, we integrate multiple specialized convolutional methods and attention mechanisms to design a new residual module, which can make full use of the contextual features. The parameter quantity of ULAF-Net is merely 0.60M. It possesses the capability of real-time segmentation and achieves competitive segmentation outcomes on public datasets.