Lite-FARNet: A Light-weight Feedback Attention Residual Network for Efficient Multi-Class Segmentation in Complex Urban Scenes

Image segmentation is one of the most important techniques in computer vision, forming the foundation for biomedical imaging and serving as a key component in urban planning and autonomous driving. In this project, we present a novel lightweight feedback attention residual network (Lite-FARNet) specifically designed to address the complex demands of multi-class segmentation in urban scenes, evaluated using the Cityscapes dataset. To effectively manage multi-class predictions, we introduce a competitive inference layer, while spatial and channel squeeze-and-excitation residual module are incorporated to enhance feature representation and improve context awareness. The streamlined architecture design significantly reduces computational complexity and memory usage, making it advantageous for deployment in resource-constrained environments. Extensive testing and evaluation demonstrate that the proposed network consistently delivers robust adaptability and strong accuracy across diverse and challenging urban environments. These findings highlight the model’s versatility, efficiency, and potential as a powerful tool for a wide range of image segmentation tasks, as well as its suitability for deployment in resource-constrained devices and real-time intelligent systems.