Cyun–Bo WangJIAN-JIUN DINGShiang–Chih Hua2024-10-142024-10-142024-04-19https://scholars.lib.ntu.edu.tw/handle/123456789/721994We developed LightSegmentor, a real-time semantic segmentation model characterized by its lightweight design, utilizing just 3.3 million parameters. The compact framework comprises an encoder, termed LightFormer, as well as a decoder. The LightFormer incorporates a novel lightweight channel mixing operation, Unified Light Channel multilayer perceptron (MLP), to efficiently exchange image information across different channels and reduce memory usage and computational cost. For the decoder, we adopted the design of EffSegmentNet to effectively integrate the detailed and contextual information from the encoder to produce the semantic segmentation result. The LightSegmentor showed 66.9% mIoU utilizing only 3.3 million parameters on the test dataset of Cityscapes. Furthermore, it attained an impressive inference speed of 151 frames per second (FPS) on an NVIDIA RTX 2080Ti card, showcasing its suitability for real-time semantic segmentation.conference paper10.1109/iceib61477.2024.10602675