LEDNet

LEDNet (Lightweight Encoder-Decoder Network) remains a pivotal architecture in the 2026 computer vision landscape, specifically optimized for real-time semantic segmentation on hardware-constrained edge devices. Originally introduced to address the computational overhead of heavy segmentation models like PSPNet or DeepLab, LEDNet utilizes an asymmetric encoder-decoder structure. The encoder employs a ResNet-based design with a focus on channel splitting and shuffling (via the SS-bt module) to maintain high representational power while drastically reducing parameter count. In 2026, it is frequently deployed in the 'Small-Model' tier of autonomous systems, providing high-frequency spatial awareness (mIoU of ~70.6% on Cityscapes) at extremely low latencies. Its technical architecture is characterized by the use of dilated convolutions and a Global Average Pooling (GAP) layer in the decoder to capture multi-scale context without the memory footprint of traditional pyramid pooling. For AI Solutions Architects, LEDNet is the go-to benchmark for industrial IoT and mobile applications where GPU memory is limited to 2GB or less and real-time performance (up to 70 FPS on modern mobile chipsets) is a non-negotiable requirement.