Taguchi-optimized DeepLabV3+ for semantic segmentation in autonomous driving applications

Semantic segmentation is a critical perception task in autonomous vehicles, enabling pixel-wise classification of road scenes. In this study, we propose a systematic optimization of DeepLabV3+ semantic segmentation model using Taguchi Design of Experiments (DoE) technique to enhance its performance for real-time deployment in autonomous driving. We explore the influence of key hyperparameters. solver type (Adam, RMSProp, SGDM), learning rate (10⁻⁵, 10⁻⁴, 10⁻³) batch size (1, 2, 3), and L2 regularization (10⁻⁵, 10⁻⁴, 10⁻³), across three backbone networks: ResNet-18, ResNet-50, and MobileNetV2. Experiments were conducted on the Cambridge-driving Labeled Video Database (CamVid), a widely used benchmark for road scene understanding. The DoE approach efficiently reduced the number of training configurations while maximizing segmentation performance. The best-performing model, DeepLabV3+ with a ResNet-50 backbone, achieved a Mean Intersection over Union (IoU) of 76.23%, surpassing recent approaches. The proposed framework offers a practical strategy for deploying semantic segmentation models in autonomous vehicle systems.