abstract
- Designing convolutional neural network (CNN) architectures for semantic segmentation is often expert-dependent. This work presents a neural architecture search (NAS) prototype that democratizes architecture design through evolutionary computation. The framework employs a dual-representation search space and a surrogate performance estimator to identify high-performing U-Net variants without the prohibitive cost of full training. A graphical user interface improves search transparency and user interaction. Testing on the Carvana and a custom road dataset shows that our automated approach consistently produces architectures that outperform the U-Net baseline, balancing IoU and model size. This prototype offers an accessible and transparent solution for researchers to generate efficient, task-specific segmentation models autonomously.