Bayesian Variational Neural Architecture Search for Lightweight Weld Seam Feature Point Detection in Laser Vision Systems.

Abstract

In laser vision-based robotic welding, the seam feature point must be located, under strong arc-light and spatter interference, in every frame acquired while the weld is being made, so that the torch can be steered in real time; the detector must therefore keep pace with the camera on the modest processor of an industrial controller. Conventional deep learning models encounter structural redundancy and high computational latency under this constraint. To address these limitations, this paper proposes a lightweight object detection model, whose box-center output is mapped to the weld seam feature point, optimized via Bayesian variational neural architecture search (BayeNAS), in which architecture parameters are modeled as Gaussian distributions and updated through natural-gradient variational inference. A novel dual-selection optimization strategy is introduced to jointly search for optimal candidate operators and network topology branches, achieving an explicit balance between model complexity and hardware latency. Furthermore, an in-search channel pruning strategy is incorporated to compress the search space and prevent the combinatorial growth of the channel configuration space. Comparative analysis and real-world welding experiments demonstrate that the proposed model achieves a high processing speed of 78 frames per second (FPS) on a legacy central processing unit (CPU)-above the 60 fps acquisition rate of the sensor-while maintaining a mean positioning error within 0.15 mm on the measurement plane. These results demonstrate that the proposed model achieves detection accuracy comparable to heavyweight detectors while offering an order-of-magnitude advantage in inference speed and model size, validating its suitability for low-cost industrial deployment.