On 7 September 2026, GRAIL-related research on the deployment of deep learning models for industrial machine vision was presented at IEEE COINS 2026, the International Conference on Omni-layer Intelligent Systems, held in Bologna, Italy from 7–9 September.
Miguel Gómez Fernández, researcher at AIMEN presented the paper “Benchmarking Edge Inference Strategies for Deep Learning Models in Industrial Machine Vision” during Session AI Architectures, Edge Systems & Connectivity. The work was co-authored with David Castro Boga, Roi Méndez-Rial and Eric López-López from AIMEN.
The study addresses a practical challenge for industrial AI: how to select the most suitable framework for running deep learning models directly on edge devices, where low latency, data control and limited connectivity can be critical requirements.
The researchers compared four widely used inference approaches across several CPU- and GPU-based hardware platforms and different computer vision models, including YOLOv8, UNet and the transformer-based Grounding DINO.
Results showed that there is no single optimal framework for every deployment scenario. OpenVINO achieved the lowest inference times on the tested Intel CPUs, while TensorRT performed best for CNN-based models on NVIDIA GPUs. For the transformer-based Grounding DINO model, however, PyTorch achieved the lowest inference times on the desktop GPUs evaluated.
The research contributes to understanding how hardware and model architecture affect the efficient deployment of AI in industrial environments. The conference brought together researchers and industry experts working across AI, intelligent systems, edge technologies and autonomous applications.
