ACKNOWLEDGMENT: REBECCA project is supported by the Chips Joint Undertaking and its members, including the top-up funding by National Authorities under grant agreement n° 101097224. Funded by the European Union. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the granting authority. Neither the European Union nor the granting authority can be held responsible for them.
How do you bring AI closer to the edge without compromising latency, throughput, and model fidelity?
Prepared by Intecs
Within the REBECCA KDT JU, this challenge is addressed by designing a secure and high-performance Edge AI stack, leveraging RISC-V architectures and reconfigurable computing paradigms to enable efficient on-device intelligence. At Intecs AI Lab, our contribution as Use Case Provider focuses on a concrete optimization pipeline for industrial applications: adapting a Deep Learning object detection model for resource-constrained embedded environments, such as UAV-based detection of photovoltaic panel defects.
The Real-time Fault Detection in PV Panels on UAVs use case, led by Intecs within the REBECCA project, aims to automate the inspection of photovoltaic plants using aerial images captured by infrared (IR) and visible (VIS) cameras. The core objective is to process data directly on-board the UAV, enabling immediate detection of defects and reducing the need for post-flight analysis, which in turn lowers O&M costs and accelerates maintenance actions.
Over the last months, Intecs has continued the development of the real-time defect detection pipeline with a focus on improving efficiency and robustness for UAV deployment. A key milestone has been the adoption of YOLOX-Tiny, a modern lightweight deep learning architecture specifically designed for real-time applications and edge deployment. Compared to the previously used YOLO-S model, YOLOX-Tiny achieves higher accuracy with a lower parameter count and significantly reduced computational complexity, making it more suitable for continuous on-board inference during flight.
Starting from a YOLO-Tiny baseline, we applied architectural and inference-level optimizations to better align with edge constraints (limited memory bandwidth, reduced compute availability), while preserving robust detection capabilities. A key step was the application of Post-Training quantization (PTQ) using ONNX Runtime, enabling reduced numerical precision without retraining. This significantly lowered memory footprint and computational complexity, with minimal impact on model accuracy.
Key outcomes:
• Model size reduced from ~30 MB (FP32) to ~5 MB (INT8)
• Substantial reduction in memory access overhead
• Improved inference efficiency for real-time, on-board execution
• Accuracy degradation kept within acceptable bounds Post-Training QuantizationThe images below illustrate detection results on photovoltaic panels affected by soiling:
• Pre-quantization model (FP32 baseline)
• Post-quantization model (INT8 optimized)
This activity aligns with REBECCA project’s broader objective: advancing deployable, scalable, and efficient Edge AI solutions capable of operating in real-world, resource-limited scenarios.
