Edge-oriented cable surface inspection: real-time multi-defect detection via OpenCV and hybrid-quantized YOLO on RaspberryPi

Surface defects in cables, including scratches, cracks, bulges, pinholes, and cuts, critically compromise electrical infrastructure reliability. Traditional manual inspections struggle with subtle defects like pinholes, while cloud-based solutions fail to meet real-time requirements in resource-limited industrial settings. This paper presents an edge-optimized defect detection system combining OpenCV-driven image processing and a compressed deep learning model, specifically tailored for deployment on Raspberry Pi 5. Key innovations include: A pruned YOLOv5s model with hybrid quantization (FP16 + INT8), achieving 97.8% mean average precision ([email protected]) across five defect categories while reducing model size to 5.1MB (79% compression); Dynamic computational resource allocation leveraging Raspberry Pi 5’s dual 4K HDMI pipelines to parallelize image acquisition and processing. Implemented with OpenCV’s DNN module and TensorFlow Lite, the system processes 1920×1080 images at 18 FPS with 65ms latency, consuming only 320MB RAM and 7W power.