Difficulty-aware edge-cloud collaborative OCR for low-latency grid mobile image recognition

Abstract

Mobile image recognition is becoming a basic capability for grid inspection, equipment testing, and field data acquisition. In practical substations and line-maintenance scenarios, workers often capture equipment test reports, nameplates, and digital-meter displays under non-ideal illumination, reflection, motion blur, and unstable wireless links. Direct cloud OCR provides high recognition accuracy but increases response time and may expose unnecessary image content, whereas edgeonly recognition is fast but unreliable for low-quality images. This paper proposes DA-ECCO, a difficulty-aware edgecloud collaborative OCR framework for low-latency grid mobile image recognition. The framework integrates terminalside image normalization, ROI extraction, lightweight OCR heads, utility-based scheduling, ROI-level cloud refinement, and structured business validation. A difficulty vector is computed from sharpness, illumination, skew, reflection, and ROI completeness. The scheduler selects local acceptance, encrypted ROI upload, or cloud/recapture fallback by minimizing a latency-error-risk-privacy-payload utility whose weights are selected on a validation split under business constraints. Experiments on 6,240 de-identified real grid field images and a 21-day on-terminal validation show that DA-ECCO achieves 94.1% report-field exact match, 98.5% meter-digit accuracy, and 238 ms median latency, while reducing upload payload by 93.4% compared with full-cloud OCR.