Deep Learning Framework for Impact Damage Quantification and Residual Flexural Strength Prediction in GFRP Composite Structures

Accurate determination of residual strength in composite materials is essential for ensuring structural reliability, particularly under low‐velocity impact conditions where damage is often barely visible yet significantly reduces load‐carrying capacity. Conventional residual strength evaluation methods are generally destructive, time‐intensive, and unsuitable for rapid in situ inspection. In this study, a vision‐based framework is proposed to estimate residual flexural strength from surface damage images by integrating surface damage images and ultrasonic C‐scan images with deep learning‐based detection, segmentation, and regression models. A comprehensive dataset of impacted composite specimens was developed using a drop‐weight impact setup. For damage localization, YOLOv8n, SSD, and Faster R‐CNN were systematically evaluated, with YOLOv8 demonstrating superior accuracy and real‐time inference capability. For pixel‐level damage characterization, the ResNet34‐U‐Net architecture was implemented, achieving improved segmentation performance with a test mIoU of 84% and an accuracy of 97% compared to U‐Net. The proposed model was validated using independent test cases, showing low prediction errors ranging from 0.86% to 2%. The framework was deployed on a Raspberry Pi 4 integrated with a camera module for real‐time image acquisition and processing, supported by a graphical user interface (GUI) that displays detection, segmentation, and predicted strength at approximately 8 FPS, demonstrating its suitability for low‐cost edge‐based structural health monitoring applications.