The on-line inspection system for pultruded GFRP plate based on machine vision and deep learning

A novel on-line inspection system predicated on the confluence of machine vision and deep learning technologies was introduced in this research to address the inefficiencies and high omission rates of traditional detection methods for pultruded glass fiber reinforced polymer (GFRP) plate. In addition, an algorithm for the defect detection of the pultruded GFRP plate synthesized through the amalgamation of image processing techniques and advanced deep learning algorithms was presented. The design of the on-line inspection system, the investigation of an image preprocessing algorithm as well as the subsequent collection, categorization and compilation of a dataset comprising the defect images of the pultruded GFRP plate were delineated. In this study, ZF-Net, Xception and squeeze-and-excitation (SE)-ResNet-101 were utilized as the feature extraction networks within the Mask region-based convolutional neural network (R-CNN) framework, and a comprehensive analysis was made of defect detection efficacy in the pultruded GFRP plate. The findings reveal that SE-ResNet-101 achieved superior performance, with an accuracy of 97.52%, an enhancement of 9.78% in precision and an improvement of 0.06 in the F1 score over traditional methods, while exhibiting robust performance across varying lighting conditions. This research pioneered a new methodology for the on-line inspection of pultruded GFRP plate, markedly augmented inspection efficiency and accuracy, and mitigated the high rates of false positives and omission rates inherent in conventional inspection methodologies.