Brain-regulated learning for classifying on-site hazards with small datasets

Abstract

Machine vision technologies have the potential to revolutionize hazard inspection, but training machine learning models requires large labeled datasets and is susceptible to biases. The lack of robust perception capabilities in machine vision systems for construction hazard inspection poses significant safety concerns. To address this, we propose a novel method that leverages human knowledge extracted from electroencephalogram (EEG) recordings to enhance machine vision through transfer learning. By pretraining convolutional neural networks with EEG data recorded during construction hazard evaluations, we investigated three common on-site hazard classifications using small datasets. Our results demonstrated that the proposed method resulted in improved accuracy (with an 11% increase) and enhanced rationality of machine learning predictions (as revealed by network visualization analysis). This research opens avenues for further exploration and industry applications, aiming to achieve more intelligent and human-like artificial visual perception, ultimately enhancing safety and efficiency in automated hazard inspection.