Deep Learning Approaches for Defect Segmentation on Composite Materials using Infrared Thermography

Abstract

Infrared Thermography (IRT) is widely used for detecting defects in composite materials. However, accurately identifying these faults presents significant challenges due to the complexity of the thermal properties of composites and the variability of inspection conditions. In this research, we propose a new approach based on deep learning for detecting defects in composite materials using infrared thermography. The proposed model integrates two neural network architectures: the DeepLabv3 spatial neural network and the BiLSTM temporal neural network. The combination of these architectures allows for efficient analysis of the spatial and temporal characteristics of thermal images, improving the accuracy of defect identification. DeepLabv3 is used to segment and highlight areas of interest in thermal images, while BiLSTM is responsible for analyzing the temporal evolution of temperatures in these areas, providing a more comprehensive view of the thermal behavior of defects over time. This approach allows for more precise and robust detection compared to traditional methods. Experiments were conducted using a real dataset composed of thermal images of Carbon-fiber reinforced polymer CFRP subjected to different test conditions. The results demonstrate that the combined use of DeepLabv3 and BiLSTM significantly improves defect detection accuracy, outperforming traditional techniques. Thus, this research contributes to the advancement of defect detection techniques in composite materials using infrared thermography and deep learning, demonstrating the potential of neural network architectures for applications in inspection and quality control.