EHSGNet: A novel edge and high-level semantic guided network for CFRP subsurface defects detection

Abstract

Infrared thermography, due to its fast detection speed, low cost, and intuitive results, is widely used in carbon fiber reinforced plastics (CFRP) materials detection. However, CFRP infrared image noise and lower resolution lead to intrinsic similarly and edge disruption, making it impossible to obtain satisfactory results. Additionally, the current methods require a lot of preprocessing and post-processing operations. Since camouflaged object detection has achieved good results in solving intrinsic similarly and edge disruption. Therefore, this paper proposes an edge and high-level semantic guided network (EHSGNet) based on camouflaged object detection for infrared thermography CFRP subsurface defect detection, achieving end-to-end detection of CFRP. This paper also constructed a dataset containing CFRP samples with both natural defects and artificially simulated internal defects to validate our method. Since the temperature change is achieved in the temperature-controlled chamber by means of hot air heating, it can heat the sample uniformly and improve the quality of CFRP infrared images, so our experiments will be carried out in the temperature-controlled chamber. The experiments were conducted on CFRP specimens with different sizes, shapes and depths of artificial defects. The experimental results demonstrate that the proposed EHSGNet outperforms current approaches significantly on our dataset, with precision (PRC) and recall (RCL) reaching 94.3% and 93.1%, respectively.