Position-Aware Network for Concealed Defect Detection

Concealed defect detection (CDD) is designed to identify defects that are not visually obvious and highly blend into the background. However, there are still major challenges in realizing real-time and high-precision hidden defect detection compared to traditional defect detection. To this end, we propose a novel Positional-Aware Network (PANet) to improve detection performance. PANet uses an optimized enhanced pyramid vision transformer to extract multi-scale global context information, and combines adaptive context encoding module to initially identify defect areas. In addition, to accurately locate the defect object, we designed a fine-grained position extraction module, which uses the attention mechanism to capture the position information and mining the subtle features to enhance the position representation ability. Finally, we introduce a boundary refine module to improve the accuracy of the final defect prediction. The experimental results show that on the CDS2K dataset, our proposed PANet is not only superior to the most advanced methods, but also significantly reduces the number of model parameters, showing higher efficiency and robustness.