An efficient defect detection method for nuclear-fuel rod grooves through weakly supervised learning

Defect detection of nuclear-fuel rod grooves plays an important role in nuclear industry. The processing quality of grooves has great influence on the safety during the nuclear reaction. This paper proposes a weakly supervised method for detecting defects of nuclear-fuel rod grooves, which mainly uses image-level labels to locate the defects at the pixel level, so as to reduce the complexity of dataset preparation. The core of the method mainly lies in the pre-positioning of defects and the generation of non-defective self-reference templates. We evaluate our method on the groove dataset, which is captured by the industrial camera. Experiments show that the proposed method outperforms existing state-of-the-art method in terms of both accuracy and efficiency. What's more, our method has been used in practical industrial engineering, and the false detection rate of grooves is only 0.171%.