Reference-Based Defect Detection Network

The defect detection task can be regarded as a realistic scenario of object in the computer vision field and it is widely used industrial field. Directly applying vanilla detector to achieve promising results, while there still exists challenging issues that have not been solved. first issue texture shift which means trained model will easily affected by unseen texture, second partial visual confusion indicates box visually similar with complete box. To tackle these two problems, we propose Reference-based Defect Detection Network (RDDN). Specifically, introduce template reference context against those respectively. Template reduce from image, feature or region levels, encourage detectors focus more on defective area result. We use either well-aligned images outputs pseudo generator references this work, they are jointly supervision normal samples. solve issue, leverage carried information reference, concentric bigger each proposal, perform accurate classification regression. Experiments datasets demonstrate effectiveness our proposed approach.