Residual shape adaptive dense-nested Unet: Redesign the long lateral skip connections for metal surface tiny defect inspection

Abstract

The state-of-the-art metal surface defect inspection methods have two problems: (1) they are sensitive to tiny defects because of their extreme small sizes, and (2) they cannot accurately locate the random appeared defects whose semantic relationship with the background context is weak. To solve these problems, Residual Shape Adaptive Dense-nested Unet, a pixel-based defect inspection method is proposed, to obtain the exact shape and location of the defect, by (1) assembling different depth Unet branches with dense skip connections as the feature extractor to combine multi-semantic level visual features; (2) adding Residual Shape Adaptive modules on the dense skip connections to help the model locate the defect regions; and (3) introducing the multi-branch training method which enables model pruning to reduce redundant parameters and accelerate the inspection speed. Experiments are conducted and demonstrated that the Residual Shape Adaptive Dense-nested Unet achieves the best performance among the state-of-the-art defect inspection methods.