In application

production lines continuously provide new training samples and defect types

requiring models to use learned knowledge and combine new samples to learn quickly and have incremental learning ability. To address this issue

an incremental learning method of defect detection for texture surface is proposed. The present algorithm consists of an adaptive convolution/transposed convolution module and a texture surface defect detection network. The former allocates weights to measure the relevance of the detection model parameters to the current training category and endows the model with incremental learning ability. The latter designs reconstruction and segmentation branches

and combines adversarial learning to improve the reconstruction quality and segmentation performance of the model for defects in texture surface. The present algorithm simulates the incremental learning on the MvTec AD texture class of defect detection public dataset

achieving an AUROC accuracy index of 97.6%. The effectiveness of the present module in ablation experiments is also verified. The research evaluates the performance of the present algorithm in real-world scenarios by collecting and testing data from printed substrates on the production line. The achieved average detection rate of 98.7% across four classes serves to validate the application of the algorithm.