Self-adaptive industrial defect visual detection algorithm

The invention relates to the technical field of computer vision, in particular to an adaptive industrial defect visual detection algorithm, which comprises the following steps: S1, adding a feature difference network (SN) on the basis of a 4Conv network commonly used for small sample classification to extract difference features; s2, introducing a position attention mechanism to eliminate feature redundancy and strengthen features; s3, using an MAML training mode to obtain a weight containing defect basic knowledge from other industrial product detection data sets; a plurality of industrial products are selected to form a task set p (T), a model is described as f theta, data of each task is divided into a support set D = {xi, yi} and a query set D '= {xj, yj}, and training of MAML is divided into two stages. The method can emphatically solve the problems of sample imbalance phenomenon, defect type diversity and incapability of pre-determining the detection difficulty of defect features, on one hand, the probl