An Automatic Design Method for Dynamic Detection Networks of Industrial Surface Defects

The advancements in artificial intelligence technology have enabled Neural Architecture Search (NAS) to become a method for the automatic design of network models, initially validated in industrial surface defect detection. However, the current automatic design methods for detection networks are built on a fixed search framework that search only for internal operational connections. Due to significant differences in defect features with different scales, this fixed framework restricts the diversity of network expression capabilities. Therefore, this paper proposes an automatic design method of dynamic detection networks for industrial surface defects. Compared to existing methods, the proposed method can provide a more flexible search framework across a wider range of search dimensions, including cell operations, cell width, network depth and scale dimensions. Experiments show that the proposed method can more effectively adapt to defect features, improve accuracy and efficiency, and overcome design limitations.