Research And Application Of Industrial Defect Visual Detection Algorithm Based On MAML Architecture

Visual inspection of the appearance defects on industrial products has always been a research hotspot pursued by industry and academia.Most of the existing product defect detection applications are task-driven systems,which can only detect the defects of a certain kind or a certain class of products.Once there is a new defect detection task,the system must be redesigned,which means the system lacks a certain generalization capability.The reason is that most of algorithms used in the designed system are designed for the detection of the specific products’ defects.Therefore,it is of practical significance to study the generalization of detection algorithm.We note that industrial products have some similarity,such as stains or texture jumps,and most of the defect detection algorithms can also be viewed as permutations and combinations among operators,which provides a theoretical basis for building a generalized and adaptive industrial defect detection system.The main work of this thesis is as follows:Based on the idea of Model-Agnostic Meta-Learning(MAML),an adaptive visual detection model Me Detction(Meta Detection)was proposed to learn basic knowledge of industrial defects form multiple known industrial defect datasets.The model uses the Siamese network to extract differential features,minimizes the influence of defect types on model generalization,highlights defect features and improves model detection performance.At the same time,the coordinate attention mechanism is added into the model to realize the feature enhancement in the two coordinate dimensions.The visual defects dataset named BS defects was constructed from real factory production environment,which supplemented the existing industrial visual defects data benchmarks.The simulation experimental results based on BC defects dataset and other public datasets have demonstrated the effectiveness of the proposed adaptive visual detection model for industrial defects.Taking industrial products on actual production lines as research objects,an industrial defect detection application system with strong generalization was designed and developed,which was applied to the related production enterprise.The system adopts a combination operators’ approach to design the algorithm of defect detection,so there is no need to redesign the system when the new product comes online.The generalizability of the system is further enhanced by embedding the Me Detection model algorithm into the detection system.Drag-and-drop,open user-interface makes the algorithm development and deployment more convenient.Good results have been obtained in the defect detection of various products on the designed system,which indicates that the platform has good generalization ability.