Weakly supervised casting defect recognition method based on attention mechanism

Abstract

The invention relates to a weak supervision casting defect recognition method based on an attention mechanism. The method comprises the following steps: firstly obtaining a ray image of a casting, carrying out the multi-label weak marking of the casting types and defects in the image, and forming a training set; and then training a ResNe-50 network model on the training sample set, forming an attention mechanism by utilizing multi-label information, enabling the model to accurately identify casting defects under the condition of weak supervision, and then calculating an activation mapping graph to obtain accurate positions of the defects. According to the method, under the condition that the casting defect recognition accuracy is guaranteed, the manpower consumption of labeling data is reduced, and finally the production cost is reduced.

Description

technical field [0001] The invention relates to a casting quality inspection technology, in particular to a weakly supervised casting defect recognition method based on an attention mechanism. Background technique [0002] The ray inspection system is a nondestructive inspection method to ensure the overall quality of industrial castings. It can detect internal defects of castings without damaging the castings. The ray system performs digital radiography on the target casting from different angles, and then manually observes the ray pictures according to the inspection personnel to judge whether the quality of the tested casting is qualified. Manual observation can detect internal defects of castings in radiographs based on inspection experience, but with the fatigue of inspectors, the detection efficiency and accuracy are greatly reduced. The realization of casting defect detection automation is of great significance for improving production efficiency, ensuring product qu...

AI Technical Summary

Problems solved by technology

Although the results of the numerical calculation method are relatively intuitive, for complex backgrounds and structures, the adjustment of the overall algorithm and threshold parameters becomes very complicated.

Although the machine learning method has good accuracy and generalization, whether it is the design of the manual algorithm or the labeling of the training samples, it is very labor-intensive and directly increases the production cost.

Images