Application of Improved VGG Net in Metallographic Recognition

The microstructure of dissimilar steel welded joints with varying degrees of damage serves as a crucial criterion for assessing material degradation. Traditionally, the identification of damage extent in metallographic images relies on manual methods. To enhance both the speed and accuracy of this recognition process, we have collected metallographic structures from dissimilar steel welded joints and employed a convolutional neural network (CNN) model for training and validation. Our findings indicate that the use of a CNN model is a viable approach for assessing the damage level in metallographic structures. Among the four CNN models evaluated for image recognition, the VGG Net model demonstrated superior overall performance in terms of accuracy and stability for tissue metallographic recognition. By incorporating an average pooling layer into the VGG Net model, tailored to the characteristics of metallographic images, we were able to further enhance the recognition accuracy and stability, making it particularly well-suited for identifying metallographic images of martensite and austenite microstructures with varying damage levels.