The Metallographic Image Description Methods Research Of Aluminum Alloy

The composition and microstructure of metal alloy can determine the properties of metal.The automatic generation of metallographic description is an important task of intelligent metallographic image analysis,which is very important for the development of nonferrous metal manufacturing industry.In this paper,the metallographic image of aluminum alloy is taken as the object,and combined with the advanced theory of image processing,natural language processing and deep learning,the metallographic image analysis method is further studied,and the network framework for automatically generating specific language description from the metallographic image of aluminum alloy is realized.The specific research contents are as follows:(1)A spatially enhanced image generation framework for aluminum alloy metallography is proposed.First,VGGNet and ResNet networks were used to extract the image features of metallography,and LSTM networks were used to extract the text features of the problem,and the extracted image and text features were fused.Aiming at the problem of feature redundancy,the feature recalibration method and attention mechanism are adopted to distribute the weight of the image features,and the useless features are suppressed.Taking the extracted features as the input of the full connection layer,the description of the given metallographic image is generated,and the feature map of the attention layer is extracted,and the visualization of the attention map is realized through bilinear interpolation and pseudo-color operation.In order to verify the validity of the method,the metallographic description data set of aluminum alloy was constructed.Two methods were constructed by combining the above image classification network and text classification network respectively to complete the comparative experiment.A large number of experimental results show that this method can generate accurate description of a given aluminum alloy metallographic image.(2)In order to solve the problem of over-fitting caused by the small amount of data,a denmobilenet-based method for the description of aluminum alloy metallographic image is proposed.Combined with DenseNet’s intensive connection operation and the deep separable convolution structure in MobileNet,Swish activation function was used to replace the commonly used ReLu activation function,and the den-mobilenet network was built to extract metallographic image features,fuse with the text features of the problem,and send into the full connection layer to generate language description and attention map after weight allocation.A large number of experimental results show that the den-mobilenet method can effectively solve the over-fitting problem and achieve better results in the aluminum alloy metallographic description data set.(3)In order to improve the learning ability of the model and solve the problem that the attention map cannot locate the metallographic structure well,a text convolutional network based metallographic description generation method for aluminum alloy is proposed.In this method,text features are extracted through the text convolution network,the image features are weighted by the dual attention mechanism,and then the features are fused by the decomposed bilinear pooling method,and the fused features are sent to the full connection layer to generate the description.A large number of experimental results show that the improvement of these three parts can effectively improve the accuracy of the model,and the generated attention map can cover the metallographic structure accurately.