Research On SEM Image Segmentation Of Mixed Minerals Based On Global Feature

The determination of the composition of mixed minerals has important significance in the mineral processing and microwave metallurgy industry.By segmenting various mineral components in SEM images of mixed minerals,we can obtain different physical information such as mineral category,grain size,inclusion content,distribution,etc.,which can provide important parameters such as sample structure and dielectric properties for subsequent mineral processing and heat treatment processes.However,due to complex situations such as adhesion of non-identical minerals and weak semantic features of small target minerals in mixed mineral images,current image segmentation methods are prone to many mineral misdetections and omissions.Therefore,this article combines the idea of global features and conducts structural design research on the U-Net model based on deep learning methods to improve the segmentation accuracy of mineral images.The main research content of this paper is as follows:(1)To address the problem of difficult balance between model size and segmentation accuracy of current deep learning-based mineral image segmentation methods,this paper proposes a hybrid mineral segmentation method based on improved dense feature mapping.Firstly,this article designs a lightweight convolutional unit based on the characteristics of deep separable convolution and SE attention mechanism.Then,the backbone network is constructed based on the characteristics of densely connected large-scale feature fusion.Finally,it is applied to U-Net encoders to improve feature extraction capabilities.The experiment shows that compared to the basic U-Net model,this method improves the segmentation indicators MIo U,Recall,and Precision by 2.17%,0.74%,and 1.73%,respectively,achieving high-precision mineral segmentation tasks with lower model parameters.(2)To address the problem that convolutional operations are difficult to obtain global feature relationships and are prone to the loss of small target features,this paper proposes a hybrid mineral segmentation method that combines global feature modeling and feature reduction.Firstly,the model constructs global feature dependencies by introducing a window self-attention mechanism to improve the model’s ability to distinguish different minerals.Then the inductive bias characteristic of convolution operation is combined to enhance the fitting ability of the model.Finally,the skip connection of U-Net is improved by compensating for small target features lost due to down sampling operations through feature cancellation and feature enhancement of channel and spatial.The experiment shows that the MIo U,Recall,and Precision of this model are 0.9455,0.9668,and 0.9765,respectively,and can effectively solve the problems of mineral misdetection and missed detection.(3)To address the problem that too many parameters are required to build a selfattentive mechanism and difficult to deploy,this paper proposes a hybrid mineral segmentation method based on Mobile ViT and generative adversarial network.Firstly,self-attention mechanism is applied in U-Net’s skip connections to guide the expression of decoder output features based on more complete feature relationships in the encoder.Then,feature selection is performed on the feature map for correlation calculation,reducing the size of the input matrix to reduce the parameter count of the self-attention mechanism.Finally,the model is trained based on the generated adversarial network,the segmentation model is optimized based on the error between the segmentation results and the actual labels.The experiment shows that the MIo U,Recall,and Precision of this method are 0.9385,0.9648,and 0.9705,respectively,which enable the segmentation model to construct a complete global feature relationship at the cost of a small amount of parameter increase.