Improvement And Application Of U-shaped Network Model For Bacterial Electron Microscope Image Recognition And Segmentatio

The morphological study of bacterial electron microscope images plays a vital role in the study of drug mechanism of action.In order to further analyze the morphology of bacterial electron microscope images,the main research work and achievements are as follows:(1)An improved attention residual network(Attention Residual UNet Plus,ARUP)is proposed.(i)By designing the network structure,the network can be integrated with multi-layer feature maps to obtain more spatial information.(ii)Improve the attention gating mechanism and cyclic residual blocks,improve or reduce the attention of the network to different regions,and avoid the waste of computing resources and the redundancy of model parameters.(iii)The weighting function of MS-SSIM,L1 loss and BCE is used as the loss function of the network to improve the segmentation and classification accuracy of the network.Experimental results show a better performance over the other networks.(2)An image segmentation network with Full-scale Jump Connection and Weight Deep Supervision(Full-scale Jump Connections with Wight deep Supervised Trans UNet,FS-Trans UNet3 +).(i)In order to improve the accuracy of image segmentation and classification,connect the connection between the encoder and the decoder of the Trans UNet network,And the decoder internal connection for refactoring,Design of a new and improved Transformer network for the FS-Trans UNet3+,Which can make full use of the semantic features of medical images,Combining the full-scale connection mode and the Transformer framework,Structure reduces the network parameters compared to Tras UNet,MS-Trans UNet++.(ii)Weight Depth Supervision is used to learn the hierarchical representation from a comprehensive aggregated feature graph,so as to improve its spatial position learning ability.Finally,the FS-Trans UNet3+ network and other networks were classified and segmented in multiple image complex scenes,and the experimental results showed that the classification and segmentation effects achieved good results.