Research On Unsupervised Segmentation Algorithm Of Character Image Imprinted By Graphite Electrode

Semantic segmentation is one of the three major research directions in the realm of computer visual and is the foundation for many image processing tasks.However,fully supervised semantic segmentation of images using deep learning technology depends on a mass of manually annotated labels,which brings huge labor costs and severely restricts the application of deep learning technology in various fields.In the industrial scenario of graphite electrode production,in order to facilitate the statistics and management of graphite electrodes in the manufacturing course,it is indispensable to pick out the characters imprinted on graphite electrodes.Semantic segmentation is an important step before recognition,and pixel-level annotation of semantic segmentation datasets will bring huge production costs to enterprises.Therefore,the unsupervised semantic segmentation of character images imprinted with graphite electrodes has important practical significance and is also extremely challenging.Aiming at the high cost of manual labeling of fully supervised semantic segmentation datasets of graphite electrode imprinted character images in industrial applications,this paper studies the unsupervised segmentation algorithm of graphite electrode imprinted character images from two aspects:recurrent generative adversarial network and domain adaptive segmentation.The main contributions of the research are as follows:（1）In the research of recurrent generative adversarial network,an unsupervised semantic segmentation network Cycle GAN-Seg is proposed for character images imprinted by graphite electrodes.First,a multi-scale feature fusion generator is constructed by combining the ideas of cross-layer connections and Atrous Spatial Pooling Pyramid（ASPP）,and an improved attention module is added to improve the network performance.At the same time,a U-shaped discriminator is proposed to discriminate the reconstructed image,and the output authenticity probability map represents the authenticity of each pixel in the reconstructed image,which acts on the cycle consistency loss function to restrain the training of the generator.In the semantic segmentation experiment of graphite electrode imprinted character dataset,the MIo U value is 5.90%higher than that of Cycle GAN,reaching 70.81%.In addition,the complexity of the Cycle GAN-Seg network is analyzed in terms of parameters,computation,and prediction time.The experimental results indicate that the network can promote the segmentation accuracy,reduce the hardware requirements for industrial deployment,and improve the prediction speed.（2）In the research of domain-adaptive segmentation,an unsupervised semantic segmentation network Db FFDA is proposed for character images imprinted by graphite electrodes.First,referring to the design idea of U-Net’s cross-layer connection,a residual domain adaptive segmentation network Res-Adp with dual-branch upsampling structure is proposed.The features of each level in the network upsampling process are upsampled through the residual branch and the convolution branch respectively.The residual branch is responsible for feature alignment to segment the image using domain-invariant features,and the convolution branch is responsible for preserving local-domain features to supplement segmentation details with unique image features in this domain.Meanwhile,using fused feature input further improves the network segmentation performance.Moreover,according to the prior knowledge of the internal continuity of various objects in the segmented image,a segmentation continuity loss function L_Con is proposed,which indirectly improves the segmentation effect of the target domain image by constraining the generation of the segmented image in the source domain.In the semantic segmentation experiment of graphite electrode imprinted character dataset,the MIo U value is improved by 6.45%compared with the unsupervised U-Net network,reaching 70.53%.（3）The effects of hyperparameters on the segmentation performance in the above two networks are theoretically analyzed and experimentally verified.Meanwhile,according to the actual needs of the medical field,the domain-adaptive unsupervised segmentation network Db FFDA is extended to the medical field.It solves the problem of repeated labeling of datasets in the application scenario of fundus blood vessel image segmentation,and reduces the high cost of manual labeling of medical datasets.The experimental results show that the actual segmentation result of the two methods raised in this thesis for unsupervised segmentation of graphite electrode imprinted character images has basically met the needs of subsequent recognition of characters,and is expected to replace the fully supervised learning method in specific industrial scenarios,thereby saving huge manual annotations cost.At the same time,the domain-adaptive unsupervised segmentation network Db FFDA achieves high-precision unsupervised segmentation of fundus blood vessels in the medical field,which is expected to save the high cost of manual annotation of medical datasets.