Research On Image Semantic Segmentation Of Pseudo-3D Residual Networks Via Visual Transformer

In recent years,with the rapid development of computer software technology and the continuous improvement of computer hardware technology,image semantic segmentation technology is becoming a very important research hot-spot in the field of machine vision.Image semantic segmentation is essentially a combination of image pixel segmentation and image content understanding,which aims to give each image pixel a specific category of semantic label.Thanks to the continuous development of hardware technology,the continuous improvement of computing ability and the blowout of image data,the image semantic segmentation technology based on deep learning has been widely used in the fields of artificial intelligence,such as film and television effects,unmanned driving,intelligent medical treatment,security monitoring,other artificial intelligence,and so on.The semantic segmentation effect of image is often influenced by many factors including color,shape,texture,and size,which may affect the final processing effect.Traditional image segmentation methods often use artificial design features to achieve pixel classification.This method not only has the problems of complex process and less robust,but also can not give each pixel certain semantic information.Because the deep learning method uses data training instead of artificial design to realize feature representation,it has obvious advantages in global feature extraction and context information description.Therefore,based on the existing work,this paper studies the image semantic segmentation,mainly carries out the following aspects:A semantic image segmentation model is constructed based on the proposed pseudo three-dimensional residual bottleneck units.Specifically,VGG-16 fully connected network is firstly replaced by pseudo three-dimensional residual neural network to obtain more accurate image segmentation results;then,1×1×1 convolution layer is added into the shortcut connection structure to further optimize the image segmentation results.A semantic image segmentation model of transparent objects is constructed based on vision transformer.Firstly,the local features of an image are extracted by convolution neural network;then,the global features of an image are obtained by using the transformer with multi attention to better understand the image context information;finally,the local and global features of an image are fused to improve the accuracy of semantic segmentation.To verify the segmentation effect of the pseudo three-dimensional residual network model proposed in this paper,experimental tests and performance evaluation are conducted on a public graph data sets,and the comparative experiments with other classical algorithms demonstrate that the proposed model has good segmentation effect.Quantitative model based semantic segmentation of transparent objects is realized.To demonstrate the practicability and real-time performance of the proposed model,raspberry pie is selected as the hardware platform and Tensor Flow Lite is used as the software platform of this experiment respectively.Experimental results show that the proposed semantic segmentation model of transparent objects based on visual Transformer can segment specific targets effectively and in real time.