Research And Realization On Technology Of Image Semantic Segmentation Based On Deep Learning

Full convolution neural network is based on convolution neural network,which replaces the last full-connection layer with 1*1 convolution layer of the same dimension,and then supplemented by deconvolution and upper sampling layer.Semantic segmentation of images can be accomplished by full convolution neural network.Semantic segmentation of images refers to extracting the object layer of images from the visual layer through certain means,and obtaining the conceptual layer information of images by synthesizing the visual layer and the object layer.That is,the semantic information of the image.The basic theory of convolution neural network is systematically studied,and the concept of image semantics segmentation is presented.Then the full convolution neural network is introduced.Taking the full convolution neural network FCN as the research object,the whole convolution neural network from structure to training is systematically explored.The training of full convolution neural network needs deep learning technology,and deep learning framework plays an important role in deep learning practice.The topic go on the training of FCN with caffe which is an open source deep leanrning framework developed by Google and the data of voc2012.The MIOU's value of training results reaches 59.65%.In order to make the training more targeted to specific scenarios,a self-made data set was added to the voc2012 data set,and the FCN was trained the second.After training,the MIOU's value for specific scenarios reached 62%.Besides FCN,the derivative versions of U-net and Sege Net of FCN are also introduced,and the structure of U-net and Sege Net is specially introduced in detail.U-net is mainly for deep learning tasks with small sample size.Seg Net is widely used in the field of automatic driving and intelligent robots.Finally,the training process of FCN is analyzed emphatically.Visualization analysis of each stage of training process is carried out.The detailed process of image semantics segmentation using FCN is vividly displayed through visualization of convolution core and training intermediate results,which makes the research of the subject more thorough.The main work are as follows:1)FCN full convolution neural network is used in the structure.The whole full convolution neural network uses VGG in the convolution part and multi-stage sampling in the upper part in order to maximize the classification accuracy.2)Data sets,for a specific application scenario,using the combination of common data sets and customized data sets,to process the original images collected by ourselves according to the format of the common data sets.At last,adding the processed original images to the common data sets,which makes the common data sets more applicable to a particular scenario.