Research On Image Semantic Segmentation Technology Based On Deep Neural Network

Image semantic segmentation is more and more concerned by computer vision and machine learning researchers.Increasing demand for applications requires accurate and effective segmentation mechanisms.Semantic segmentation is also widely used in various fields.The design of semantic segmentation models based on deep neural networks usually revolves around enhancing detailed information and reducing the loss of semantic information.Continuously improving segmentation accuracy and speed is still the main problem at this stage.The main work of this thesis is to deepen the research of semantic segmentation technology based on deep neural network,and aim at the problem of improving the accuracy and speed of semantic segmentation,respectively designing a semantic segmentation model focusing on two aspects.In terms of improving the accuracy,this thesis designs the Ince-DRes Aspp Net model for the problem of rough edges of the segmentation results and missing partial segmentation pixels.This model introduces Dense in the feature extraction and feature fusion parts of the model encoder in order to enhance the representation of the network.The idea of dense convolution is to construct Ince-DRes dense separation convolution structure and D-Aspp dense cavity convolution structure based on coprime factor.The new model can not only widen the number of channels to explore new features based on feature reuse,but also combine multi-scale features to increase the receptive field,collect more dense pixels,and strengthen contextual semantic information connections.The model Ince-DRes Aspp Net conducted experiments on the dataset PASCAL VOC 2012 and the dataset City Scapes.The experimental results show that the model achieves 83.3% and 78.1% segmentation accuracy on the m Io U indicator,respectively,and is more accurate than the basic model under the same training conditions.Increased by 2.7% and 3.6% respectively.At the same time,in order to reduce the loss of semantic information,this thesis further improves the structure of a single decoder into a multi-branch decoder,and combines the low,medium and high levels of feature information to design the Ince-DRes Aspp-Decoder model,which is based on the PASCAL VOC 2012 and City Scapes datasets.The m Io U index has reached 83.7% and 78.8% respectively,and the segmentation accuracy has been improved by 0.4% and 0.7% compared with the previous one.In terms of improving the speed,improving the accuracy of segmentation will increase the amount of model parameters and calculations,and reduce the efficiency of segmentation,this thesis applies the knowledge distillation method to the semantic segmentation technology and designs the Teacher-Student real-time semantic segmentation model.The model uses Ince-DRes Aspp-Decoder as the Teacher network,extracts some of the layers as the Student network,and constructs the Teacher-Student real-time semantic segmentation model by pixel-wise distillation distillation,pair-wise distillation,and Holistic distillation.Compared with the Teacher network,the Student network in this model reduces the space complexity by 63.86 M,the time complexity by 49.59 B,and increases the segmentation speed by 2.35sample/sec.Under the guidance of the Teacher network,the m Io U index of the Student network on the City Scapes data set is 73.7%,which is 0.6% higher than the original lightweight network Res Net18(1.0).At the same time,the segmentation performance also exceeds other lightweight real-time semantic segmentation network.