Research On Lightweight Semantic Segmentation Method Based On Mutual Learning

In our daily life,the human eye is an important channel for human to receive information.When people deal with information in daily life,images occupy a very important position.With the development of computer vision,it has replaced the human eye in many aspects,and even has stronger judgment than the human eye in some aspects.Especially with the proposal and rapid development of Deep Neural Network,the computer has completed many complicated tasks instead of human beings.For image analysis,segmentation is a very important task.Its main goal is to recognize the contour of the object in the image,and then analyze its category and behavior by clipping a certain kind of object.Semantic segmentation is a kind of task in image segmentation.Its main function is to recognize the contour of all the objects in an image and the category to which the objects belong.In recent years,with the rapid development of semantic segmentation,networks such as PSPNet and DeepLab have achieved very good results in accuracy.But in practical application,we will pay more attention to the running speed and parameter size of the model besides the accuracy of the model.In order to meet the real-time processing requirements of the model on different hardware platforms.In view of this,the research of lightweight semantic segmentation network is a popular research direction,and it is a very meaningful direction.This paper proposes a new training method for lightweight semantic segmentation network,which aims to improve the accuracy of the model through mutual learning strategy without increasing model parameters.There are three main works and innovations in this paper:1)in this study,the deep mutual learning training strategy is transferred to the semantic segmentation network,and the network accuracy is improved by mutual supervision training between two lightweight semantic segmentation models;2)in view of the imbalance of data sets,a new loss function:AWLoss is proposed.To some extent,the loss function maintains the positive and negative sample equilibrium in the gradient update by dynamically monitoring the proportion of positive and negative samples in the training batch;3)aiming at the difficult cases in the data set,this study proposes a new loss function:AFLoss.By analyzing the distribution characteristics of the loss function in a certain period of time,the loss function can dynamically monitor the difficulties in the training samples.4)after testing in Pascal VOC 2012 and cityscapes,the accuracy of the above strategy has been improved by nearly 1%mIOU.