Research On Quantification Method Of Uncertainty In Semantic Segmentation Model Based On Evidential Reasoning

In the past decade,deep neural networks(DNN)has penetrated into almost all fields of science.Especially in safety critical areas,deep learning models are often fatal if they make wrong decisions.Nowadays,semantic segmentation models are more and more widely used in safety-critical fields.Therefore,semantic segmentation models are not only required to give high confidence decisions,but also need to clearly point out the reliability of model decisions,that is,the uncertainty of the model decisions can be expressed,then the category of prediction errors can be given a signal through high uncertainty,and the possible catastrophic consequences can be prevented in time.Therefore,it is of great and far-reaching significance to study the uncertainty quantification method of semantic segmentation model for safety scenes such as autonomous driving.Evidential reasoning,as a mathematical method for dealing with uncertainty,can carry out uncertainty modeling and reasoning without prior information,and has significant advantages in the quantification of uncertainty information.This paper conducts research on the uncertainty quantification method of semantic segmentation model based on evidential reasoning.The main research contents of the paper include:Firstly,this paper conducts uncertainty modeling on the semantic segmentation model based on evidential reasoning,which proves the applicability of the method based on evidential reasoning in the field of semantic segmentation.As the core achievement of evidential reasoning,the evidence classifier is very effective for quantifying the uncertainty of the convolutional neural network(CNN)model in the classification task.However,the evidence classifier is currently only applicable to the fully connected layer,and there is no work to study its applicability of convolutional layers.Therefore,based on the basic principle of the evidence classifier,this paper firstly models the evidence of the fully convolutional neural network,and realizes the extension of the evidence classifier to the fully convolutional neural network commonly used in semantic segmentation tasks.Secondly,this paper proposes an uncertainty quantification method for semantic segmentation model based on evidential reasoning.The core idea is to map the pixel-bypixel feature vector of the input sample to the evidence weight,and then convert the evidence weight into a mass function reasonably.This function can make decisions and has the ability to quantify model uncertainty.By combining the evidence classifier with the semantic segmentation model,this paper first puts forward the uncertainty quantification framework of the semantic segmentation model,and then puts forward the uncertainty quantification algorithm of the semantic segmentation model based on this framework,which mainly quantifies the uncertainty of the semantic segmentation model from different perspectives by using the conflict and the ignorance.Experiments show that this method does not need to change the original model structure,training process and loss function,and can quantify the uncertainty of the model without affecting the accuracy of the model.In addition,compared with the current classical methods based on MC Dropout,the quantization method proposed in this paper has significant advantages in time complexity,and the quantization speed is increased by at least 3-4 times.Finally,based on the proposed uncertainty quantification method,this paper makes a quantitative research on the aleatoric uncertainty of semantic segmentation model.Through experiments on Cam Vid dataset,it is fully proved that the conflict and the ignorance can effectively quantify the aleatoric uncertainty of the semantic segmentation model,and the conflict is more sensitive to quantify the uncertainty of the edge of the object,and the ignorance has the ability to identify areas of high uncertainty.