Research On Uncertainty Estimation In Deep Learning And Its Application On Object Detection

Neural networks,which emerged in the 1960 s,have flourished in the past decade.The effectiveness of the deep neural network as a complex function approximator makes us achieve a breakthrough in many fields.Using some form of deep learning model is widely used in the real world,such as face recognition,voice assistants and autonomous driving.However,most of the models in deep learning are deterministic models,whose output prediction lacks uncertainty estimation and is an overconfident prediction.This can be deadly for serious decisions,ranging from economic loss to loss of life.Uncertainty estimation is an important research direction in information theory,which reflects the confidence of prediction.Understanding uncertainty estimation in prediction can help us make better decisions.The main research contents of this thesis are as follows:Firstly,the theoretical basis of deep learning is introduced and the general paradigm of classification task and regression task is analyzed.For these two kinds of tasks,the commonly used deep learning techniques are introduced,including regularization method,gradient descent optimization method,weight initialization,etc.Secondly,this thesis analyzes two main sources of uncertainty in deep learning:epistemic uncertainty and aleatoric uncertainty.The causes and related forms of these two kinds of uncertainty are described in detail,and the motivation and method of measuring these two kinds of uncertainty are introduced respectively for classification task and regression task.Thirdly,aiming at the problem that the traditional convolutional network cannot estimate the uncertainty in the field of image classification,a bayesian convolutional neural network is proposed,which introduces the gaussian distribution to the weight of the network,so that its weight is no longer a single point estimation,so that the prediction can estimate the uncertainty.Through experiments on the classification data set of the benchmark,it is proved that the accuracy of the bayesian convolutional network is comparable to that of the traditional convolutional neural network,and the uncertainty measurement of its prediction can be carried out,which proves the feasibility and effectiveness of the proposed algorithm.Finally,aiming at the high requirements of target detection in the field of automatic driving,namely,high speed is maintained while high precision is detected,uncertainty measurement is introduced to improve the YOLOv3 algorithm,and gaussian modeling YOLOv3 is proposed.In this algorithm,gaussian modeling is carried out on the predicted target frame coordinates in YOLOv3,and the loss function is reconstructed.Then,the uncertainty estimation of the coordinate positioning is used to improve the evaluation standard of detection,so as to improve the detection accuracy of YOLOv3 without losing the detection speed of YOLOv3.Experiments show that the average accuracy of the proposed algorithm for gaussian modeling of YOLOv3 is 3.07 percentage points higher on the KITTI data set and 3.4percentage points higher on the BDD data set than that of YOLOv3,and its frame rate reaches 43.07 FPS,which is only 0.5FPS lower than that of YOLOv3,and it can reach the standard of real-time detection.