Trustworthy Object Detection Method Based On Second-Order Distribution Modeling

Object detection task is an important branch of computer vision which aims to predict the category and location of objects simultaneously.It has a wide range of applications in automatic driving,medical image recognition,defect detection and other fields.In some security sensitive scenarios,it is very important that the object detection model can give reliable uncertainty estimation.There are three main types of trustworthy object detection methods: Bayes trustworthy object detector,Ensemble trustworthy object detector,First-order distribution modeling trustworthy object detector.The calculation cost of Bayes Neural Network and Ensemble method is too high due to the need for multiple forward propagation.The disadvantage of the latter is that it can’t reflect the degree of confidence of the probability given.In order to solve the problems in the above related work,this paper proposes a trustworthy object detection method based on the second-order distribution modeling method,which includes the following research contents:(1)A trustworthy object detector incorporating metric learning is proposed for the task of detecting targets that are sensitive to classification uncertainty,by introducing Dirichlet distribution and deep metric learning.This approach establishes the Dirichlet distribution as the conjugate prior of categorical distribution,enhances the classification branch of Retinanet to estimate the parameters of Dirichlet distribution,and transforms these parameters into equivalent subjective opinions based on subjective logic theory.As a result,the model can quantify the degree of uncertainty in prediction.In addition,to prevent the model from producing misleading evidence,this method employs an proxy-based deep metric learning approach that aims to increase the distance between different semantic target features.This enhances the distinctiveness of the learned features for various types of targets.Ultimately,our experimental results on real data demonstrate that our proposed approach not only outperforms other methods in terms of detection performance but also provides more effective uncertainty estimation.(2)Aiming at the object detection task that is both sensitive to classification and location uncertainty,we employ Gaussian inverse gamma distribution to model the regression branch of the aforementioned model.And then,by introducing the Cluster-NMS method,we propose a trustworthy object detector incorporating Non-Maximum Suppression of uncertainty perception.By modifying the Retinanet regression branch to predict the parameters of a Gaussian inverse gamma distribution,this method is capable of predicting both aleatoric and epistemic uncertainty during one forward propagation.Additionally,to effectively and precisely eliminate redundant detection boxes during the Non-Maximum Suppression stage,this approach integrates model prediction uncertainty estimation with the Cluster-NMS method,thereby simultaneously considering the accuracy and reliability of predictions.Ultimately,experimental results on real datasets validate the efficacy and superiority of this proposed method.(3)Based on the proposed trustworthy object detector,a intelligent verification system for secondary wiring junction caps has been designed to replace manual inspections.This prototype system is capable of effectively detecting misconnections,missed connections,and multiple connections of power cables by providing the function to reject difficult samples.And simultaneously,this system incorporates user information isolation,member information management,and exception sample backup to address other service requirements that may arise in this scenario.