Research On Classification Regression Inconsistency And Localization Uncertainty In Object Detection

Object detection is one of the key problems in computer vision research and the basis for understanding high-level semantic information in images.The model trained for object detection should have the generalization capability to new samples,but the inconsistency between the classification and regression tasks in object detection can affect the model’s generalization capability.The results of object detection are influenced by the perceptual uncertainty and accidental uncertainty of the model,with the latter causing localization uncertainty.This paper investigates in depth the reasons for the inconsistency and localization uncertainty between the classification and regression tasks in object detection,respectively proposing solutions and developing systems for relevant application scenarios.This literature focuses on the following contributions:1.To address the inconsistency between classification and regression tasks in object detection,the Feature Fusion Network Model(FMRNet)was proposed,which included Feature Fusion Module(FFM),regression loss function(RMAE)and inconsistency loss function(Lin).A Feature Fusion Module was proposed in order to enhance the capabilities of the model’s feature extractions.This study also further considered the inconsistency between the loss functions and the proposed regression loss function based on Mean Absolute Error(MAE)for the purpose of improving the location quality.Furthermore,in order to solve the problem of the lack of information regarding the interactions between detection heads,an inconsistency loss function was added on the basis of the Feature Fusion Module.The experimental results demonstrated that this study’s proposed methods had surpassed the accuracy of some existing detectors when FMRNet was adopted.2.To address the localization uncertainty in object detection,the Uncertainty Network Model(DPMNet)was proposed,which include Distribution Processing Module(DPM)and uncertainty loss function(UMAE).The sharpness of probability distribution can characterize the uncertainty of data,and thus the Distribution Processing Module was proposed.The uncertainty loss function was proposed to fully utilize the uncertainty of data to optimize the model.The experimental results demonstrated that this study’s proposed methods had surpassed the accuracy of some existing detectors when DPMNet was adopted.3.This paper develops a fall detection system based on an improved YOLO v5 model,which added the Coordinate Attention(CA)and uncertainty loss function(UMAE).To improve the feature extraction capability of model,the coordinate attention module was added.To address the localization uncertainty problem in the system,the uncertainty loss function was added.Through extensive experimental training and testing,the fall system is shown to have improved performance.