Research On Object Detection Algorithm Based On Anchor Boxes And Loss Function

Object detection technology has developed rapidly in recent years,evolving from two-stage algorithms to single-stage algorithms,and from anchor-based algorithms to anchor-free algorithms.The field is moving towards more streamlined and efficient solutions.Anchor-free algorithms,with their simpler structures and faster detection speeds,represent the current development trend in object detection technology.However,these algorithms face issues such as suboptimal bounding box regression and unreasonable label assignment strategies.To address these problems,this thesis presents the following main contributions:(1)The pseudo anchor-free algorithm is proposed for the problem of low accuracy of the bounding box regression of the anchor-free detection algorithm.we introduce anchor mechanism for the regression task based on the FCOS.The K-means clustering algorithm is used to obtain anchor boxes sizes that more closely resemble actual targets,and an Io U prediction branch is added to guide prediction boxes to regress more accurately.S-NMS is proposed to select high-quality predictors more accurately in the testing phase.The pseudo anchor-free algorithm reduces the number of anchor boxes and computational effort while improving accuracy by 2.4%-3.0% compared to the benchmark FCOS algorithm.(2)To address the issue of unreasonable label assignment strategies for anchor-free object detection algorithms,a dynamic label assignment strategy based on joint loss minimization(JLM)has been proposed.The dynamic label assignment strategy uses the weighted sum of sample regression loss and classification loss as the label assignment metric.The average joint loss of candidate samples is used as the minimum positive sample boundary for the label assignment strategy to filter out low-quality positions in the positive sample space.A more advanced regression loss function(D-EIOU)is designed to participate in the calculation of the joint loss and suppress low-quality prediction frames during the training process.Experimental results indicate that the dynamic label assignment strategy based on joint loss minimization can lead to improved algorithm performance(3)To meet the requirements of detection speed and accuracy in object detection systems,a lightweight object detection algorithm is proposed.This algorithm combines the innovative points of chapters three and four of this thesis,and applies the JLM label assignment to the pseudo anchor-free algorithm to improve the algorithm performance.The backbone network is replaced by Mobile Net V2 to reduce algorithm parameters and improve detection speed.Based on the algorithm designed in this chapter,a object detection system is developed to make detection target selection more convenient and detection results more intuitive,which has practical application value.