| The current particle physics experiment has some issues,including a large amount of data,complicated particle correlations,a variety of particle types,complex data,and a sluggish rate of data processing.Problems such as low particle recognition accuracy and insufficient storage space are urgent problems.In particle physics experiments,cluster detection can determine the trajectory,energy,momentum,track reconstruction,timeof-flight measurement,reduce the incident position and direction of particles,calculate energy deposition and transfer,improve particle detection,etc.At present,charged particle trajectory,energy,momentum,data processing,and analysis are the key applications for particle cluster detection technology.This paper conducts in-depth research on particle cluster detection algorithms,including beam experiments on HIRFL(Heavy Ion Research Facility in Lanzhou)to collect data,preprocess the data,and then use different methods for detection.(1)Employ traditional image processing methods are used for detection;(2)Traditional machine learning methods are used for detection;(3)One-stage and two-stage detection algorithms based on deep learning are built,then the position and classification of clusters are realized.For data preprocessing,the paper uses the 3 criteria for background modeling.Compared with the multi-frame average technique to build a background,targeted frames can be excluded from the background modeling,which improves the robustness of the background.After that,clustered data frames are created by using data filtering and background difference,which can reduce the amount of data.The traditional image processing and machine learning methods are designed to detect clusters,which include the steps of enhancement,segmentation,classification,and recognition.Segmentation is where the two approaches diverge most.The image processing method uses threshold-based segmentation,and the machine learning approach employs hierarchical clustering for segmentation.In addition,the classification criteria according to the aspect ratio are designed.Both one-stage and two-stage detection models based on deep learning are built,Conv Ne Xt and its variants based on convolutional neural network architecture and Swin Transformer and its variants based on Transformer architecture are used as backbone networks for feature extraction,and finally,decoupled head is used to divide the task into regression task and classification task.The results of the two models are also compared and analyzed. |
