Classification of High Energy Track Using CNNs

Abstract

Image recognition is a topic that focuses on finding and identifying various specified objects, classes of objects, features in a given input image or video frame, even it's no so clear to see. Deep Convolutional Neural Networks (CNN) is a state-of-the-art technique for image recognition. These convolutional neural networks may require long training time. This training time is depended on the number of classes, the number of training images in those classes, computing power, and the complexity of the neural network.

This master's thesis studies an application of a convolutional neural network in the tracking high energy particles that were produced at CMS detector at CERN.

There are many applications for image recognition. These applications span various domains, such as Face recognition, OCR, Manufacturing inspection and Quality Control, Medical diagnosis, and Autonomous vehicle. The application discussed in the thesis is to identify particles that have "high" energy that the "low" energy particles using images of the hit pattern formed in the tracker of CMS detector.

The importance of this application is that this image identification process is faster once the model is trained. In the future, LHC will go into phase 3, because of this, a huge amount of data will get generated per collision. Traditional methods will fail to read and keep data in time as the hardware's writing speed is limited. This thesis tries to show the use of a convolutional neural network at a trigger level by using images data of tracks of particles per event.

Two models have been created for analysis. One model is called the toy model, and this is a simplified model. The second model is based on data of simulated hits taken from CMS tracker. The performance of both models has been tested by putting different conditions on CNN architecture and tracker geometry.