Automatic detection of boosted Higgs and top quark jets in event image

We build a deep neural network based on the Mask R-CNN framework to detect the Higgs jets and top quark jets in any event image. We propose an algorithm to assign the top quark final states at the ground truth level so that the network can be trained in a supervised manner. A new jet branch is added to the network, which uses constituent information to predict the four-momenta of the original parton, thus intrinsically implementing the pileup mitigation. The network can predict both the shapes and the momenta of target jets. In terms of reconstruction efficiency and accuracy, we demonstrate that the network outperforms an idealized traditional analysis. We also show that the performance of the network does not degrade much when applied to events from a different process than the trained one and to events with overlapping jets.