Cluster-Locating Algorithm Based on Deep Learning for Silicon Pixel Sensors

Abstract

The application of silicon pixel sensors provides an excellent signal-to-noise ratio, spatial resolution, and readout speed in particle physics experiments. Therefore, high-performance cluster-locating technology is highly required in CMOS-sensor-based systems to compress the data volume and improve the accuracy and speed of particle detection. Object detection techniques using deep learning technology demonstrate significant potential for achieving high-performance particle cluster location. In this study, we constructed and compared the performance of one-stage detection algorithms with the representative YOLO (You Only Look Once) framework and two-stage detection algorithms with an RCNN (region-based convolutional neural network). In addition, we also compared transformer-based backbones and CNN-based backbones. The dataset was obtained from a heavy-ion test on a Topmetal-M silicon pixel sensor at HIRFL. Heavy-ion tests were performed on the Topmetal-M silicon pixel sensor to establish the dataset for training and validation. In general, we achieved state-of-the-art results: 68.0% AP (average precision) at a speed of 10.04 FPS (Frames Per Second) on Tesla V100. In addition, the detection efficiency is on the same level as that of the traditional Selective Search approach, but the speed is higher.

Advances in nuclear detection and readout techniques

“…[517] reports a resourceefficient FPGA-based neural network regression model for ATLAS hardware muon trigger system, which achieves a simulated network latency and dead-time of 122 and 25 ns, respectively. [517,518] proposed an FPGA-based L1 trigger system to perform similarly to the vertex fitting trigger but with fewer logic resources. Trigger decisions are created by a multilayer perceptron (MLP) model to distinguish the signal events from noise events and achieve an efficiency of higher than 99%, with a latency of ∼ 128 ns.…”