A convolutional neural network accelerator on FPGA for crystallography spot screening

The X-rays from a light source produce a lot of data in a very short period of time, and only a tiny fraction of it is meaningful to study. Traditional statistics-based methods rely heavily on experienced experts to classify it. How to obtain these data efficiently and accurately is one of the challenges that light sources are now facing. To solve this problem, a new FPGA-based Convolutional Neural Network (CNN) accelerator is introduced for screening Bragg spots in diffraction images produced at electron laser light sources. The proposed method enhances the target by filtering the data and aims at the region of interest through heat map analysis. A lightweight hardware-friendly convolutional neural network architecture called BraggNet is proposed. The proposed network contains a feature extractor and a classifier. Quantization and layer fusion method is proposed to compression the model. The accelerator uses pipelining between layers and on-chip quantized weight reuse to reduce the time of data transfer process. Data streaming is used between layers in the accelerator. The proposed accelerator is implemented using HLS and FPGA ZeU104. The results show that the proposed method achieves 82.66% accuracy of BraggNet, and 82.16% accuracy of quantized BraggNet, which is higher than the existing method. The proposed method is suitable for data screening at the receiving end of the light source detector.