FPGA Implementation of Convolutional Neural Network for Defect Identification on Swiven Cap

Abstract

AbstractIn this work, Convolutional Neural Network (CNN) is applied for defect identification of Swiven Cap (one type of medical component) based on Field-Programmable Gate Array (FPGA) implementation. Caffe is used as the platform to develop the CNN model. After training phase, a confusion matrix is used to validate the accuracy, precision and recall rate of the trained model. Once the validation is confirmed, OpenCL software is used to develop cross-compile algorithm for System on Chip FPGA (SoC FPGA) implementation on SoC FPGA. This hardware implementation is then used to perform inference of the trained model. Considered surface defects are dent, scratch and black dot. By using a confusion matrix as the evaluation criteria, an accuracy of 88.9% is achieved. The well-trained model is further implemented on FPGA using Altera DE1-Soc development board. PipeCNN framework is used to support FPGA-based inference of the trained CNN model. The inference time for one test image using Altera DE1-SoC is around 300 to 400 milisecond (ms) whereas the inference time using CPU is around 19 to 21 s(s). The performance of Altera DE1-SoC is nearly 63 times faster than the performance of CPU. This prove that the FPGA implementation is applicable alternative for hardware accelerator to speed-up the CNN inference process.KeywordsConvolutional neural networkCaffeDefect identificationPipe CNN