Convolutional Neural Network Hardware Implementation for Flower Classification

Abstract

Flower classification becomes more and more important as the medical and industrial world grows. Based on that emergency, Convolutional Neural Network (CNN) proposed a way for computer to recognize flowers in place of human as the data becomes enormous. This study proposes the hardware architecture for CNN which is tested with FPGA. Numbers and type of layers, as well as their properties are also proposed for effective hardware implementation. Math functions that engine the CNN are also well-cared for the smoothness of both feed forward and back propagation processes. Measurements were taken on the proposed CNN; its accuracy and yield were verified. It also appeared that the classification accuracy of the CNN is strongly affected by the training conditions as well as the flower characteristics. This indicates that further image pre-processing can improve the accuracy of the CNN, which can be implemented separately with the CNN or embedded in CNN's first layers by controlling the weights.