Vlsi Implementation of a Convolutional Neural Network for Binary Image Classification

Abstract

Convolutional neural networks (CNNs) are very popular in computer vision, speech processing, image processing and classification applications. In this work, a CNN model is developed using VHDL for classifying gray-scale images. FPGA is used for efficient implementation of a CNN architecture. CNN architectural components, convolution, pooling, activation functions and fully connected layers are designed using VHDL and implemented on Kintex-7 FPGA. Sigmoid activation function is used in the final layer of CNN for classifying gray-scale images. The LUT(Look-Up-Table)-based sigmoid function is designed using VHDL and the final output of a particular class is shown in normalized form. Firstly, RGB images (dogs and cats) of size 8X8 are converted into gray-scale images and then applied to the CNN architecture. The synthesis and simulation of the proposed CNN architecture is being carried out using the Xilinx Vivado 2015.2 IS tool. The simulation output shows that the proposed CNN architecture is able to classify cat and dog images with a probability of value greater than 90%.