Comparative analysis of efficient implementation of CNN on CPU and GPU for image processing and implementation of basic CNN operations in Verilog

---

Abstract

Artificial Neural Networks can assist machines make intelligent and smart decisions with very limited human interference. This is because these networks can learn quickly and build a model based on relationships between input and output data which are nonlinear and complex in nature. The main goal of implementing these neural networks is to obtain the highest possible accuracy of output with low latency along with low training and testing period. ANNs have shown their potential particularly for the analysis of image data, which is the base in the biomedical field for detection of diseases. CNN (Convolutional Neural Network) is a proven neural network for image dataset according to many studies, in this paper we use CNN model on biomedical dataset- for detection of pneumonia using real chest scans. In this paper we have compared performances of the same algorithm on CPU and GPU based on a few parameters. Research in the area of hardware implementation of CNN using FPGA (Field Programmable Gate Array) has received attention due to the inherent advantages of FPGA. There are several hardware descriptive languages for FPGA like VHDL, Verilog, System Verilog etc. We have implemented certain basic building blocks of CNN in Verilog (a hardware descriptive language used to model electronic systems). In this paper we successfully implemented selected tasks on FPGA which can be adapted to increase the speed and reduce the cost of the system.

---

Keywords

CNN, CPU, GPU, VERILOG, FPGA Pneumonia Detection