Accelerating CNN Model Training and Real-Time Face Recognition on GPU Optimized with OpenCL
Abstract
Currently, it is proposed to use CUDA (Compute Unified Device Architecture) technology on GPU (Graphics Processing Unit) to accelerate tasks such as deep learning. However, the fact that CUDA supports only NVIDIA GPUs causes resource problems and additional costs for the user. In this article, a real-time face recognition system was developed. The system was built based on a lightweight CNN (Convolutional Neural Network) model and was trained on image samples of 160×160 pixels. When the model training process was performed on a device without an NVIDIA GPU, speed problems were eliminated using OpenCL (Open Computing Language). The CNN model training process built for face recognition was accelerated on GPU using OpenCL. The training speed, accuracy, and real-time face recognition performance were compared on CPU (Central Processing Unit), a regular GPU, and a GPU optimized with OpenCL based on FPS (frames per second) and latency (ms). According to the experimental results, the optimization performed on the GPU using OpenCL significantly increased the model’s performance to 17 FPS. This is a more effective result than the CPU (4 FPS) and GPU (9 FPS) benchmarks. The proposed approach allows for stable and fast face recognition even on devices with limited resources.