FPGA BASED HARDWARE ACCELERATOR FOR CONVOLUTION NEURAL NETWORK

Abstract

Transport layer of IoT ecosystem relies on optics that work as frame relays in diverse areas. Real-Time images being captured find advantages in terms of latency and decision outcome when the edge computing is preferred keeping in consideration restriction on mobility and increased privacy concerns. These objectives can be achieved by Convolution Neural Networks (CNN) at specialized speed and precisely the right amount of adaptability to handle various computer vision algorithms. It is challenging to design a control accelerator for a CNN that requires a lot of recurrent feature extraction process. Driver assistance systems' real-time, safety-critical face emotion detection classification task is accomplished using a deep learning approach. In this study, an effective overlay for LeNet-5-based face emotion detection and classification utilizing the ZYNQ architecture is proposed. Using the Facial Expression Recognition dataset (FERD) and a pre-trained LeNet-5 Model, a multimodal ZYNQ evaluation system assesses the Overlay's capacity towards inference. Combining Hardware and Software by co-design reduces bandwidth requirement by 30%. Efficiency of per image inference reduces from 2.9 seconds to 30 milli seconds, an uplift of almost 98%. Impact on throughput is extensively visible from 0.5 GOP/sec to 3.6 GOP/sec. Experimental results show similar output in terms of classification accuracy by reducing the amount of time taken to complete classification, thereby making resources available for more data loads. Classification accuracy in both the methods are approximately 98%, thereby making the hypothesis of using Convolution Neural Networks (CNN) with specialized speed a better solution for real-time traffic sign classification.