Implementation of High-speed and High-Accuracy Convolutional Neural Network Accelerator for Target Detection Applications

Deep Convolutional Neural Network (CNN) for real-time applications heavily depends on computing efficiency. The efficiency of the CNN algorithm has been improved by modifying Convolutional Neural Network accelerator. Thus, the proposed work development of improved CNN accelerator using Residue Number System (RNS) and implement in FPGA for performance investigation of throughput, latency, accuracy. The non-weighted nature of RNS makes it suitable for high-speed applications. As a result, it requires less space to house input feature maps and kernels. In addition to that, the proposed method will allow for parallel computation and required less power consumption. According to test results, the planned CNN accelerator can run the AlexNet at 165.4 frames per second, 200.5 GOPs per second at its max, and 22.5 GOPs per Watt. Additionally, it can be changed to support VGG-16 for complicated object recognition.