Objects Detection Deep Learning System Based on 2-D Winograd Convolutional Neural Network
Fangyi Liu, Che-Lun Liao, Po-Wen Chou, Yu‐Cheng Fan · 2021 IEEE 10th Global Conference on Consumer Electronics (GCCE) · 2021
In recent years, artificial intelligence (AI) grows exponentially. Traditional chips cannot satisfy modern applications of AI requirements because AI chips must provide significant computation speed on deep learning. In order to solve this problem, 2-D Winograd Convolutional Neural Network integrated circuit that decreases the amount of multiplier and reduces the complexity of computation is proposed to speed up convolution operations. This chip is designed as the core circuit of Yolov4 object detection system. To improve the performance of Yolov4, the Yolov4 structure based on Modified Spatial Attention Module (MSAM) is designed in this paper, and batch size 4 is selected to achieve high-precision object recognition. In our experiment, the presented system can reach exceed of mean Average Precision (mAP) 73.01%.