| Title |
A Research on Quantized Lightweight Convolutional Neural Network Structure for Facial Emotion Recognition |
| Authors |
김재명(Jaemyung Kim) ; 강진구(Jin-Ku Kang) ; 김용우(Yongwoo Kim) |
| DOI |
https://doi.org/10.5573/ieie.2020.57.12.51 |
| Keywords |
Face emotion recognition; Deep learning; Data augmentation; Lightweight CNN; Quantization |
| Abstract |
Recently, a study on facial emotion recognition using CNN, which has an excellent performance in the field of computer vision, is being conducted. To obtain high classification accuracy, a CNN structure with a large number of parameters and high computational complexity is required. However, such a CNN model is not suitable in an environment where the use of hardware resources is limited. In this paper, we designed a lightweight CNN structure with a small number of parameters and low computational complexity for an optimized implementation under a limited environment and proposed a quantization technique that can reduce the parameter size and computational complexity while maintaining accuracy. Also, for high classification accuracy, a data augmentation technique using various image processing algorithms was proposed. As a result of evaluating the performance by applying the FERPlus dataset to the proposed floating-point trained CNN model(FP32), the number of parameters was about 1.98M, FLOPs were about 31MFLOPs, and accuracy was about 86.87%. The highest accuracy was achieved compared to other lightweight models. In addition, two quantized CNN models(INT8, INT4) proposed by applying the proposed quantization technique as the number of parameters in the 8-bit model(INT8) are about 495K and the number of parameters in the 4-bit model(INT4) is about 247.5K. Compared to the proposed FP32 CNN model, it was confirmed that the number of parameters was as small as 1/4 and 1/8, but the accuracy drop was less than 0.54%. |