摘要
为提高自然光下人脸表情识别精度,提出了一种结合残差网络与通道注意力机制的人类面部表情识别方法.采用表情分类与人类面部表情特征提取融合至一个端到端的深度卷积神经网络的结构进行识别人类面部表情,由ResNet18 网络作为主干网络,输入的图片直接进入嵌入通道注意力机制的残差基础块中,通过加入Dropout策略,提高算法的鲁棒性,选择Cosine Decay学习率调整方式,提高模型精度.结果表明,文中所提出的模型在FER2013 数据集上取得了76.09%的准确率,与其他模型相比具有更好的识别效果.
Abstract
This paper proposes a facial expression recognition method combining channel attention mechanism and residual network,designed to improve the accuracy of facial expression recognition in nat-ural light.The method works by extracting expression classification and facial expression feature and blend them into an end-to-end deep convolutional neural network structur to achieve facial expression rec-ognition;using the ResNet18 network as the backbone network;embeding the input images come directly into the residual basis block of the channel attention mechanism;improving the robustness of the algo-rithm by adding Dropout strategy;and improving the model accuracy by choosing Cosine Decay as learn-ing rate adjustment method.The experimental results show that the model proposed in this paper achieves an accuracy of 76.09%on the FER2013 dataset,which has a better recognition effect than other models.
基金项目
黑龙江省省属本科高校基本科研业务费专项高层次培育项目(2022-KYYWF-0565)
黑龙江科技大学教学研究重点项目(2023)(JY23-40)
国家科技期刊平台
NSTL
万方数据