戴沁璇 | AMiner学者主页

全面呈现 [戴沁璇] 的学术影响力

深度挖掘高质量成果与权威引用，让优秀看得见

研究兴趣

暂无数据

论文（6）

专利

基金项目

按年份排序按引用量排序主题筛选期刊级别筛选合作者筛选合作机构筛选

1Expression recognition method based on improved dense convolutional neural network

DAI Qinxuan,LUO Xiaoshu,MENG Zhiming,HUANG Yuanqin

人的表情包含大量信息,可用于显示人的很多情感状态,例如疲劳和疼痛的表情等.卷积神经网络(CNN)是一种识别人脸表情的有效方法,它可以同时执行特征提取和分类,并可以自动发现数据中的多个表情特点.针对卷积神经网络参数大以及传统表情识别方法准确率不高的问题,提出一种基于改进的稠密卷积神经网络的面部表情识别模型.首先通过使用Gabor滤波器初始化第一层卷积层;然后采用一种对数线性函数(LLU)进行网络优化,该模型中的特征重用和参数压缩技术提高了网络的学习能力,大大减少了模型参数;最后基于此模型设计了一个表情识别系统,该系统能够准确地识别照片上的表情和在线识别人脸表情.实验结果表明,该模型可以显著提高三个表情数据集的准确率,并能很好地识别人脸表情.

2022Modern Electronics Technique（2022）

引用

AI阅读

2Research on Face Super-Resolution Reconstruction Algorithm Based on Generative Adversarial Networks

罗晓曙,蒋文杰,戴沁璇

由于受到光照和成像设备等条件因素的影响,采集到的单帧人脸图像分辨率低,无法进行准确人脸识别,所以需要图像超分辨率重建.而利用SRGAN模型在进行人脸超分辨率重建过程中,易出现梯度消失或爆炸的问题,严重影响了重建图像的精度和质量.针对上述问题,提出了基于生成对抗网络的改进人脸超分辨率重建算法,在SRGAN结合WGA-N的基础上引入Wasserstein散度,并将其最大化得到最优化标量函数T,实现去掉Lipschit-z约束能够直接得到Wassertein距离,并通过最小化Wassertein距离得到生成网络的目标函数,最终改进的模型提高了重建图像的质量.实验结果表明,该方法能够生成高分辨率的人脸图像,在主观视觉和客观评价指标均同比优于DRCN、FSRCNN、SRGAN_WGAN、VDSR和DRRN模型.

2021Computer Engineering and Applications（2021）引用：5

引用

AI阅读

3LS-FIR Filter Based on Least Square Method

武康康,董翠,陆叶, 周闰昌,戴沁璇,周鹏,朱旭飞

本文提出一种基于最小二乘法的LS-FIR滤波器.根据色散补偿滤波器与有限长单位冲激响应的频域函数之间的关系构造线性方程组,线性方程组中独立方程的个数大于未知量个数,所以此方程组无解.但在最小范数意义下,能够找到一组解使得估计误差模的平方和最小,得到的解是最小二乘解.用所得到的最小二乘解作为滤波器的抽头权值,即是所要设计的滤波器.滤波器仿真结果表明:不同调制格式(QPSK、16QAM、64QAM)和信噪比(1～20)条件下,当全频带且滤波器的抽头数为315时,均方误差稳定于2.6109×10-6;当窄频带滤波器的抽头数为197时,均方误差稳定于2.5566×10-6;与FIR和Hm-FIR滤波器对比,LS-FIR滤波器算法具有较高的稳定度、较好的逼近效果,且滤波效果更佳.

2021Journal of Guangxi Normal University(Natural Science Edition)（2021）引用：1

引用

AI阅读

4Research on an Improved Method of Generative Adversarial Networks Image Coloring

Wen-jie JIANG,Xiao-shu LUO,Qin-xuan DAI

利用手持云台对黑白照/图像进行拍摄再对其进行上色是耗时耗力的工作,为了提高对黑白照片/图像上色的效率和视觉效果,利用深度学习的卷积神经网络提出了一种基于改进生成对抗网络的上色算法.采用了原模型pix2pix(image-to-image translation with conditional generative adversarial networks)的U型结构的生成器并在其中引入了自注意力机制提高输出图像的色彩多样性;其次是使用实例归一化对网络结构作进一步的优化处理,生成器和判别器在网络训练期间不断进行相互对抗学习,同时模型不断学习并优化黑白照片/图像到对应彩色图像的映射关系;最后实现了对黑白照片/图像的自动化上色,同时使用客观量化指标(MSE、PSNR和SSIM)对实验结果进行评价.实验结果表明:该算法能快速有效实现对黑白照片/图像的无监督上色,同时保持良好的视觉效果.

2020Computer Technology and Development（2020）引用：5

引用

AI阅读

5Research on Face Recognition Based on Gabor-LeNet Convolutional Neural Network Model

Qinxuan Dai,Xiaoshu Luo,Zhiming Meng

Aiming at the problems of slowing convergence speed, and low recognition accuracy in the face recognition training of the existing LeNet-5 convolutional neural network, an improved LeNet-5 convolutional neural network model is proposed and applied for face recognition. The main improvement is the use of Gabor filter to initialize the first convolutional layer and the activation function of Parametric Rectified Linear Unit (PReLU). The recognition accuracy of the improved model on ORL and GT face datasets has reached more than 98%. At the same time, the recognition accuracy rate on the face data set AR with occlusion has reached more than 90%, indicating that the improved model has strong robustness.

2020Journal of Physics: Conference Series（2020）引用：2

引用

AI阅读

6Research on Super-resolution Reconstruction Algorithm of Remote Sensing Image Based on Generative Adversarial Networks

JIANG Wenjie,LUO Xiaoshu,DAI Qinxuan

Due to natural conditions and hardware manufacturing processes, the resolution of remote sensing images is generally low. Obtaining high-definition remote sensing images by simply improving hardware and manufacturing processes is not only costly and technically challenging but also cannot be deployed on a large scale. Aiming at the limitations of the traditional methods, this paper studies the image super-resolution reconstruction method for improving the generated anti-network. Firstly, the generator network is optimized, and an RRDB (Residual-in-Residual Dense without BN (Batch Normalization) is used. Block) module; secondly, the related idea of relativistic GAN (relativistic generative adversarial network) is introduced, the relative value of the discriminator is not the absolute value; finally, the sensation loss is improved, and the feature is used before the function is activated. The test results show that the proposed algorithm is better than SRGAN (Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network), SRCNN (Super-Resolution Convolutional Neural Network) and FSRCNN (Fast Super-Resolution Convolutional Neural Network). The clarity of the reconstructed image is improved, and the reconstructed image quality is significantly improved.

2020Computer Engineering and Applications（2020）引用：5

引用

AI阅读