【作者】 胡敏；

【导师】 郭克华；

【作者基本信息】 中南大学 ， 计算机技术（专业学位）， 2022， 硕士

【摘要】 人脸图像是公共安全和刑事调查领域的关键数据基础。然而,由于光线和拍摄角度等原因,在低光场景下采集的人脸往往目标较小、亮度较低、分辨率较低,导致人脸关键特征难以识别,无法为公共安全和刑事调查等领域提供数据支撑。本文针对小目标人脸图像亮度低、分辨率低两个问题,提出了两种研究方法。(1)针对低光照场景下小目标人脸难以辨识的问题,本文提出了一种面向低光照场景的小目标人脸图像亮度增强方法。首先,通过多级特征提取模块获取不同层级的丰富人脸图像特征,以尽可能挖掘出隐藏在暗处的人脸图像信息。其次,设计了亮度增强模块,通过引入自注意力机制,学习捕获低光照人脸图像中的远距离的像素依赖关系并辅助增强人脸图像中的亮度细节,使增强之后的图像更加友好。最后,通过特征融合模块集成所有亮度增强的特征图,并使用卷积层进行特征图融合。实验结果表明,与其他方法相比,本文提出的方法能更有效的提升低光场景下小目标人脸图像的亮度。(2)针对低光照场景下小目标人脸图像分辨率低、细节缺失的问题,本文在人脸图像增强的基础上,提出一种基于谱归一化的人脸超分辨率方法,通过渐进式人脸生成的方法,实现超低分辨率人脸图像的高倍数重建。首先,设计了低分辨率图像编码模块,充分提取亮度增强的低分辨率人脸图像的特征信息。其次,通过引入通道注意力机制并设计一系列带噪声的样式块,重建出视觉友好的清晰且逼真的高质量人脸。最后,在鉴别器中引入频谱归一化、自注意力机制和两时间尺度更新规则,提高模型训练稳定性。实验表明,与现有方法相比,本文的方法在主观视觉感知和客观指标上均产生了更优越的结果。图23幅,表6个,参考文献66篇更多还原

【Abstract】 Face images have always been a key data foundation in security and criminal investigations.However,due to lighting and shooting angles,faces captured in low-light scenes often suffer from low resolution and low luminance,making it difficult to identify facial features and unable to provide data support for related tasks.In this paper,we propose a small target low-light face image enhancement method and a face superresolution method based on spectral normalization for improving low brightness and low resolution of target face images.(1)For the problem that small target faces are difficult to recognize in low-light scenes,this paper proposes a small target face image light enhancement method for low-light scenes.Firstly,the multi-level feature extraction module is used to obtain rich face image features at different levels to dig out the face image information hidden in the dark as much as possible.Secondly,a self-attention mechanism is introduced in the luminance enhancement module to learn to capture pixel relationships at a distance in low-light face images and make the enhanced image more friendly.Finally,the feature fusion module integrates the luminanceenhanced feature maps and uses convolutional layers for fusion.Experiments prove that the proposed method can enhance the brightness of small target face images in low-light scenes more effectively compared with existing methods.(2)For small target luminance-enhanced face images with low resolution and missing details,this paper proposes a face super-resolution method based on spectral normalization to achieve a large upscaling factor of ultra-low resolution face images by a progressive face generation method.Firstly,a low-resolution image encoder is designed to fully extract the feature information of low-resolution face images with enhanced brightness.Secondly,we introduce a channel attention mechanism and design a series of noisy style blocks to reconstruct visually friendly and realistic high-quality faces.Finally,to improve the stability of model training while speeding up the model training,the spectral normalization,self-attentiveness mechanism,and two-time scale update rule are introduced in the discriminator.Experiments show that the proposed method yields superior results in terms of both subjective visual results and evaluation metrics compared with existing methods.更多还原