论文 Low Resolution Face Recognition Using a Two-Branch Deep Convolutional Neural Network Architecture

Introduction

Three general approaches for low resolution face recognition
对于待识别人脸图像为低像素 LR probe image,训练数据为高像素 HR gallary image 的情况,常见的处理方式:

  1. 训练数据也 down-sampling 到低像素。丢失有用信息。
  2. 从低像素的目标人脸图像生成高像素数据用于识别。对应方法关注点往往在于生成图像的质量,而非人脸识别的性能。[10]-[13]
  3. 将 LR probe image 和 HR gallary image 同时转换到一个共同空间,使两者距离接近。[14]-[17] 这篇论文采取的方式。

此论文的关键在于找到 nonlinear transformation from LR and HR to common space。 Two deep CNN。
由于包含一个 super-resolution CNN,同时还可以生成 LR 对应的 HR 图像。
Object Function 是 Distance of transformed low and high resolution images in the common space。
Dataset 是 FERET。
占用内存小。

Method

Architecture of two deep convolutional neural networks in two branches
Training set: pairs of LR and HR of same person on different conditions.

Networks Architecture

VGGnet: 13 CONV + 3 FC

  • HR images -> common space: FECNN(feature extraction onvolutional neural network). 224x224 image -> 4096 feature vector. VGGnet - 2 FC
  • LR images -> common space: SRFECNN = SRnet(super-resolution net) + FECNN. 224x224 -> 4096 feature vector

由于去掉两层 FC,比 VGGnet parameter 少,可以放进内存。
SRFECNN weights

Common Subspace Learning

3 steps:

  1. Use trained VGGnet on face dataset and then dropped the last two FC. 因为这两层是为 classification task 特别起作用的。称为 pre-trained FECNN。
  2. Train the SRnet of the bottom branch with a dataset of high and low resolution face image pairs.
  3. Merge SRnet and FECNN and a training dataset that contains pairs of LR and HR of same persons was fed into the brached.

HR 所在上层固定,只训练 LR 所在下层的 FECNN 和 SRnet。
Distance between LR and HR images of the same subjects is the error, backpropagated into the bottom branch net (both FECNN and SRnet).

Configurations with different super-resolution modules

Reconstruct Input Image

SRnet 输出。SRnet 的主要作用看上去是让极低分辨率的图片不至于表现太坏。
生成图片的效果并不是很好,为 better recognition performance 做出了牺牲。
Reconstructed Faces by different configurations

Datasets

Datasets Used

Ref

https://arxiv.org/abs/1706.06247