基于Siamese网络的人脸识别 2019-3-9
- 将CAS-PEAL-R1数据集放入datasets文件夹
- 建立FaceRecognition的类实例
- 调用prepare方法进行数据处理
+-<CAS-PEAL-R1> ... The root directory of this CDROM.
|
|
+-<Documents> ... The documentation directory. In this directory the pdf version of
| the technical report of CAS-PEAL database and the CAS-PEAL
| Database Release Agreement can be found.
|
+-<Evaluation Prototype> ... This directory contains the evaluation prototype of the
| | CAS-PEAL-R1 database.
| |
| +-<Gallery> ... This directory contains one text file which records the image names
| | in the Galley.
| |
| |
| +-<Probe Sets> ... This directory contains seven text files. Each text file records
| | the image names of a specific probe set.
| |
| |
| +-<Training Set> ... This directory contains one text file which records the image names
| in the Training set.
|
|
+-<FRONTAL> ... Frontal Subset of CAS-PEAL-R1
| |
| +-<Accessory> ... images with accessory variations
| |
| +-<Background> ... images with background variations
| |
| +-<Distance> ... images with distance variations
| |
| +-<Expression> ... images with expression variations
| |
| +-<Lighting> ... images with lighting variations
| |
| +-<Aging> ... images with aging variations
| |
| +-<Normal> ... images under normal conditions
|
+-<POSE> ... Pose Subset of CAS-PEAL-R1.
|
+-<000001> ... All the pose images of one person are stored in one directory
|
+-<000002> ... Each directory's name is one person's ID.
|
...
|
+-<001042>
- 建立FaceRecognition的类实例
- 调用train方法进行训练
- 若载入权重，将load_pretrained置为True
- 建立FaceRecognition的类实例
- 调用predict方法进行预测，输入为图片文件路径，输出1为同一人，输出0为不同人
- 训练好的权重下载地址
- 链接：https://pan.baidu.com/s/1YinGcQ8_4FKRKrIo0fd7Xw 提取码：7iog
- 该权重在数据集上正确率为98.72%
- 注意：预测图片尽量与CAS-PEAL-R1图片的size类似（360*480）
- 借鉴ResNet50
- 采用多个小卷积核级联的形式提高感受野
- 使用inception block形式将不同感受野结合
- 采用leaky relu加快训练
- 使用L1距离加上logistic回归单元进行预测
