一、二分类模型
1.（yu）数据集：使用nuaa数据集的 crop的人脸图片；
网络结构：mobilenet；
训练结果：收敛至val acc 0.85左右
实时视频测试：mtcnn 检测人脸，效果很不好，完全没有区分能力
2.使用nuaa+replayattack数据集（train和val），mtcnn统一预处理
网络结构：mobilenet、inception Resnet、unet接二分类输出层训练模型
loss函数：categorical_crossentropy
训练方式：all layers trainable
优化器：adam
训练结果：
1）截图测试（1000张）：
inception Resnet: total:1082; correct:737; percent:0.681146
mobileNet: total:1082; correct:555; percent:0.512939
unet: total:1082; correct:550; percent:0.508318
3.使用nuaa数据集（train、val）
训练结果：
1）replayattack数据集（val）测试：
inceptionResnet: total:10503; correct:4163; percent:0.396363
mobileNet: total:10503; correct:4654; percent:0.443111
unet: total:10503; correct:4304; percent:0.409788
2）截图测试（1000张）：
inceptionResnet: total:1082; correct:338; percent:0.312384
mobileNet: total:1082; correct:345; percent:0.318854
unet: total:1082; correct:174; percent:0.160813
4.使用replayattack数据集（train、val和test）
训练结果：
1）test 数据集测试：
inceptionResnet: total:12198; correct:11111; percent:0.910887
mobileNet: total:12198; correct:12001; percent:0.983850
unet: total:12198; correct:11745; percent:0.962863
gram_vgg16: total:12198; correct:11991; percent:0.983030
gram_vgg16_mobilenet:total:12198; correct:7965; percent:0.652976
gram_vgg16_mobilenet_normalization：total:12198; correct:11900; percent:0.975570
gram_vgg16_mobilenet_normalization_2：total:12198; correct:12073; percent:0.989752
gram_model: total:12198; correct:12178; percent:0.998360
gram_model_2:total:12198; correct:12001; percent:0.983850
2）nuaa数据集（val）测试：
inceptionResnet:total:2193; correct:1146; percent:0.522572
mobileNet: total:2193; correct:1168; percent:0.532604
unet: total:2193; correct:1065; percent:0.485636
gram_vgg16: total:2193; correct:1143; percent:0.521204
gram_vgg16_mobilenet:total:2193; correct:1179; percent:0.537620
gram_vgg16_mobilenet_normalization：total:2193; correct:1119; percent:0.510260
gram_vgg16_mobilenet_normalization_2：total:2193; correct:1114; percent:0.507980
gram_model: total:2193; correct:1088; percent:0.496124
gram_model_2:total:2193; correct:1053; percent:0.480164
3）截图测试（1000张）：
inceptionResnet:total:1082; correct:592; percent:0.547135
mobileNet:total:1082; correct:734; percent:0.678373
unet:total:1082; correct:675; percent:0.623845
gram_vgg16: total:1082; correct:672; percent:0.621072
gram_vgg16_mobilenet: total:1082; correct:745; percent:0.688540
gram_vgg16_mobilenet_normalization：total:1082; correct:636; percent:0.587800
gram_vgg16_mobilenet_normalization_2：total:1082; correct:730; percent:0.674677
gram_model: total:1082; correct:706; percent:0.652495
gram_model_2:total:1082; correct:637; percent:0.588725
二、FDA
数据集：replayattack +FDA预处理
网络结构：mobilenet
训练结果：
1）test 数据集测试：
total:12198; correct:10408; percent:0.853255
2）nuaa数据集（val）测试：
total:2193; correct:930; percent:0.424077
3）截图测试（1000张）：
total:1082; correct:553; percent:0.511091
结论：较未使用FDA的效果反而更差了
三、统计