文件列表:

Output ScreenShots (0, 2017-06-13)
Output ScreenShots\2400 iteration 92.5% accuracy.png (951287, 2017-06-13)
Output ScreenShots\4000_iteration_95.5% accuracy.png (953157, 2017-06-13)
Output ScreenShots\Testing Adam_sandler_0.6137.png (960453, 2017-06-13)
Output ScreenShots\Testing Adam_sandler_0.9972.png (630966, 2017-06-13)
Output ScreenShots\Testing Will_smith_.5510.png (964683, 2017-06-13)
Output ScreenShots\Testing Will_smith_1.000.png (966096, 2017-06-13)
Output ScreenShots\Testing Zac_efron_0.8769.png (962524, 2017-06-13)
Output ScreenShots\Testing Zac_efron_0.9949.png (962442, 2017-06-13)
createTestDataset.py (507, 2017-06-13)
createlist.py (314, 2017-06-13)
faceDetection.py (1027, 2017-06-13)
face_recog_camera_feed.py (3149, 2017-06-13)
face_recog_single_image.py (1460, 2017-06-13)
solver.prototxt (284, 2017-06-13)
synset_FR.txt (1424, 2017-06-13)
test_840.txt (32589, 2017-06-13)
test_mean.binaryproto (786446, 2017-06-13)
train_8770.txt (357982, 2017-06-13)
train_mean.binaryproto (786446, 2017-06-13)
vgg_face_deploy.prototxt (4870, 2017-06-13)
vgg_face_pubfig.prototxt (5874, 2017-06-13)

# Face-Recognition-using-VGG_FaceNet Trained a VGG net for face recognition. ## Dataset Dataset has images of 84 individuals which includes faces of 83 celebrities and myself. Training set has 8770 images and the testing set has 840 images in total. Dataset Link: http://www.briancbecker.com/blog/research/pubfig83-lfw-dataset/ ## Model I have used VGG Net which includes 13 convolutional layers, 3 fully connected layers, and ReLu, Max-Pooling, Dropout layers in between. ## Training For training, I have used Transfer Learning to train the network and to achieve more accuracy in fewer iterations. In the training, I have used the pre-trained weights instead of initializing the weights randomly i.e trained caffemodel on the same network/model but on the different dataset. To use Transfer Learning, I have executed the following command from Caffe root folder to train the model. build/tools/caffe train --solver = face_recognition/solver.prototxt --weights = face_recognition/VGG_FACE.caffemodel Note: Below in References section, I have included the link to pre-trained caffemodel. ## Output/Accuracy After running for 4000 iterations I achieved the accuracy of 95.5% which is pretty good and very close to the accuracy of Facebook Face Recognition model which uses billions of images to train their network. "Output ScreenShots" folder includes few snapshots that show correct classifications done by the newly trained network. Below is the command I executed on new images of celebrities for classification using the newly trained model. build/examples/cpp_classification/classification.bin face_recognition/vgg_face_deploy.prototxt face_recognition/face_recog_iter_4000.caffemodel face_recognition/train_mean.binaryproto face_recognition/synset_FR.txt face_recognition/Testing_data/test_000082-000018.jpg ## Demo Videos Below are the links of videos that is demonstrating Face Detection, Face Tracking and Face Recognition of myself, Zac Efron and Will Smith. 1.) Nishank(Myself): https://drive.google.com/open?id=0B7zeI-3IdrKLY3dUYmhhQjZqZWc
2.) Zac Efron: https://drive.google.com/open?id=0B7zeI-3IdrKLTkoxc3EyckVpOG8
3.) Will Smith: https://drive.google.com/open?id=0B7zeI-3IdrKLQ2U1MnNLUDhsSXc ## References http://www.robots.ox.ac.uk/~vgg/software/vgg_face/ ## Deep Learning Platform Used Caffe

近期下载者：

相关文件：

评论：[我要评论] [举报此文件]

收藏者：