Face Recognition на Андроид

Face Recognition can be used as a test framework for several face recognition methods including the Neural Networks with TensorFlow and Caffe.

It includes following preprocessing algorithms:

- Grayscale

- Crop

- Eye Alignment

- Gamma Correction

- Difference of Gaussians

- Canny-Filter

- Local Binary Pattern

- Histogramm Equalization (can only be used if grayscale is used too)

- Resize

You can choose from the following feature extraction and classification methods:

- Eigenfaces with Nearest Neighbour

- Image Reshaping with Support Vector Machine

- TensorFlow with SVM or KNN

- Caffe with SVM or KNN

The manual can be found here https://github.com/Qualeams/Android-Face-Recognition-with-Deep-Learning/blob/master/USER%20MANUAL.md

At the moment only armeabi-v7a devices and upwards are supported.

For best experience in recognition mode rotate the device to left.

For best performance use "Image Reshaping with Support Vector Machine" (0.5 s / image).

For best accuracy use the "VGG Face Descriptor" model (the performance is very bad though - 6.5 s/ image)

_______________________________________________________________

TensorFlow:

If you want to use the Tensorflow Inception5h model, download it from here:

https://storage.googleapis.com/download.tensorflow.org/models/inception5h.zip

Then copy the file "tensorflow_inception_graph.pb" to "/sdcard/Pictures/facerecognition/data/TensorFlow"

Use these default settings for a start:

Number of classes: 1001 (not relevant as we don't use the last layer)

Input Size: 224

Image mean: 128

Output size: 1024

Input layer: input

Output layer: avgpool0

Model file: tensorflow_inception_graph.pb

---------------------------------------------------------------------------------------------------------

If you want to use the VGG Face Descriptor model, download it from here:

https://www.dropbox.com/s/51wi2la5e034wfv/vgg_faces.pb?dl=0

Caution: This model runs only on devices with at least 3 GB or RAM.

Then copy the file "vgg_faces.pb" to "/sdcard/Pictures/facerecognition/data/TensorFlow"

Use these default settings for a start:

Number of classes: 1000 (not relevant as we don't use the last layer)

Input Size: 224

Image mean: 128

Output size: 4096

Input layer: Placeholder

Output layer: fc7/fc7

Model file: vgg_faces.pb

_______________________________________________________________

Caffe:

If you want to use the VGG Face Descriptor model, download it from here:

http://www.robots.ox.ac.uk/~vgg/software/vgg_face/src/vgg_face_caffe.tar.gz

Caution: This model runs only on devices with at least 3 GB or RAM.

Then copy the files "VGG_FACE_deploy.prototxt" and "VGG_FACE.caffemodel" to "/sdcard/Pictures/facerecognition/data/caffe"

Use these default settings for a start:

Mean values: 104, 117, 123

Output layer: fc7

Model file: VGG_FACE_deploy.prototxt

Weights file: VGG_FACE.caffemodel

_______________________________________________________________

The license files can be found here https://github.com/Qualeams/Android-Face-Recognition-with-Deep-Learning/blob/master/LICENSE.txt and here https://github.com/Qualeams/Android-Face-Recognition-with-Deep-Learning/blob/master/NOTICE.txt</br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br></br>

- Switch from building Tensorflow from source to using the Jcenter library
- Included optimized_facenet model and changed default settings to use TensorFlow by default