Please read the full article on LinkedIn or Medium. The github repo provides a docker image with necessary ML base models and python utilities for face recognition, with easy to use command-line api. Each tool has a corresponding jupyter notebook so you can test with your images, tune parameters, or just visualize results.
- Install Docker on your local machine
- Download docker image from docker hub using
docker pull samirsdoshi/faceml:latestOR - Build image locally by running build-image.sh and copy large model files
a) Download a converted darknet YOLO model to keras here and copy to yolo_keras/yolo.h5. This is used in detectobjects.py for object detection
b) Download keras-facenet from here and copy under keras-facenet. This is used for face detection in train_keras.py and detectface_keras.py
c) Download a Torch deep learning model which produces the 128-D facial embeddings here file and save as opencv/openface_nn4.small2.v1.t7.
d) Also download a pre-trained Caffe deep learning model to detect faces here and save as opencv/res10_300x300_ssd_iter_140000.caffemodel.
- Setup a folder with images. for. e.g
|--training
|-person1
|-person2
|-person3
...
|-validation
|-person1
|-person2
|-person3
...
|-allimages
image1.jpg
image2.jpg
...
|-outdir
|-logdir
- Change the volume line in faceml.sh to mount your directory with images. Run faceml.sh to launch docker container. The container also runs jupyter notebook. It will wait for few seconds for the container to start. You can access the jupyter notebook from your browser using http://localhost:8888/?token=token
- Run
docker exec -it faceml bashto enter docker container
- To detect various types of objects and move images with objects in a seperate folder
root@afc13fdc17ef:/faceml# python detectobjects.py --help
Using TensorFlow backend.
usage: detectobjects.py [-h] -i IMAGEDIR -c CLASS [-k [CONFIDENCE]]
[-s [SIZE]] [-n [COUNT]] [-p [PORTRAIT]] [-g [GROUP]]
[-o OUTDIR] [-l LOGDIR] [-v LOG_LEVEL]
optional arguments:
-h, --help show this help message and exit
-i IMAGEDIR, --imagedir IMAGEDIR
path to input directory of images
-c CLASS, --class CLASS
object class expression to search for as per
yolo_keras/coco_classes.txt
-k [CONFIDENCE], --confidence [CONFIDENCE]
minimum confidence percentage for object detection.
Default 80
-s [SIZE], --size [SIZE]
minimum percentage size of the object in the image.
-n [COUNT], --count [COUNT]
select images containing n count of class object
-p [PORTRAIT], --portrait [PORTRAIT]
portrait mode. select images only if smallest of the n
count objects is bigger by p percentage over next
biggest object
-g [GROUP], --group [GROUP]
group mode. select images only if n count objects are
within g percentage
-o OUTDIR, --outdir OUTDIR
path to output directory where selected images will be
moved
-l LOGDIR, --logdir LOGDIR
path to log directory
-v LOG_LEVEL, --loglevel LOG_LEVEL
log level: DEBUG, INFO, ERROR. Default INFO
Class expression supports expressions like
- person
- not [person]
- [person] and [car]
- [person] and ([car] or [boat])
- [person] and [flower] and not [bird]
- ([person] and [car]) or ([flower] and [bird]) etc..
Also you can find potrait photos of a single person and use it for training the model. You can set big enough size and count to 1 to isolate images of single person, even if there are few other people in the background.
- To detect similar images and group them under a folder, so that those can be reviewed and sorted appropriately.
root@afc13fdc17ef:/faceml# python detectsimilar.py --help
Using TensorFlow backend.
usage: detectsimilar.py [-h] -d IMAGEDIR [-c CLASS] [-t [THRESHOLD]]
[-o OUTDIR] [-l LOGDIR] [-v LOG_LEVEL]
optional arguments:
-h, --help show this help message and exit
-d IMAGEDIR, --imagedir IMAGEDIR
path to input directory of images
-c CLASS, --class CLASS
object class to filter as per
yolo_keras/coco_classes.txt
-t [THRESHOLD], --threshold [THRESHOLD]
minimum confidence percentage for similarity. Default
70
-o OUTDIR, --outdir OUTDIR
path to output directory where selected images will be
moved
-l LOGDIR, --logdir LOGDIR
path to log directory
-v LOG_LEVEL, --loglevel LOG_LEVEL
log level: DEBUG, INFO, ERROR. Default INFO
- To extract faces from images and save as seperate files (could be used as inputs for training)
root@afc13fdc17ef:/faceml# python extract_faces.py --help
Using TensorFlow backend.
usage: extract_faces.py [-h] -i IMAGESDIR -o OUTDIR [-p [MARGIN]] [-l LOGDIR]
[-v LOG_LEVEL]
optional arguments:
-h, --help show this help message and exit
-i IMAGESDIR, --imagesdir IMAGESDIR
path to input directory of images
-o OUTDIR, --outdir OUTDIR
path to output directory to store face images
-p [MARGIN], --margin [MARGIN]
margin percentage pixels to include around the face
-l LOGDIR, --logdir LOGDIR
path to log directory
-v LOG_LEVEL, --loglevel LOG_LEVEL
log level: DEBUG, INFO, ERROR. Default INFO
See example
- To train model (using Keras/Facenet)
root@7f87ba121d08:/faceml# python train_keras.py --help
Using TensorFlow backend.
usage: train_keras.py [-h] -t TRAINDIR [-p [MARGIN]] -v VALDIR -o OUTDIR
[-l LOGDIR] [-g LOG_LEVEL]
optional arguments:
-h, --help show this help message and exit
-t TRAINDIR, --traindir TRAINDIR
path to input directory of images for training
-p [MARGIN], --margin [MARGIN]
margin percentage pixels to include around the face
-v VALDIR, --valdir VALDIR
path to input directory of images for training
-o OUTDIR, --outdir OUTDIR
path to output directory to store trained model files
-l LOGDIR, --logdir LOGDIR
path to log directory
-g LOG_LEVEL, --loglevel LOG_LEVEL
log level: DEBUG, INFO, ERROR. Default INFO
- To detect faces and move files (using Keras/Facenet)
root@87b636b3469e:/faceml# python detectface_keras.py --help
Using TensorFlow backend.
usage: detectface_keras.py [-h] -i IMAGESDIR -m MODELPATH -c PERSON
[-p [MARGIN]] [-k [CONFIDENCE]] [-o OUTDIR]
[-n NOMATCHDIR] [-l LOGDIR] [-v LOG_LEVEL]
optional arguments:
-h, --help show this help message and exit
-i IMAGESDIR, --imagesdir IMAGESDIR
path to input directory of images
-m MODELPATH, --modelpath MODELPATH
directory with trained model
-c PERSON, --person PERSON
person name expression to filter - <name1> | not
[name1] | [name1] and [name2] | [name1] and ([name2]
or [name3]) and not [name4] etc
-p [MARGIN], --margin [MARGIN]
margin percentage pixels to include around the face
-k [CONFIDENCE], --confidence [CONFIDENCE]
minimum confidence percentage for face recognition.
Default 70
-o OUTDIR, --outdir OUTDIR
path to output directory to store images having filter
class
-n NOMATCHDIR, --nomatchdir NOMATCHDIR
directory to move image to, if not matched.
-l LOGDIR, --logdir LOGDIR
path to log directory
-v LOG_LEVEL, --loglevel LOG_LEVEL
log level: DEBUG, INFO, ERROR. Default INFO
- To train model (using OpenCV/Caffe)
root@7f87ba121d08:/faceml# python train_cv2.py --help
usage: train_cv2.py [-h] -t TRAINDIR -v VALDIR [-p [MARGIN]] -o OUTDIR
[-l LOGDIR] [-g LOG_LEVEL]
optional arguments:
-h, --help show this help message and exit
-t TRAINDIR, --traindir TRAINDIR
path to input directory of images for training
-v VALDIR, --valdir VALDIR
path to input directory of images for training
-p [MARGIN], --margin [MARGIN]
margin percentage pixels to include around the face
-o OUTDIR, --outdir OUTDIR
path to output directory to store trained model files
-l LOGDIR, --logdir LOGDIR
path to log directory
-g LOG_LEVEL, --loglevel LOG_LEVEL
log level: DEBUG, INFO, ERROR. Default INFO
- To detect faces and move files (using OpenCV/Caffe)
usage: detectface_cv2.py [-h] -i IMAGESDIR -m MODELPATH -c PERSON
[-p [MARGIN]] [-k [CONFIDENCE]] [-o OUTDIR]
[-n NOMATCHDIR] [-l LOGDIR] [-v LOG_LEVEL]
optional arguments:
-h, --help show this help message and exit
-i IMAGESDIR, --imagesdir IMAGESDIR
path to input directory of images
-m MODELPATH, --modelpath MODELPATH
directory with trained model
-c PERSON, --person PERSON
person name expression to filter - <name1> | not
[name1] | [name1] and [name2] | [name1] and ([name2]
or [name3]) and not [name4] etc
-p [MARGIN], --margin [MARGIN]
margin percentage pixels to include around the face
-k [CONFIDENCE], --confidence [CONFIDENCE]
minimum confidence percentage for face recognition.
Default 70
-o OUTDIR, --outdir OUTDIR
path to output directory to store images having filter
class
-n NOMATCHDIR, --nomatchdir NOMATCHDIR
directory to move image to, if not matched.
-l LOGDIR, --logdir LOGDIR
path to log directory
-v LOG_LEVEL, --loglevel LOG_LEVEL
log level: DEBUG, INFO, ERROR. Default INFO
- Utility to sort images into folders on year/month/day (hierachical or flat directories)
root@7f87ba121d08:/faceml# python sortimages.py --help
usage: sortimages.py [-h] -i IMAGEDIR -o OUTDIR -s SORTMODE -f FOLDERMODE
[-l LOGDIR] [-v LOG_LEVEL]
optional arguments:
-h, --help show this help message and exit
-i IMAGEDIR, --imagedir IMAGEDIR
path to input directory of images for sorting
-o OUTDIR, --outdir OUTDIR
path to output directory to store sorted files
-s SORTMODE, --sortmode SORTMODE
sort mode (y,m,d,ym,ymd). Create folders by y (year),
m (month), d (day), ym, ymd
-f FOLDERMODE, --foldermode FOLDERMODE
h (hierarchical) or f (flat)
-l LOGDIR, --logdir LOGDIR
path to log directory
-v LOG_LEVEL, --loglevel LOG_LEVEL
log level: DEBUG, INFO, ERROR. Default INFO
I have the combined training/detection code as jupyter notebooks for both methods. Review the end of both files below for results.
https://machinelearningmastery.com/how-to-develop-a-face-recognition-system-using-facenet-in-keras-and-an-svm-classifier/
https://www.pyimagesearch.com/2018/02/26/face-detection-with-opencv-and-deep-learning/
https://pjreddie.com/darknet/yolo/
https://mc.ai/face-detection-with-opencv-and-deep-learning/
https://sb-nj-wp-prod-1.pyimagesearch.com/2018/09/24/opencv-face-recognition/
http://cmusatyalab.github.io/openface/
https://cmusatyalab.github.io/openface/models-and-accuracies/