Face Recognition and Classification With FaceNet On Android

Store images of people who you would like to recognize and the app, using these images, will classify those people. We don't need to modify the app/retrain any ML model to add more people ( subjects ) for classification

Message from the developer,

You may also like my latest project -> Age + Gender Estimation in Android with TensorFlow. I'm open for freelancing in Android + ML projects as well as Technical Blogging. You may send me an email/message on Google Chat at [email protected].

What's New

Updates - December 2021

• Users can now control the use of GpuDelegate and XNNPack using useGpu and useXNNPack in MainActivity.kt,

 // Use the device's GPU to perform faster computations.
 // Refer https://www.tensorflow.org/lite/performance/gpu
 private val useGpu = true

 // Use XNNPack to accelerate inference.
 // Refer https://blog.tensorflow.org/2020/07/accelerating-tensorflow-lite-xnnpack-integration.html
 private val useXNNPack = true

Major Updates - October 2021

• The app now has a face mask detection feature with models obtained from achen353/Face-Mask-Detector repo. You may off it by setting isMaskDetectionOn in FrameAnalyser.kt to false.

• The source of the FaceNet model is now Sefik Ilkin Serengil's DeepFace, a lightweight framework for face recognition and facial attribute analysis. Hence, the users can now use two models, FaceNet and FaceNet512. Also, the int-8 quantized versions of these models are also available. See the following line ine MainActivity.kt,

private val modelInfo = Models.FACENET

You may use different configurations in the Models class.

• The app will now classify users, whose images were not scanned from the images folder, as UNKNOWN. The app uses thresholds both for L2 norm and cosine similarity to achieve this functionality.

• For requesting the CAMERA permission and access to the images folder, the request code is now handled by the system itself. See Request app permissions.

Major Updates - July 2021

• We'll now use the PreviewView from Camera instead of directly using the TextureView. See the official Android documentation for PreviewView

• As of Android 10, apps couldn't access the root of the internal storage directly. So, we've implemented Scoped Storage, where the user has to allow the app to use the contents of a particular directory. In our case, users now have to choose the images/ directory manually. See Grant access to a directory's contents.

• The feature request #11 for serializing the image data has been considered now. The app won't load the images everytime so as to ensure a faster start.

• The feature request #6 has also been considered. After considering the use of PreviewView, the app can now be sed in the landscape orientation.

• The project is now backwards compatible to API level 25. For other details, see the build.gradle file.

• The lens facing has been changed to FRONT and users won't be able to change the lens facing. The app will open the front camera of the device as a default.

• The source of the FaceNet Keras model -> nyoki-mtl/keras-facenet

• The image normalization step is now included in the TFLite model itself using a custom layer. We only need to cast images to float32 using the CastOp from TFLite Support Library.

• A TextView is now added on the screen which logs important information like number of images scanned, similarity score for users, etc.

June 2021

• The source of the FaceNet model has been changed. We'll now use the FaceNet model from sirius-ai/MobileFaceNet_TF
• The project is now backwards compatible to API level 23 ( Android Marshmallow )

minSdkVersion 23

December 2020

• Lens Facing of the camera can be changed now. A button is provided on the main screen itself.
• For multiple images for a single user, we compute the score for each image. An average score is computed for each group. The group with the best score is chosen as the output. See FrameAnalyser.kt.

images ->  
    Rahul -> 
         image_rahul_1.png -> score=0.6 --- | average = 0.65 --- |
         image_rahul_2.png -> score=0.5 ----|                    | --- output -> "Rahul"
    Neeta ->                                                     |
         image_neeta_1.png -> score=0.4 --- | average = 0.35 --- |
         image_neeta_2.png -> score=0.3 ----|             

• Cosine similarity can be used alongside L2 norm. See the metricToBeUsed variable in FrameAnalyser.kt.
• A new parameter has been added in MainActivity.kt. The cropWithBBoxes argument allows you to run the Firebase MLKit module on the images provided. If you are already providing cropped images in the images/ folder, set this argument to false. On setting the value to true, Firebase ML Kit will crop faces from the images and then run the FaceNet model on it.

---

If you're ML developer, you might have heard about FaceNet, Google's state-of-the-art model for generating face embeddings. In this
project, we'll use the FaceNet model on Android and generate embeddings ( fixed size vectors ) which hold information of the face.

The accuracy of the face detection system ( with FaceNet ) may not have a considerable accuracy. Make sure you explore other options as well while considering your app's production.

The FaceNet Model

So, the aim of the FaceNet model is to generate a 128 dimensional vector of a given face. It takes in an 160 * 160 RGB image and
outputs an array with 128 elements. How is it going to help us in our face recognition project?
Well, the FaceNet model generates similar face vectors for similar faces. Here, by the term "similar", we mean
the vectors which point out in the same direction. In this app, we'll generate two such vectors and use a suitable metric to compare them ( either L2norm or cosine similarity ). The one which is the closest will form our desired output.

You can download the FaceNet Keras .h5 file from this repo and TFLite model from the assets folder.

Usage ( Intended file structure for the app )

So, an user can store images in his/her device in a specific folder. If, for instance, the user wants the app to recognize
two people namely "Rahul" and "Neeta". So the user needs to store the images by creating two directories namely "Rahul" and "Neeta"
and store their images inside of these directories. For instance, the file structure for the working example ( as shown above in the GIF ),

The app will then process these images and classify these people thereafter. For face recognition, Firebase MLKit is used which
fetches bounding boxes for all the faces present in the camera frame.

For better performance, we recommend developers to use more images of the subjects, they need to recognize.

Working

The app's working is described in the steps below:

1. Scan the images folder present in the internal storage. Next, parse all the images present within images folder and store
   the names of sub directories within images. For every image, collect bounding box coordinates ( as a Rect ) using MLKit. Crop the face from the image ( the one which was collected from user's storage ) using the bounding box coordinates.

2. Finally, we have a list of cropped Bitmap of the faces present in the images. Next, feed the cropped Bitmap to the FaceNet
   model and get the embeddings ( as FloatArray ). Now, we create a HashMap<String,FloatArray> object where we store the names of
   the sub directories as keys and the embeddings as their corresponding values.

See MainActivity.kt and FileReader.kt for the code.

The above procedure is carried out only on the app's startup. The steps below will execute on each camera frame.

1. Using androidx.camera.core.ImageAnalysis, we construct a FrameAnalyser class which processes the camera frames. Now, for a
   given frame, we first get the bounding box coordinates ( as a Rect ) of all the faces present in the frame. Crop the face from
   the frame using these boxes.
2. Feed the cropped faces to the FaceNet model to generate embeddings for them. We compare the embedding with a suitable metric and form clusters for each user. We compute the average score for each cluster. The cluster with the best score is our output. The final output is then stored as a Prediction and passed to the BoundingBoxOverlay which draws boxes and
   text.

See FaceNetModel.kt and FrameAnalyser.kt for the code.

Limitations

Predictions may go wrong as FaceNet does not always produce similar embeddings for the same person. Consider the accuracy of the FaceNet model while using it in your apps. In that case, you may learn to use the FaceNetModel class separating for using FaceNet in some other tasks.

Important Resources

• FaceNet: A Unified Embedding for Face Recognition and Clustering
• MLKit for face recognition.
• TensorFlow Lite Android
• TensorFlow Lite Android Support Library
• CameraX