Deep Learning in Practice III: Face Recognition

Explore how facial recognition differs from traditional classification and detection, and learn how convolutional networks enable one-shot classification, ACMD network, face encodings with triplet loss, and evaluating similarity between faces.

Demonstrates face recognition access demo where a web app captures an image, sends it to the cloud for authentication, and signals door opening, highlighting touchless biometric authentication used in airports.

Differentiate face verification from face recognition by comparing a new image to a single reference versus searching a database for the best match, as mobile phones verify you.

The lecture contrasts traditional classification with face recognition, highlighting how changing object counts and limited images force retraining and changes to output layers, limiting practicality.

Discover one-shot learning, where recognition comes from comparing features and similarities between images rather than learning the object from many examples, enabling face verification and recognition.

One-shot learning example for face recognition shows using embeddings and distance thresholds to identify a new face, avoiding retraining a softmax classifier for each person.

Train face embeddings using triplet loss with anchor, positive, and negative images, minimizing anchor-positive distance while maximizing anchor-negative distance, controlled by an alpha hyperparameter.

Generate a triplet dataset of facial images to train embeddings for one-shot face recognition. Use anchor, positive, and negative samples, and hard-negative mining to improve embedding discrimination.

Explore how a siamese network trained with triplet loss uses anchors, positives, and negatives to shape embeddings that cluster same-person faces while separating different faces by distance.

Explain triplet loss on image embeddings, separating anchor, positive, and negative samples with a margin to distinguish faces; leverage pre-treated recognition models to generate embeddings and perform distance-based clustering.

Learn binary face classification with a siamese network that computes embeddings and outputs a probability via a sigmoid logistic regression to decide whether two images show the same person.

Provide an overview of face detection and recognition, showing how a system detects faces with bounding boxes, creates embeddings, and identifies individuals by distance against a database.

Create a face embedding and save it to a file in Google Colab by installing empty CNN libraries, mounting Google Drive, and configuring project and system paths for face recognition.

Extract faces from images using the MTCNN Python library, which implements joint face detection and alignment via a multitask cascading convolutional network, and create a detector instance with default parameters.

Extract faces from an image using the detector, returning metadata with bounding box coordinates, confidence, and key points for left eye, right eye, nose, left mouth, and right mouth.

Read the four values of the face bounding box as x, y, width, and height to draw the box around the detected face.

Analyze the first detected face by extracting the bounding box coordinates from the face metadata, draw a rectangle around it with OpenCV, and visualize the detected face.

Select the second detected face by setting phase index to one, apply a rectangle to its coordinates, and visualize the result, noting false positives or negatives in deep learning.

Apply a convolutional neural network face recognition model to generate a face embedding from the detected face, using a public model with options for 128 or 512 dimensions.

Load FaceNet with the TensorFlow Keras library, load the model from the H5 file, and convert the input image into an embedding with the neural network.

Create a face embedding with FaceNet by converting image pixels to float32, normalizing by mean and standard deviation, flattening to a 1D vector, and obtaining the embedding from the model.

Load the face embedding from a file using the load method, obtaining the embedding and its label from the data array. Confirm the embedding matches the stored vector.

Identify a face in an image by comparing its embedding to a saved face database. Extract faces, generate embeddings with Fastnet, and find the closest match by minimum distance.

Use face embedding to identify a person by comparing embeddings with a vector distance; the smallest distance reveals the closest match, a footprint-like representation from the ACMD network.