BitLaw

Example 39: Method for Training a Neural Network for Facial Detection

Background

Facial detection is a computer technology for identifying human faces in digital images. This technology has several different potential uses, ranging from tagging pictures in social networking sites to security access control. Some prior methods use neural networks to perform facial detection. A neural network is a framework of machine learning algorithms that work together to classify inputs based on a previous training process. In facial detection, a neural network classifies images as either containing a human face or not, based upon the model being previously trained on a set of facial and non-facial images. However, these prior methods suffer from the inability to robustly detect human faces in images where there are shifts, distortions, and variations in scale and rotation of the face pattern in the image.

Applicant’s invention addresses this issue by using a combination of features to more robustly detect human faces. The first feature is the use of an expanded training set of facial images to train the neural network. This expanded training set is developed by applying mathematical transformation functions on an acquired set of facial images. These transformations can include affine transformations, for example, rotating, shifting, or mirroring or filtering transformations, for example, smoothing or contrast reduction. The neural networks are then trained with this expanded training set using stochastic learning with backpropagation which is a type of machine learning algorithm that uses the gradient of a mathematical loss function to adjust the weights of the network. Unfortunately, the introduction of an expanded training set increases false positives when classifying non-facial images. Accordingly, the second feature of applicant’s invention is the minimization of these false positives by performing an iterative training algorithm, in which the system is retrained with an updated training set containing the false positives produced after face detection has been performed on a set of non-facial images. This combination of features provides a robust face detection model that can detect faces in distorted images while limiting the number of false positives.

Claim:

A computer-implemented method of training a neural network for facial detection comprising:

• collecting a set of digital facial images from a database;
• applying one or more transformations to each digital facial image including mirroring, rotating, smoothing, or contrast reduction to create a modified set of digital facial images;
• creating a first training set comprising the collected set of digital facial images, the modified set of digital facial images, and a set of digital non-facial images;
• training the neural network in a first stage using the first training set;
• creating a second training set for a second stage of training comprising the first training set and digital non-facial images that are incorrectly detected as facial images after the first stage of training; and
• training the neural network in a second stage using the second training set.