Face detection using machine learning

Patent metadata

Field	Value
Publication number	US-10268950-B2
Application number	US-201414406927-A
Country	US
Kind code	B2
Filing date	Nov 15, 2014
Priority date	Nov 15, 2014
Publication date	Apr 23, 2019
Grant date	Apr 23, 2019

Abstract

A disclosed face detection system (and method) is based on a structure of a convolutional neural network (CNN). One aspect concerns a method for automatically training a CNN for face detection. The training is performed such that balanced number of face images and non-face images are used for training by deriving additional face images from the face images. The training is also performed by adaptively changing a number of trainings of a stage according to automatic stopping criteria. Another aspect concerns a system for performing image detection by integrating data at different scales (i.e., different image extents) for better use of data in each scale. The system may include CNNs automatically trained using the method disclosed herein.

First claim

What is claimed:

1. A method implemented on a computer system to automatically detect faces in images, the method comprising:
training a convolutional neural network (CNN) on face detection, the CNN including convolution, pooling and softmax functions, the training using a pool containing face images and non-face images, the training occurring in stages, the stages comprising:
formulating a training set for the stage, the training set having face images and non-face images derived from the images in the pool, the non-face images including non-face images that were false positives in earlier stages; and
training the CNN using the training set for that stage to optimize weights for the convolution, pooling, and softmax functions;
wherein training continues for additional stages until a number of new false positives for a stage falls below a threshold, and the new false positives were not also false positives in a previous stage;

incorporating the trained CNN in a face detection system; and
using the face detection system with trained CNN to detect faces in images.

2. The method of claim 1 wherein the training further comprises:
deriving synthetic face images by combining different face images from the pool;
the training set further including synthetic face images derived from the face images in the pool; wherein inclusion of the synthetic face images in the training set increases a total number of face images used in training.

3. The method of claim 2 wherein the stages further comprise:
formulating a validation set of face images and non-face images for the stage, wherein the training set is used to train the CNN during that stage and the validation set is used to calculate the validation cost for that stage, and the validation cost is used to determine whether to continue training for the stage.

4. The method of claim 2 wherein training for a stage continues until a validation cost for that stage falls below a threshold cost.

5. The method of claim 2 wherein training for a stage continues until a validation cost for that stage does not improve within a predetermined training duration.

6. The method of claim 5 wherein the predetermined duration is a predetermined number of trainings for the stage.

7. The method of claim 2 wherein training continues for additional stages until a number of total false positives for a stage falls below a threshold.

8. The method of claim 2 wherein, for each successive stage, a percentage of false positives from earlier stages is an increasing percentage of the non-face images in the training set.

9. The method of claim 2 wherein the training set for a stage includes all of the false positives from the immediately preceding stage.

10. The method of claim 2 wherein the training set for a stage includes false positives from only the immediately preceding stage.

11. The method of claim 2 wherein the training set for a stage includes false positives from multiple preceding stages.

12. The method of claim 2 further comprising:
deriving additional face images by altering face images in the pool, wherein the training sets for the stages further include the additional face images.

13. The method of claim 12 wherein the additional face images in the training sets are derived by rotating, blurring, mirroring or distorting face images in the pool.

14. The method of claim 2 wherein the computer system includes a GPU, and training the CNN is implemented on the GPU.

15. The method of claim 2 wherein the CNN is a bi-scale CNN having a first CNN and a second CNN that share weights.

Assignees

• Beijing Kuangshi Tech Co Ltd

Inventors

• Yin Qi
• Cao Zhimin
• Jia Kai

Classifications

• G06V10/764

  using classification, e.g. of video objects · CPC title

• G06N20/00

  Machine learning · CPC title

• G06N3/045

  Combinations of networks · CPC title

• G06F18/2431

  Multiple classes · CPC title

• G06V40/172 Primary

  Classification, e.g. identification · CPC title