Research Efforts on Face Detection

In the early stage of our research, we mainly studied the Eigen-features, skin model, rule-based method. Later on, we have focused on statistical learning-based method. Especially, we have investigated the GAM-based training set data re-sampling problem, the spatial histogram-based object detection method, and ECOC-based multi-class classification method.

  • Re-sampling for for face detection

Problem Description

Though the design of a classifier is important, the performance of the trained detector is highly depended on the distribution of the training set. In general, the larger the training set is, the better generalization of the detector has. However, how to collect a well distributed training set is time consuming. Different from many researchers who pay more attention to the re-sampling of the negative samples based on the Bootstrap, we lay more emphasis on the re-sampling of the positive samples. 

Solution 1:

Given a training set, we swell face database through re-sampling from existing faces by the genetic algorithm (GA)-based method, such as crossover and mutation. And we use the optimal selection to discard the bad samples and keep the good ones to improve the performance of the training set. Meanwhile, the crossover is performed by crossover the components of the faces, such as: eyes, noses, mouths etc., and the mutation includes kinds of the variations of similarities and gray value. See our FG2004 paper or contact Mr.Jie Chen.

Solution 2:

There are always noise data in the collected face samples or it may be biased due to the innocence about the distribution of the faces when one collects training set. It is necessary to resample and purify the noisy training set. We use the Isomap (isometric feature mapping) to carry out this operation because of the nonlinear distribution of the faces in the high dimension. See our AVBPA2005 paper or contact Mr.Jie Chen.

Solution 3:

SVM tells us that only those boundary examples, called support vectors (SVs), are useful for the final decision. Also, another popular algorithm AdaBoost pays more attention to difficult training samples, which lie close to the decision boundary and thus are more likely to be misclassified, by assigning them larger weights. We adopt a nonlinear technique based on the RS method to generate new examples, which lie close to the face/non-face class boundary. Then we add these examples to the original database to enhance it. (submitted to ICPR2006)

  • Spatial histogram features and their discriminating analysis

Motivation

Feature extraction for object representation plays an important role in automatic object detection systems. Previous methods have used many representations for object feature extraction, such as raw pixel intensities, edges, wavelets, rectangle features, and local binary pattern. However, what kinds of features are stable and flexible for object detection still remains an open problem.

Methods

An object representation method based on local texture analysis (i.e. a kind of improved spatial histogram features) is proposed in this project. Motivated by the observation that objects have texture distribution and shape configuration, we propose spatial histogram based features (termed as spatial histogram features) to represent objects. As spatial histograms consist of marginal distributions of an image over local patches, the information about texture and shape at different scales of the object can be encoded simultaneously. We present the extraction method of spatial histogram features and gives quantitative analysis of the discriminative ability of spatial histogram features. See our PCM2004 paper or contact Mr.Hongming Zhang.

  • Object detection using spatial histogram features

Motivation

At present, most object detection approaches are designed for one specific object class and they are not easy to extend to deal with other object classes. Moreover, how to learn informative features for object detection is an unsolved problem.

Methods

A generic object detection approach based on spatial histogram features is presented. This approach uses spatial histogram features as object representation, employs histogram matching and support vector machine to construct an object detector based on a coarse-to-fine strategy. The proposed method is efficient and robust to object detection, and it can be extended to generic object detection for different object types including rigid objects, non-rigid objects and texture rich objects. (to appear in Image and Vision Computing or contact Mr.Hongming Zhang)

A feature selection approach based on class separability and feature correlation is presented. The proposed method uses Fisher criterion to measure class separability of each feature and employs mutual information to calculate features correlation. Optimal feature subset is constructed by selecting uncorrelated and discriminative features through minimizing the classification error rate. The experiment results show that the proposed feature selection method is efficient to extract informative and class-specific features for object detection. See our IJCNN2005 paper or contact Mr.Hongming Zhang.

  • Multi-view face detection using Error Correcting Output Codes

Motivation

In real applications, a perfect face detection method should deal with face pattern with multi views. We aim to provide novel method for multi-view face detection and multi-class object detection.

Methods

A multi-view face detection method based on spatial histogram features and ECOC (error correcting output codes) multi-class classification is presented. First, we use the ECOC multi-class classification framework to extend the spatial histogram features based object detection approach to solve the multi-view face detection problem. Second, we present an approach for designing efficient binary classifiers by learning informative features through minimizing the error rate of the ensemble ECOC multi-class classifier. (Accepted by ECCV2006 as an oral paper or contact Mr.Hongming Zhang.)