A Fast Mobile Face Recognition System for Android OS Based on Eigenfaces Decomposition

Charalampos Doukas, Ilias MaglogiannisIFIP 2010 · DOI · OpenAlex

Abstract

This paper describes the implementation of a computationally efficient face recognition system for the Android mobile platform (version 1.6). By leveraging Eigenface decomposition based on Principal Component Analysis (PCA) and optimizing the classification process using Mahalanobis Cosine distance, the system achieves real-time performance on legacy mobile hardware. To balance autonomy and precision, a hybrid architecture is proposed, integrating local on-device processing with a remote cloud-based classification service. Experimental results show that the system can perform local recognition in approximately 1.2 seconds with 80% accuracy, while cloud-based SVM classification improves accuracy to 98%.

1. Introduction

In the early 2010s, mobile devices faced significant computational and memory constraints compared to desktop systems. Despite these limitations, there was a growing need for mobile assistive technologies, particularly for individuals with cognitive impairments such as dementia. This work addresses the challenge of implementing a robust face recognition system on mobile hardware (specifically the HTC G1) that can operate in real-time to assist users in identifying people.

2. Methodology

The proposed system utilizes a multi-stage pipeline designed for efficiency.

2.1 Preprocessing and Alignment

Initial face detection is performed using the native Android API to locate eye coordinates. An affine transformation is then applied to align the face horizontally. The image is resized to a 24x24 pixel template, and an elliptical mask is applied to exclude non-facial features like hair and background noise, which can negatively impact PCA performance.

2.2 Eigenface Decomposition

The core of the recognition engine is the Eigenface method. By applying Principal Component Analysis (PCA) to the training set, the high-dimensional image data is projected into a lower-dimensional subspace. Each face is represented as a weight vector corresponding to the most significant eigenvectors (eigenfaces).

Visual representation of the Eigenface decomposition process: (a) original face image, (b-d) examples of eigenfaces representing principal components of the dataset.

2.3 Classification and Hybrid Architecture

For local recognition, the system calculates the Mahalanobis Cosine distance between the input face vector and the stored training vectors. This metric was chosen for its superior balance of accuracy and low computational overhead on mobile CPUs.

To enhance accuracy, the system can offload classification to a Java-based web service via SOAP. This backend utilizes the Weka machine learning library to employ more complex classifiers, including Support Vector Machines (SVM), Neural Networks, and Bayes classifiers.

3. Experimental Evaluation

The system was evaluated using a dataset of 126 images featuring 7 subjects (4 males, 3 females). Each subject provided 18 training images (cropped to 100x100 pixels) and 5 separate testing images. The training phase, involving the calculation of eigenvectors for the entire dataset, was completed on-device in 65 seconds.

4. Results

The system demonstrated high efficiency and respectable accuracy:

  • Speed: Total local processing time (preprocessing, decomposition, and distance calculation) averaged 1.2 seconds per image.
  • On-Device Accuracy: The system achieved a 96% face detection rate and 80% recognition accuracy locally.
  • Cloud-Based Accuracy: Offloading to the web service significantly improved performance, with the SVM classifier reaching 98% accuracy. Neural Networks and Bayes classifiers achieved 96.2% and 95% accuracy, respectively.
  • Latency: While cloud processing was fast (<1s), network transmission was identified as the primary bottleneck for the hybrid approach.

5. Conclusion

This work demonstrates that real-time face recognition was achievable on early Android devices through careful algorithmic selection and a hybrid edge-cloud architecture. While the Eigenface approach is sensitive to lighting and pose variations, the system provides a viable foundation for mobile assistive technologies. Future improvements will focus on handling non-uniform illumination and pose variations, as well as implementing online learning for real-time database updates.

Summary updated

BibTeX
@inproceedings{doukas2010fast,
  author = {Charalampos Doukas and Ilias Maglogiannis},
  title = {{A Fast Mobile Face Recognition System for Android OS Based on Eigenfaces Decomposition}},
  booktitle = {IFIP},
  year = {2010},
  doi = {10.1007/978-3-642-16239-8\_39},
}
Figure