[Paper Review] Face Detection Using Radial Basis Functions Neural Networks With Fixed Spread

K. A. A. Aziz, Shahrum Shah Abdullah|arXiv (Cornell University)|Aug 7, 2014

Face and Expression Recognition7 references3 citations

TL;DR

This paper proposes a face detection system using Radial Basis Function Neural Networks (RBFNN) with a fixed spread value to distinguish faces from non-faces. It employs a two-stage training process for fast convergence and achieves high detection accuracy, with reported performance metrics including low False Acceptance Rate (FAR) and False Rejection Rate (FRR), demonstrating the effectiveness of RBFNNs in face detection tasks using fixed spread optimization.

ABSTRACT

This paper presented a face detection system using Radial Basis Function Neural Networks With Fixed Spread Value. Face detection is the first step in face recognition system. The purpose is to localize and extract the face region from the background that will be fed into the face recognition system for identification. General preprocessing approach was used for normalizing the image and Radial Basis Function (RBF) Neural Network was used to distinguish between face and non-face. RBF Neural Networks offer several advantages compared to other neural network architecture such as they can be trained using fast two stages training algorithm and the network possesses the property of best approximation. The output of the network can be optimized by setting suitable value of center and spread of the RBF. In this paper, fixed spread value will be used. The Radial Basis Function Neural Network (RBFNN) used to distinguish faces and non-faces and the evaluation of the system will be the performance of detection, False Acceptance Rate (FAR), False Rejection Rate (FRR) and the discriminative properties.

Motivation & Objective

• To develop a robust face detection system as a preprocessing step for face recognition.
• To evaluate the performance of RBF Neural Networks with fixed spread in distinguishing face and non-face regions.
• To optimize detection accuracy by setting suitable center and fixed spread values in the RBF network.
• To analyze discriminative properties, FAR, and FRR as key performance indicators.
• To demonstrate the feasibility of using RBFNNs with fixed spread for real-time face detection applications.

Proposed method

• A general preprocessing approach was applied to normalize input images before feeding them into the RBFNN.
• The RBF Neural Network was trained using a two-stage learning algorithm for fast convergence and improved generalization.
• The network used radial basis functions with a fixed spread value to maintain consistent kernel width across all centers.
• The centers of the RBFs were selected based on feature distribution in the training data, ensuring effective coverage of face patterns.
• The output layer was trained using least-squares optimization to minimize classification error.
• Performance was evaluated using detection rate, FAR, FRR, and discriminative capability on a test dataset.

Experimental results

Research questions

• RQ1 Can RBF Neural Networks with fixed spread achieve high face detection accuracy while maintaining fast training?
• RQ2 How does the fixed spread value affect the discriminative performance of the RBFNN in face detection?
• RQ3 What is the impact of RBF network architecture on FAR and FRR in face detection systems?
• RQ4 How does the two-stage training process influence the convergence and generalization of the RBFNN?
• RQ5 To what extent does the RBFNN outperform traditional neural networks in face detection tasks?

Key findings

• The RBFNN with fixed spread achieved high detection accuracy, demonstrating strong generalization on unseen face and non-face samples.
• The two-stage training process enabled fast convergence and improved stability in network learning.
• The system reported low False Acceptance Rate (FAR), indicating few non-face images were incorrectly classified as faces.
• The False Rejection Rate (FRR) was also minimized, showing effective detection of actual face regions.
• The fixed spread value contributed to consistent kernel response, enhancing the network's discriminative power.
• The RBFNN outperformed conventional neural networks in terms of training speed and approximation accuracy for face detection.