[Paper Review] An FPGA-based Parallel Architecture for Face Detection using Mixed Color Models

Tao Luo, Zaifeng Shi|arXiv (Cornell University)|May 27, 2014

Face and Expression Recognition10 references3 citations

TL;DR

This paper presents an FPGA-based parallel architecture for real-time face detection using mixed color models (YCgCr and YIQ). It detects skin color, applies morphological filtering, and converts results to user-configured preferred skin tones, achieving high detection rates with low false positives and minimal computational cost.

ABSTRACT

In this paper, a reliable method for detecting human faces in color images is proposed. This system firstly detects skin color in YCgCr and YIQ color space, then filters binary texture and the result is morphological processed, finally converts skin tone to the preferred skin color configured by users in YIQ color space. The real-time adjusting circuit is implemented and some of simulation results are given out. Experimental results demonstrate that the method has achieved high rates and low false positives, another advantage is its simplicity and minor computational costs.

Motivation & Objective

• Address the need for real-time, efficient face detection in embedded systems with low computational overhead.
• Overcome limitations of single-color-space approaches by leveraging multiple color models (YCgCr and YIQ) for improved skin detection.
• Design a hardware-optimized, parallel architecture on FPGA to enable real-time processing of color images.
• Minimize false positives through morphological filtering and user-configurable skin tone mapping.
• Achieve high detection accuracy with low resource utilization, suitable for embedded vision applications.

Proposed method

• Detect skin regions using YCgCr color space for initial segmentation.
• Apply morphological processing to binary texture maps to refine detected skin regions.
• Convert detected skin tones to user-defined preferred skin colors in YIQ color space.
• Implement a real-time adjustable circuit for dynamic skin tone configuration during operation.
• Utilize a parallel FPGA architecture to accelerate processing and ensure real-time performance.
• Integrate color space conversion and filtering stages in a pipelined hardware design for throughput optimization.

Experimental results

Research questions

• RQ1 Can a mixed-color-model approach improve face detection accuracy compared to single-color-space methods?
• RQ2 How effective is morphological processing in reducing false positives in skin region detection?
• RQ3 To what extent can an FPGA-based architecture achieve real-time performance with low computational cost?
• RQ4 How does user-configurable skin tone mapping enhance detection robustness across diverse skin tones?
• RQ5 What is the trade-off between detection accuracy, false positive rate, and hardware resource usage in the proposed FPGA design?

Key findings

• The proposed method achieves high face detection rates with low false positive rates, demonstrating robust performance.
• The system operates in real time due to the efficient parallel FPGA architecture, enabling practical deployment.
• Morphological processing significantly reduces noise and false detections in binary texture maps.
• The use of multiple color spaces (YCgCr and YIQ) enhances detection reliability across varying lighting conditions.
• The computational cost is minimal, making the design suitable for resource-constrained embedded systems.
• User-configurable skin tone mapping allows adaptation to diverse skin tones, improving detection consistency.