Introduction
Face detection is a fundamental task in computer vision, but traditional "sliding window" approaches are computationally expensive. These methods scan an entire image across multiple scales, leading to significant redundancy. While pre-processing techniques like skin color segmentation can narrow the search space, they often fail under varying lighting or across diverse skin tones. This paper proposes a robust, geometry-based pre-processing technique that identifies human silhouettes—specifically head and shoulder curves—to reduce the search space for face classifiers by up to 80%.
Methodology
The proposed method leverages the structural consistency of the human upper body. It utilizes Gradient Angle Histograms (GAH) to detect dominant linear edge segments within small image blocks.
The algorithm operates in three primary stages:
- Gradient Angle Histograms (GAH): The image is partitioned into small blocks, and edge maps are generated via Sobel filters. GAH identifies linear "anchor" edges by finding peaks in specific angular ranges (e.g., slanting up vs. down to the right).
- Piecewise Linear Curve Tracing: Curves are modeled as a sequence of shifting tangents. Starting from an anchor block, the algorithm searches neighboring blocks for edges where the gradient angle changes by a predefined delta ($\delta$), effectively tracing the convex shape of a head or shoulder.
- Macroblock Association: Detected curves are grouped into larger "macroblocks." A spatial check ensures that at least three of the four key curves (left/right head, left/right shoulder) are present in the correct geometric configuration to trigger a detection window.
Experimental Results
The technique was evaluated on the CASIA Face Image Database (constrained) and the Buffy dataset (unconstrained) to test robustness across complex backgrounds.
- Search Space Reduction: The method consistently reduced the pixels requiring full classification by approximately 80%.
- Detection Accuracy: On the CASIA dataset, a detection rate of 99.5% was achieved with a 6-pixel block size. On the more challenging Buffy dataset, the rate was 94.4% with an 8-pixel block size.
- Robustness: The system maintained low False Positives Per Frame (FPPF), with fewer than 2 for CASIA and approximately 6 for the unconstrained Buffy images.
Discussion and Future Work
By prioritizing structural geometry over color, this approach avoids common pitfalls related to illumination and skin tone. The grid-based approximation makes the process highly efficient and parallelizable. Furthermore, the requirement of only 3 out of 4 curves provides built-in robustness against partial occlusions, such as a hidden shoulder.
Current limitations include a focus on front-facing poses and a dependency on fixed scales. Future research will explore hierarchical implementations across image pyramids to handle varying human scales dynamically and extend the model to accommodate profile views where the standard silhouette is broken.
Summary updated