Robust Multi-View Face Detection

As one of the most fundamental techniques in the field of computer vision and human-c omputer interaction, face detection has drawn much attention since the invention of digital im age processing. In the recent decade, the great progress made in face detection is mainly due t o the Viola-Jones face detection framework based on boosting method and cascade structure. However, as the real world scenario becomes more and more complicated, face detection algo rithms are facing huge challenges and traditional ones cannot meet the current application de mands.To solve this problem, this paper introduces a novel feature representation called aggreg ated channel feature which is suitable for face, fast to compute and has strong representation capacity. The channel feature means that it calculates features on gradients of different orient ations, magnitude and color space. This paper also makes a modification of the learning algor ithm used in Viola-Jones framework to further improve the efficiency and effectiveness of the classifiers. Besides, we exploit decision tree based boosting algorithm and soft-cascade to fur ther improve the efficiency and effectiveness. Based on these techniques, we train models for multi-view face detection.We tune our algorithm through many experiments. When evaluated on two face detectio n benchmarks, AFW and FDDB, the proposed multi-view face detection algorithm surpasses current state-of-the-art detectors by a considerable margin, while runs at 62 FPS for VGA ima ge. On AFW, it achieves 90% recall rate when the accuracy is about 100% and achieves 95% recall rate when the accuracy is about 90%. Experiments show that the proposed method can handle with various unconstrained settings and therefore is of great practical value.