Face detection with a Viola–Jones based hybrid network

Abstract

Face detection is the determination of the positions and sizes of faces, primarily human, within digital images and videos, often as a component of a broader facial recognition system. It is seen as technologically mature, yet its operational performance typically remains sub-optimal, even within the less difficult frontal face detection tests. Empirical evidence shows that the Viola–Jones framework, a standard face detection solution with generally superior performance and other desirable properties, underdetects in some instances. Some true faces survive all but the final stages of the rejection cascade, resulting in missed faces. A hybrid framework consisting of a neural network following a truncated Viola–Jones cascade is constructed in an attempt to recover the undetected faces. Presumably, the neural network could fine tune and augment the face decision. Its inputs are a subset of the thresholding (detection) values of a rejection cascade's intermediate stages. Experiments reveal significantly improved performance, with increased detection rates if no false alarm increases are tolerated, with a greater detection rate increase if some false alarm increases are acceptable, and with a substantial false alarm reduction with no detection reduction. These improved face detection results could address shortcomings in widely-varying applications.