The goal of face detection is to determine whether or not an image or video frame contains faces and, if present, return the number of instances of each face object and their location within an image space. Face detection is an important computer vision task as it is the building block for more sophisticated face processing algorithms such as face recognition and facial expression tracking. However, robust and reliable face detection in completely unconstrained settings remains a very challenging task. For example, while the human brain performs face detection and recognition robustly and with apparent ease, computer algorithms continue to find this a difficult task due to the huge variation of facial appearance in still images and video sequences. The existing literature documents extensive work on face detection utilising different classical machine learning and traditional algorithmic techniques. Given that challenges such as invariance to facial pose still remain with these traditional machine learning approaches, an exploration of biologically representative solutions that behave adaptively and autonomously through learning may help account for the well documented superior human and primate detection performance. In an effort to implement a more biologically plausible approach to invariant multi-view face detection, this thesis presents a novel hierarchical Spiking Neural Network (SNN) framework that adopts a hybrid approach to learning. This is achieved by combining a bottom-up unsupervised Spike-Timing Dependent Plasticity (STDP) feature extraction and filtering phase with a supervised feature selection process that provides feedback to the framework in an effort to select the most diagnostic neurons for accurate face detection. The detection accuracy of the hybrid system is further enhanced through two biologically plausible mechanisms of error control; namely threshold potential adaptation and spike latency thresholding. The broadly tuned behaviour of the neurons allows for a small but expressive set of multiview neurons to achieve efficient and robust detection for multi-view face poses. The merged, multi-view face detection system is further adapted through a competitive lateral inhibition mechanism to achieve accurate in-plane and out-of-plane face pose estimation.
| Identifer | oai:union.ndltd.org:bl.uk/oai:ethos.bl.uk:722684 |
| Date | January 2017 |
| Creators | McCarroll, Niall |
| Publisher | Ulster University |
| Source Sets | Ethos UK |
| Detected Language | English |
| Type | Electronic Thesis or Dissertation |
Page generated in 0.002 seconds