The performance of most face recognition systems (FRS) in unconstrained environments is widely noted to be sub-optimal. One reason for this poor performance may be due to the lack of highly effective image pre-processing approaches, which are typically required before the feature extraction and classification stages. Furthermore, it is noted that only minimal face recognition issues are typically considered in most FRS, thus limiting the wide applicability of most FRS in real-life scenarios. Thus, it is envisaged that developing more effective pre-processing techniques, in addition to selecting the correct features for classification, will significantly improve the performance of FRS.
The thesis investigates different research works on FRS, its techniques and challenges in unconstrained environments. The thesis proposes a novel image enhancement technique as a pre-processing approach for FRS. The proposed enhancement technique improves on the overall FRS model resulting into an increased recognition performance. Also, a selection of novel hybrid features has been presented that is extracted from the enhanced facial images within the dataset to improve recognition performance.
The thesis proposes a novel evaluation function as a component within the image enhancement technique to improve face recognition in unconstrained environments. Also, a defined scale mechanism was designed within the evaluation function to evaluate the enhanced images such that extreme values depict too dark or too bright images. The proposed algorithm enables the system to automatically select the most appropriate enhanced face image without human intervention. Evaluation of the proposed algorithm was done using standard parameters, where it is demonstrated to outperform existing image enhancement techniques both quantitatively and qualitatively.
The thesis confirms the effectiveness of the proposed image enhancement technique towards face recognition in unconstrained environments using the convolutional neural network. Furthermore, the thesis presents a selection of hybrid features from the enhanced image that results in effective image classification. Different face datasets were selected where each face image was enhanced using the proposed and existing image enhancement technique prior to the selection of features and classification task. Experiments on the different face datasets showed increased and better performance using the proposed approach.
The thesis shows that putting an effective image enhancement technique as a preprocessing approach can improve the performance of FRS as compared to using unenhanced face images. Also, the right features to be extracted from the enhanced face dataset as been shown to be an important factor for the improvement of FRS. The thesis made use of standard face datasets to confirm the effectiveness of the proposed method. On the LFW face dataset, an improved performance recognition rate was obtained when considering all the facial conditions within the face dataset.