A Bimodal Approach for Partially Occluded Face Detection and Recognition for Crime Control in Nigeria Using Deep Learning and Machine Learning Algorithms

For the purpose of crime prevention and control, much effort has been made in literature on accurate face recognition using several approaches. However, little had been achieved on accurate identification of partially occluded faces, which is now the growing trend among criminals, as literature reveals. In this study, first, Deep Learning Multi-Task Cascaded Convolutional Neural Networks were used for face detection and face alignment, while VGG16 Convolutional Neural Network architectures were used for feature learning and classification. Secondly, and for result comparison, the Machine Learning Histogram of Orientated Gradients (HOG) with the Support Vector Machine algorithm was used for face detection as well. The feature vectors generated by the HOG descriptor were used to train Support Vector Machines (SVM), and the results were validated against given test input. The model was trained with datasets obtained from Disguised Faces in the Wild and with primary data of African facial images (occluded and non-occluded) comprising diverse occlusion patterns. Obtained results showed that the Convolutional Neural Network produces a recognition accuracy confidence level of 96% for occluded faces as opposed to the Histogram of Gradients. Convolutional Neural Network is recommended therefore for detecting and recognising partially occluded faces. For improved performance results, using larger datasets is recommended.