Abstract:

Abstract Face recognition remains a general biometric verification approach employed for assessing the face images and excerpting beneficial identification data out of them that is consistently named as a feature vector, which is employed for differentiating the biological features. The face recognition procedure starts with excerpting the coordinates of features like mouth’s width, eyes’ width, pupil, and correlating these with a saved face template. The objective of the proffered scheme remains to craft an independent security system, which executes face recognition-based surveillance alongside a hardware mechanism for locking up the protected area.  Surveillance camera photographs of people tend to be of Low Resolution (LR), making it difficult to match them with High Resolution (HR) images. Super resolution, linked mappings, multidimensional scales, and convolutional neural networks are only moderately effective in practice.This study proffers Multi-Head Attention with VGG19 Backbone (MHA_VGG19) that is trained by face images of 3 remarkably disparate resolutions that are employed for excerpting distinctive features strong to the resolution alteration. This as well gives a quantization of the image specimens into a topological region in which inputs, which remain close in the original region as well as remain close in the output region; consequently, they give size decrement and invariability to petty alterations in the image specimen. The proffered methodology is widely analyzed employing LFW and Color FERET datasets by correlating with the advanced methodologies concerning different criteria. Subsequently, the proffered MHA_VGG19 attains 98.69% of accuracy, 99.06% of precision, 98.51% of recall, 98.75% of F1-score, and 100% of ROC for colour FERET database. By employing the LFW database, the proffered MHA_VGG19 attains 95.96% of accuracy, 96.41% of precision, 95.73% of recall, 96% of F1-score, and 99.83% of ROC.

Keywords: