Enhancing facial geometry analysis by DeepFaceLandmark leveraging ResNet101 and transfer learning

The face recognition is the pivotal component in the surveillance system. To comprehensively assess facial structures, acquiring facial features becomes indispensable. This information is effectively represented by facial landmarks, which serve as specific locations of key characteristics. This manuscript proposes a novel DeepFaceLandmark algorithm by incorporating a transfer learning approach with the ResNet architecture on the Dlib iBUG 300-W dataset to detect the coordinates of 68 facial landmarks. This proposed work investigates the rationale behind encompassing DeepFaceLandmark in conjunction with facial feature extraction and face geometry analysis. The proposed work detects facial landmarks in challenging situations such as illumination, expression, head-pose, and occlusion to extract the facial features. In addition to these challenges, the proposed work also detected the facial landmarks on the face which are occluded via glass and partially visible due to the impact of light on glass. Furthermore, the author also specifically delves into elucidating the categories of face landmarks, detailing the employed algorithms, and evaluating their performance. Performance evaluation resulted in an impressive accuracy of 98.76%. The recorded training loss was 0.0003, with a validation loss of 0.0007 for the train-test split, and a minimum loss of 0.0006. Furthermore, the analysis extends its performance evaluation by conducting comparisons with other comparable state-of-the-art techniques.