| Abstrakt: |
In surveillance application systems, detecting and recognizing the face assumes a significant role. Therefore, some useful information about the face may be helpful in this situation to assess the face and its structure. This information can be defined in terms of face landmarks. These landmarks represent the localization of key characteristics in terms of point. With the help of deep learning techniques, the identification of landmarks is got more accuracy in various challenging situations. These challenging situations are different postures of humans such as sitting or sleeping or standing positions. Apart from these, the expressions, illumination, occlusion, and shadow are also other factors. This paper is targeted to outline the type of face landmarks, and algorithms that are used in it with performance. The author applied a transfer learning approach with ResNet architecture on the iBUG 300-W dataset to detect the coordinates of the 68-face landmarks at left & right eyebrows, left & right eyes, nose, lips, chin, and jawline. Haar cascade classifier and DNN are experimented with to detect and localize the face. The performance is evaluated through the training loss and validation loss in each iteration on the original image and image behind the glass. The observation indicates that the train-test split training loss is 0.0008 and the validation loss is 0.0011, with the model saving 0.000995 of the minimum loss. After that, the performance is also checked on the image at the back of the glass with shadows, face pose, and occlusion for more accuracy. [ABSTRACT FROM AUTHOR] |
