A Novel Hierarchical Training Architecture for Siamese Sequential Network-LSTM Based Fauld Kinship Recognition for the Microexpression of Indonesian Faces

Analysis of facial images decoding familial features has been attracting the attention of researchers to develop a computerized system interested in determining whether a pair of facial images have kinship. Verifying kinship based on analyzing facial dynamics makes it possible to assess the level of resemblance between individuals within the context of kinship. Therefore, this research examined the use of microexpressions, widely used as a reference, as an alternative parameter for kinship detection. The primary focus of this research was to recognize lineage through microexpressions analysis using hybrid Long Short-Term Memory (LSTM) and Siamese Neural Networks (SNNs). The dataset was independently collected from Indonesian faces, which are fully microexpression images, to ensure its relevance. There was an imbalance in the dataset, with a clear difference in the number of data pairs between the various labels. The classifiers were tested on an independently created dataset and the Families in the Wild (FIW) dataset, where the model starts with the transfer learning model to extract features in each image and then continues with the classification network model to predict kinship relations based on the Euclidean distance of each two kinship images represented by a feature vector. These classifiers were designed with the underlying assumption that each kinship relationship can be treated as a binary classification problem with two outputs. The proposed method outperformed various state-of-Art methods.