A Review of Interpretability of Facial Features Based on Deep Learning

With the intensive integration of deep learning and computer vision, a series of advanced technologies such as facial recognition, image (video) generation, and image classification, have made rapid progress. However, deep learning models are considered “black box models” due to their difficulty in explaining internal processes and predicting results, which poses a serious challenge to the interpretability of image evidence in the field of forensic science. Based on this, this review outlines an overview of interpretability issues based on deep learning. Emphasis was placed on the theoretical and methodological research on the interpretability of facial features based on deep learning both domestically and internationally, such as saliency maps method, perturbation-based method, and score/statistics-based method. Their applications in facial recognition and other related fields, especially in the field of forensic science portraits, were summarized. This review proposes the problems of facial feature interpretability methods based on deep learning models, and looks forward to the future development direction of facial feature interpretability based on deep learning.