Deep Feature Representation and Similarity Matrix based Noise Label Refinement Method for Efficient Face Annotation

Abstract

Face annotation is a naming procedure that assigns the correct name to a person emerging from an image. Faces that are manually annotated by people in online applications include incorrect labels, giving rise to the issue of label ambiguity. This may lead to mislabelling in face annotation. Consequently, an efficient method is still essential to enhance the reliability of face annotation. Hence, in this work, a novel method named the Similarity Matrix-based Noise Label Refinement (SMNLR) is proposed, which effectively predicts the accurate label from the noisy labelled facial images. To enhance the performance of the proposed method, the deep learning technique named Convolutional Neural Networks (CNN) is used for feature representation. Several experiments are conducted to evaluate the effectiveness of the proposed face annotation method using the LFW, IMFDB and Yahoo datasets. The experimental results clearly illustrate the robustness of the proposed SMNLR method in dealing with noisy labelled faces.