CASK-Net fusion: Multi branch approach for cross-age sketch face recognition

Abstract

Cross-Age Sketch Face Recognition targets the collective problem of Cross-Age Face Recognition (FR) and Sketch Face Recognition. Existing works discuss these problems individually, but no attempts towards collective version of these problems have been observed. In real life law enforcement, criminal and forensic investigations; the age and facial appearance of a subject may be different at sketch generation time and recognition time (present day). We therefore address this issue and propose a CASK-Net fusion approach to solve the collective problem of Cross-Age FR and Sketch FR. This paper presents a novel CASK-Net fusion architecture to capture discriminative features using multiple feature extractor branches including HOG, SIFT, CNN, LBP, ORB and Inception ResNetV2 (SOTA) respectively. The proposed approach grounds on extraction of age invariant features from sketch images of an individual for effective recognition. Our approach eliminates the requirement of modality conversion (sketch-photo) for recognition and provides less complex (transformation complexity is eliminated) solution. We also propose a benchmark Cross-Age Sketch (CASK) dataset to serve as a standard towards collective problem of Cross-Age FR and Sketch FR. The quantitative and ablation results highlight 95.52 % AUC-ROC performance and the fusion model achieved 93.37 % training accuracy (last epoch). Moreover, the SOTA comparison and dataset analysis confirms the model superiority with validation accuracy of 60.89 % on challenging and intrinsically hard CASK dataset.