Evolutionary data augmentation in deep face detection

We present an evolutionary approach for Data Augmentation (DA) in deep Face Detection (FD). The approach is fully automatic and creates new face instances by recombining facial parts from different faces. We explore the selection of the facial parts that construct each new face instance using two strategies: random and evolutionary. The evolutionary strategy employs a Genetic Algorithm (GA) with automatic fitness assignment based on a pre-trained face detector. The evolutionary approach is able to find new face instances that exploit the vulnerabilities of the detector. Then we add these new instances to the training dataset, retrain the detector, and analyse the improvement of the performance of the detector. The presented approach is tested using deep object detectors, trained with instances from the CelebFaces Attributes (CelebA) dataset. The experimental results show that the presented approach improves face detection performance when comparing to base models trained using standard DA techniques. Also, the approach generates new realistic faces with interesting and unexpected features.