An Improved Siamese Network for Face Sketch Recognition

Abstract

Face sketch recognition identifies the face photo from a large face sketch dataset. Some traditional methods are typically used to reduce the modality gap between face photos and sketches and gain excellent recognition rate based on a pseudo image which is synthesized using the corresponded face photo. However, these methods cannot obtain better high recognition rate for all face sketch datasets, because the use of extracted features cannot lead to the elimination of the effect of different modalities' images. The feature representation of the deep convolutional neural networks as a feasible approach for identification involves wider applications than other methods. It is adapted to extract the features which eliminate the difference between face photos and sketches. The recognition rate is high for neural networks constructed by learning optimal local features, even if the input image shows geometric distortions. However, the case of overfitting leads to the unsatisfactory performance of deep learning methods on face sketch recognition tasks. Also, the sketch images are too simple to be used for extracting effective features. This paper aims to increase the matching rate using the Siamese convolution network architecture. The framework is used to extract useful features from each image pair to reduce the modality gap. Moreover, data augmentation is used to avoid overfitting. We explore the performance of three loss functions and compare the similarity between each image pair. The experimental results show that our framework is adequate for a composite sketch dataset. In addition, it reduces the influence of overfitting by using data augmentation and modifying the network structure.