Sketches obtained from eyewitness descriptions of criminals have proven to be useful in apprehending criminals, particularly when there is a lack of evidence. Automated methods to identify subjects depicted in sketches have been proposed in the literature, but their performance is still unsatisfactory when using software-generated sketches and when tested using extensive galleries with a large amount of subjects. Despite the success of deep learning in several applications including face recognition, little work has been done in applying it for face photograph-sketch recognition. This is mainly a consequence of the need to ensure robust training of deep networks by using a large number of images, yet limited quantities are publicly available. Moreover, most algorithms have not been designed to operate on software-generated face composite sketches which are used by numerous law enforcement agencies worldwide. This paper aims to tackle these issues with the following contributions: 1) a very deep convolutional neural network is utilised to determine the identity of a subject in a composite sketch by comparing it to face photographs and is trained by applying transfer learning to a state-of-the-art model pretrained for face photograph recognition; 2) a 3-D morphable model is used to synthesise both photographs and sketches to augment the available training data, an approach that is shown to significantly aid performance; and 3) the UoM-SGFS database is extended to contain twice the number of subjects, now having 1200 sketches of 600 subjects. An extensive evaluation of popular and state-of-the-art algorithms is also performed due to the lack of such information in the literature, where it is demonstrated that the proposed approach comprehensively outperforms state-of-the-art methods on all publicly available composite sketch datasets.
