The document introduces the topic of photo sketching and matching sketches to photos. It discusses how face recognition has improved with algorithms like Eigenfaces but challenges remain with conditions like lighting, pose and occlusion. A new challenge is matching sketches created by witnesses to criminal photos, which could help law enforcement but current techniques cannot do this. The document outlines two approaches - intra-modality which aims to transform sketches into pseudo-photos or vice versa, and inter-modality which uses modality-invariant features. It notes limitations of the intra-modality approach and issues with sketch and photo quality that can affect matching. The coming chapters will discuss databases, challenges in data collection, deep learning techniques like GANs, training and testing a model
The document introduces the topic of photo sketching and matching sketches to photos. It discusses how face recognition has improved with algorithms like Eigenfaces but challenges remain with conditions like lighting, pose and occlusion. A new challenge is matching sketches created by witnesses to criminal photos, which could help law enforcement but current techniques cannot do this. The document outlines two approaches - intra-modality which aims to transform sketches into pseudo-photos or vice versa, and inter-modality which uses modality-invariant features. It notes limitations of the intra-modality approach and issues with sketch and photo quality that can affect matching. The coming chapters will discuss databases, challenges in data collection, deep learning techniques like GANs, training and testing a model
The document introduces the topic of photo sketching and matching sketches to photos. It discusses how face recognition has improved with algorithms like Eigenfaces but challenges remain with conditions like lighting, pose and occlusion. A new challenge is matching sketches created by witnesses to criminal photos, which could help law enforcement but current techniques cannot do this. The document outlines two approaches - intra-modality which aims to transform sketches into pseudo-photos or vice versa, and inter-modality which uses modality-invariant features. It notes limitations of the intra-modality approach and issues with sketch and photo quality that can affect matching. The coming chapters will discuss databases, challenges in data collection, deep learning techniques like GANs, training and testing a model
The document introduces the topic of photo sketching and matching sketches to photos. It discusses how face recognition has improved with algorithms like Eigenfaces but challenges remain with conditions like lighting, pose and occlusion. A new challenge is matching sketches created by witnesses to criminal photos, which could help law enforcement but current techniques cannot do this. The document outlines two approaches - intra-modality which aims to transform sketches into pseudo-photos or vice versa, and inter-modality which uses modality-invariant features. It notes limitations of the intra-modality approach and issues with sketch and photo quality that can affect matching. The coming chapters will discuss databases, challenges in data collection, deep learning techniques like GANs, training and testing a model
The true identity of an individual is invaluable and extremely private information. While the average person has no concerns with their identity being known, some individuals would prefer to keep such knowledge secret despite the negative connotations attached to it. Typically, the motivation for an individual to hide his/her identity is to evade detection by law enforcement agencies for some type of criminal activity. Ongoing research in biometric recognition has offered a crucial method to help identify who a person truly is. Solving unconstrained face recognition often requires a considerable amount of research in face modelling, feature extraction and matching. The past two decades have witnessed major landmarks in face recognition algorithms. Turk and Pentland’s holistic Eigenface matching algorithm, for example, served as the basis for modern face recognition engines. Since the introduction of the Eigenface algorithm almost two decades back, face recognition accuracy has increased exponentially, to the point where the face recognition rates under controlled imaging conditions such as ambient lighting, frontal pose, neutral expression and uniformity in background, are comparable to fingerprint and iris matching rates. Unfortunately, real world face recognition scenarios seldom satisfy such controlled conditions. This has prompted researchers to turn the focus of their research in face recognition to more difficult problems such as varying degrees of illumination, non-frontal pose and occlusion. 1.2 Sketches and Images A new and very recent problem of interest in face recognition has emerged that deals with matching the sketched image of an individual to his/her photograph that may be stored in government records. The consequence of being able to develop a robust algorithm for the purpose of matching sketches to photos would be of great value to law enforcement agencies. When a witness sees a crime being committed, in many instances, the verbal description of the suspect elicited by the witness is used by a police artist to create a sketch. Many criminals have been apprehended when they are identified based on such sketches. State of the art face recognition techniques are not able to successfully match a sketch to the photographs which are stored in law enforcement databases. The sketch to photo matching capability of an algorithm implies a camera is not always mandatory to capture the face biometrics. 1.3 Modality 1.3.1. Intra-modality approach Previous approaches to face sketch-photo recognition have mostly been concentrated on reducing the modality gap by finding a transformation from sketch to photo aka pseudo-photo generation or from photo to sketch aka pseudo-sketch generation and subsequently matching with a photo or sketch respectively (intra-modality approach). Such an ideal transformation from sketch to photo or otherwise may not exist in the first place. This is because the photo and sketch come from two independent sources, i.e. a camera and an artist. Thus, it is not feasible to enforce an assumption of the existence of a perfect transformation from one to another. Although, an approximate transformation may be learnt from training a set of sketch-photo pairs via deep learning techniques, but it is likely to fall short to the intrinsic drawing style of a particular sketch artist and may not hold good for sketches drawn by other artists, who again, might have their own signature strokes. Thus, researchers are constantly trying to bridge the gap between the photo and sketch via either the inter modality approach or the intra modality approach. The intra-modality approach, as discussed above, aims to bring both sketch and image in the same modality by generating a pseudo sketch. 1.3.2. Inter-modality approach Whereas, the inter modality approach uses modality-invariant features to represent the images and perform the similarity measure based on this representation. For the inter modality approach, apart from having features that are independent of modality to represent the image, the sketch and photo quality must also be taken into account because they may affect the retrieval rate. To shine some more light, a sketch is drawn with no consideration of lighting conditions (i.e., no illumination, just a greyscale version of an image) but it may suffer from slight to moderate shape exaggeration (especially for forensic sketches, because eyewitnesses may have a different perception from the artist). While for photos, there is no chance of occurrence of shape exaggeration, but there is potential of them being exposed to lighting variations. Disregarding these imperfections will obviously sacrifice performance. Also, keeping in mind that sketches are drawn with no regard to the lighting conditions, matching the features from such representations is inaccurate. If mugshot photos are free from illumination variance, a better retrieval rate is expected because the extracted features are devoid of any illumination effects. 1.4 Organisation of Report In the coming chapters, we will see how important databases and their huge size are to any study, how to overcome challenges faced while trying to obtain data, various deep learning techniques, how GANs work to enhance performance, types of GANs, why we used the GAN we used, how to train a model and subsequently, test it, and many other topics of interest regarding this project.