University of Malta, Msida, MSD2080, Malta, Europe,

Transcription

1 Dataset and paper available at: and This is the longer version of the paper published in the BIOSIG2016 proceedings, and is almost identical to the one that was reviewed and accepted for publication. Please cite the published paper (with the same title) if you use information from this document. 1 A Large-Scale Software-Generated Face Composite Sketch Database Christian Galea 1 Reuben A. Farrugia 2 Abstract: Numerous algorithms that can identify suspects depicted in sketches following eyewitness descriptions of criminals are currently being developed because of their potential importance in forensics investigations. Yet, despite the prevalent use of software-generated composite sketches by law enforcement agencies, there still exist few such sketches which can be used by researchers to adequately evaluate face photo-sketch recognition algorithms when using these composites. The main contribution of this paper is the creation of the University of Malta Software-Generated Face Sketch (UoM-SGFS) database that will be made publicly available and which contains the largest number of viewed software-generated sketches, that also exhibit several deformations and exaggerations to mimic sketches obtained in real-world investigations. Further, in contrast to sketches found in other databases, all sketches in the UoM-SGFS database are represented in colour. Lastly, state-of-the-art recognition algorithms are found to perform worse on the software-generated composites than on hand-drawn sketches, while recognition accuracies still lag far behind those achieved for traditional photo-to-photo comparisons. Keywords: Software-generated composite sketches, dataset, face photos. 1 Introduction Sketches created from eyewitness descriptions of criminals have proved to be useful for law enforcement agencies, especially when there is insufficient evidence to identify the perpetrators [JKP12, Kl14]. Sketches are currently disseminated to the media and law enforcement officers so that any persons recognising the suspect in the sketch come forward with information leading to an arrest, a process that is time-consuming and not guaranteed to be successful. Therefore, automated methods that can compare these sketches with the mug-shot photographs maintained by law enforcement agencies can enable faster apprehension times, increase chances of success and improve utilisation of readily available resources [Kl14]. However, comparison of face photos to sketches has been described as perhaps the hardest face recognition scenario owing to the typical deformations, exaggerations and lack of detail in the face sketch due to factors such as memory loss of eyewitness [KLJ11, ZWT11, GS12, KJ13, MVS15]. 1 Faculty of Information & Communication Technology, Department of Communications & Computer Engineering, University of Malta, Msida, MSD2080, Malta, Europe, christian.galea.09@um.edu.mt 2 Faculty of Information & Communication Technology, Department of Communications & Computer Engineering, University of Malta, Msida, MSD2080, Malta, Europe, reuben.farrugia@um.edu.mt

2 2 Christian Galea an Reuben A. Farrugia (a) (b) (d) (c) (e) Fig. 1: Comparison of viewed, forensic and semi-forensic sketches and the corresponding photographs: (a) Viewed sketches from the CUFSF database [ZWT11] (b) Viewed sketches from the CUFS database [TW04, WT09, KLJ11], (c) Viewed software-generated composites from the PRIPVSGC dataset [Ha13, Kl14], (d) Semi-forensic sketches from the IIIT-D dataset [Bh12], (e) Forensic sketches from the PRIP-HDC dataset [Kl14] As shown in Figure 1, sketches can either be hand-drawn by forensic sketch artists and are thus termed hand-drawn composite sketches, or created using computer software such as FACES [FA] and EFIT-V [Vi] and are known as software-generated composite sketches. In addition, there are three types of hand-drawn and software-generated sketches [Ha13, KLJ11, Kl13, Bh12]: the first is forensic sketches as obtained from eyewitness descriptions in real-world criminal investigations, whose availability for researchers is typically limited. This problem is circumvented by creating sketches whilst viewing a subject or face photo, termed viewed sketches, of which a virtually unlimited number can be generated and have in fact been the predominant means of evaluating face photo-sketch recognition algorithms. Lastly, semi-forensic sketches are created when forensic sketch artists view a subject or face photograph for a few minutes and then create the sketch from memory.

3 A Large-Scale Software-Generated Face Composite Sketch Database 3 Unfortunately, most currently available datasets contain only hand-drawn sketches, which are nowadays used less often than software-generated sketches due to lower cost and training requirements [MTM06, Ha13]. Consequently, evaluation of face photo-sketch recognition algorithms is often performed on hand-drawn sketches only. Thus, the contributions of this paper are twofold: first, a new large-scale database called the Software-Generated Face-Sketch (UoM-SGFS) database containing 600 software-generated face sketches of 300 subjects is described, with sketches having several deformations and exaggerations to mimic real-world forensic sketches. To the best of the authors knowledge, the UoM-SGFS database (i) is the largest software-generated sketch database that will be made publicly available since the two other available databases contain only 123 images each (of the same subjects), and (ii) is the only dataset containing sketches represented in full colour, thus enabling the use of colour information for face photo-sketch recognition. The final contribution is the evaluation of leading algorithms on this new dataset, where it is shown that recognition using composite sketches still lags behind photo-to-photo face recognition. The rest of this paper is organised as follows: in Section 2, an overview of related databases and algorithms designed for face photo-sketch recognition are outlined, followed by a description of the new UoM-SGFS database in Section 3. A description of the methodology used to evaluate algorithms on the UoM-SGFS database and their results are given in Section 4. Concluding remarks and directions for future work are finally given in Section 5. 2 Related Work Algorithms designed for the identification of subjects depicted in sketches must be evaluated on datasets containing not only subjects sketches but also the corresponding photographs. An overview of the databases currently available and the methods used for face photo-sketch recognition will now be discussed. 2.1 Existing face photo-sketch databases There exist five main publicly available databases for face photo-sketch recognition, with examples from each database shown in Figures 1 and 2. The first is the CUFS database 3 [TW04, WT09] consisting of 606 sketches, depicting subjects in three face databases. The CUFSF database 4 [ZWT11] contains 1194 sketches of subjects in the FERET database, which has been superseded by the Color FERET database 5 [PWR98]. Although both the CUFS and CUFSF databases contain viewed hand-drawn sketches, those in the CUFSF database are closer to real-life sketches since they were designed to contain several deformations and shape exaggerations [ZWT11]. 3 Available at: 4 Available at: 5 Available at:

4 4 Christian Galea an Reuben A. Farrugia The IIIT-D sketch database 6 [Bh12] contains 238 viewed, 190 forensic and 140 semiforensic sketches of subjects present in a number of databases. However, only 55 of the forensic sketches are publicly available and all sketches are hand-drawn. The remaining two datasets are the PRIP Hand-Drawn Composite (PRIP-HDC) [Kl14] and the PRIP Viewed Software-Generated Composite (PRIP-VSGC) [Ha13, Kl14] datasets 7. PRIP-HDC contains forensic sketches, although these are hand-drawn and only 47 photosketch pairs are available. PRIP-VSGC contains viewed software-generated composite sketches created using Identi-Kit [Id]. Although other sketches were created by another user and by another software program (FACES [FA]), these have not been made publicly available. As a result, this database only contains one sketch for each of the 123 subjects considered. The Extended PRIP (EPRIP) database [Mi14, MVS15] contains 123 sketches of the same subjects, created by an Indian software operator. Other databases have been used in literature, but these have not been made publicly available. As a result, there is only one database (PRIP-VSGC), and an extended version of it (EPRIP), that contain software-generated composite sketches available for researchers. However, the number of sketches available is quite low for robust algorithm evaluation and the similarity in appearance of sketches in the PRIP-VSGC dataset with respect to the corresponding photographs is generally quite poor. Indeed, the creators of the database themselves use the sketches generated with the FACES software for algorithm evaluation instead due to the better representational accuracy [Ha13]. This is mostly because the PRIP-VSGC sketches were created by an Asian operator, whereas over 95% of subjects considered are Caucasian [Ha13, Mi14]. Therefore, the synthesised sketches tend to look like Asian people due to the other-race effect, where people belonging to a particular ethnic group have difficulty in recognising and processing faces of people in other races and therefore the Asian user selected components that he is more familiar with [Ha13]. 2.2 Face photo-sketch recognition algorithms Traditional Face Recognition Systems (FRSs) such as Eigenfaces [TP91] and the VGG- Face algorithm [PVZ15] that has achieved state-of-the-art performance for unconstrained face recognition can be used for identification of faces depicted in sketches. However, as shown in Section 4, they do not achieve good performance since the images to be compared reside in different modalities. Consequently, methods have been designed specifically for face photo-sketch recognition which can generally be classified into two categories: the first set of approaches are known as intra-modality methods, which reduce the modality gap by transforming an image from one domain to another (e.g. photo-to-sketch transformation) such that the images to be compared lie within the same domain. Some of the most prominent intra-modality approaches include Eigentransformation [TW04] where an optimal linear combination of weights to reconstruct a photo (or sketch) are derived using photos (sketches) in a training set which are then replaced with the corresponding sketches (photos) to yield pseudo-sketches (pseudo-photos), the Eigenpatches extension proposed 6 Available at: 7 Available at:

5 A Large-Scale Software-Generated Face Composite Sketch Database 5 in [GF15] which synthesises local regions rather than the whole face as in Eigentransformation, and the Multiscale Markov Random Fields approach [WT09] that models the relationships among patches. A more detailed review of face hallucination methods, encompassing both face super-resolution and photo-sketch synthesis, may be found in [Wa14]. Intra-modality methods have been criticised for being too complex and computationally intensive, since they attempt to solve a problem that is arguably harder than the recognition task itself whilst producing images containing several artefacts that inhibit face recognition performance [ZWT11, Ha13, GF16]. As a result, there has recently been a shift towards inter-modality methods, which learn or extract modality-invariant features from the photos and sketches to be compared with the aim of maximising inter-class separability whilst minimising intra-class differences [GS12, Ha13]. Leading inter-modality methods include the Histogram of Averaged Orientation Gradients (HAOG) [GS12] approach which emphasises orientations belonging to strong edges and achieved a 100% recognition rate on the CUFS database, the Prototype Random Subspaces (P-RS) approach [KJ13] which uses a cosine kernel for comparison of feature descriptors obtained from filtered images, the Direct RS (D-RS) approach that uses a similar method to P-RS but compares features directly, and the LGMS approach [GF16] that uses Spearman correlation to compare Multi-Scale Local Binary Pattern (MLBP) features extracted from log-gabor-filtered images. 3 UoM-SGFS Database The new UoM-SGFS database contains software-generated sketches of 300 subjects in the Color FERET database, created using the EFIT-V software [Vi]. An overview of this software will first be given, followed by a more detailed description of the UoM-SGFS database. 3.1 EFIT-V overview Most facial composite systems involve the selection and spatial configuration of individual facial features, thus relying on the witness ability to recall and describe these features. However, it is not only difficult for witnesses to recall and describe a suspect s face, but human recognition and synthesis of faces is a global process relying on the interaction of all features in the face [Ge08]. EFIT-V was designed to counteract these drawbacks by using a global face model which allows witnesses to recognise perpetrator s faces rather than recall faces during the composite generation process. This is done by presenting the eyewitness with a set of faces, who retains or rejects them if they appear similar or different to the criminal, respectively. A genetic algorithm then learns the eyewitness s choices and presents a new set of faces through a process of mutations. The procedure is repeated again until the eyewitness is satisfied that a better likeness cannot be generated. Adjustments such as translation, scaling, and rotation may also be performed for each facial component. EFIT-V allows features other than the facial components to be depicted, including earrings, hats, necklaces and glasses which may not only help the eyewitness during the

6 6 Christian Galea an Reuben A. Farrugia composite generation process but also allow the creation of sketches which can depict discriminative and therefore critical information about the perpetrators quite well. Clothes and the general appearance of the criminal s physique can also be modified. Image editing software is directly supported in EFIT-V, which can be used to fine-tune details [Fr05]. The older EFIT software has proven to be effective in generating a good likeness to suspects, and was shown to be able to outperform not only other software systems but also hand-drawn sketches [Fr05]. As a result, the enhanced EFIT-V is nowadays used by numerous law enforcement agencies worldwide. More information may be found in [Ge08, Vi]. 3.2 Database details The face photographs for the new UoM-SGFS database were obtained from the Color FERET database 5 [PWR98], chosen since it contains a large number of good quality frontal images and a variety of different age ranges and ethnicities. As shown in Table 1, most subjects selected are White to minimise the other-race effect, since the software operator creating the sketches was also White. However, any law enforcement officers creating face sketches will typically take the description for any suspect irrespective of his race, and eyewitnesses will also be asked to describe the criminal even if they belong to different ethnic groups. Therefore, subjects belonging to different races to that of the operator were still included to represent real-life conditions where both law enforcement officers and eyewitnesses may belong to a different race than the perpetrator. In addition, the EFIT-V operator was trained by a qualified forensic scientist from a local police force so as to ensure that practices adopted in real-life were also used in the creation of the UoM-SGFS database. Moreover, photos in the Color FERET database often contain nonuniform lighting and slight head rotations, making the recognition task more challenging but also more realistic. The UoM-SGFS database contains two viewed sketches for each of the 300 subjects considered and is thus partitioned into two sets, where each set contains the sketch of one subject. Set A contains those sketches created using EFIT-V, where the number of steps Gender Race Male 61.00% White 56.67% Female 39.00% Asian 22.33% Black-or-African-American 6.33% Hispanic 5.33% Asian-Middle-Eastern 7.67% Pacific-Islander 1.00% Native-American 0.33% Other 0.33% Tab. 1: Demographic statistics of subjects in the UoM-SGFS database. Labels are the same as those given in the Color FERET dataset.

7 A Large-Scale Software-Generated Face Composite Sketch Database 7 performed in the program (e.g. number of evolutions, and selection and modifications of components) was minimised so as to lower the risk of producing composites that are overly similar to the original photo. In fact, the average time taken to create sketches varied between approximately 30 to 45 minutes. The sketches in Set A were then edited using the Corel PaintShop Pro X7 image editing software [Co] to fine-tune details which cannot be easily modified with EFIT-V, yielding Set B. Consequently, sketches in Set B are generally closer in appearance to the original face photos. On average, editing spanned approximately 15 to 30 minutes only, to retain inaccuracies as found in real-life forensic sketches. It should be noted that the Corel software was also used for sketches in Set A, but only to modify the hair component. Despite the extensive number of hair shapes and textures available in EFIT-V, the complexity of hair is such that it is often necessary to first select the most similar hair component and then modify it using the image editing software. The edited hair is included as part of the Set A sketches since it is also often done in real-life scenarios. The EFIT-V software also allows the depiction of shoulders in the face sketch, which can indicate both the type of clothes that the perpetrator was wearing and the physique (e.g. fat, muscular, etc.). While the type of clothing is important, more emphasis was given to correctly representing the physique of the subject since it provides more salient information. In addition, any accessories such as jewellery and hats are generally slightly different to those shown in the original photograph and sometimes omitted in the UoM-SGFS database sketches, to mimic memory loss effects of eyewitnesses. Lastly, to the best of the authors knowledge, all sketches currently available are represented in grayscale. However, the sketches in the UoM-SGFS database are all colourful. Colour has been shown to yield improved performance for face recognition [JA06] and therefore its use may now also be considered for photo-sketch recognition to potentially improve performance. Some examples of sketches in the UoM-SGFS database may be found in Figure 2. 4 Implementation Methodology & Results Some of the most popular and state-of-the-art algorithms are evaluated on both sets individually of the UoM-SGFS database. The protocol used and results will now be given. 4.1 Evaluation Methodology The two sets of sketches in the UoM-SGFS database are evaluated separately, with 150 subjects selected at random for training and the remaining 150 subjects used for testing. Twenty-five cross-validation folds are performed, with the means and standard deviations of results reported. Sketches and the corresponding photos form the probe and gallery sets, respectively, and all images are rotated such that the angle between the two eye centres is zero degrees and scaled such that the distance between the eyes and mouth are at the same location for all images. Lastly, all images are cropped to a height of 250 pixels and a width of 200 pixels. Since the top 50 to 200 subjects are generally given the same importance as

8 8 Christian Galea an Reuben A. Farrugia Set A Photos Set B Set A Photos Set B Fig. 2: Eight subjects from the Color FERET database and the corresponding sketches from the two sets of the UoM-SGFS database the best match (Rank-1) in criminal investigations [Ha13, KJ13, GF16], the Rank-retrieval rates are used as performance metrics together with the Area under the Receiver Operating Characteristics (ROC) Curve (AuC) that is typically used to evaluate FRSs. The gallery is extended with 1522 subjects to simulate the mug-shot galleries maintained by law-enforcement agencies, including 510 subjects from the MEDS-II database8, 476 subjects from the FRGC v2.0 database9, 337 subjects from the Multi-PIE database [Gr08], and 199 subjects from the FEI database10. As a result, the test gallery set contains = 1672 subjects. The algorithms chosen for evaluation include the classic Eigenfaces [TP91] and the stateof-the-art VGG-Face [PVZ15] FRSs and five of the best face photo-sketch recognition Available at: Available at: Available at:

9 A Large-Scale Software-Generated Face Composite Sketch Database 9 True Accept Rate (TAR) PCA VGG-Face HAOG P-RS D-RS D-RS + P-RS LGMS True Accept Rate (TAR) PCA VGG-Face HAOG P-RS D-RS D-RS + P-RS LGMS False Accept Rate (FAR) (a) False Accept Rate (FAR) Fig. 3: ROC curves of the algorithms considered, averaged over 25 set splits on (a) Set A (b) Set B (b) algorithms as outlined in Section 2. The parameters of all algorithms are the same as those described in literature, and the provided pre-trained model was used in the case of the VGG-Face algorithm. 4.2 Results The results of the algorithms considered are shown in Figure 3, Table 2 and Tables 3a and 3b. As expected, all algorithms perform better on Set B than Set A. This is because Set B contains sketches that were modified using an image editing software program to make the sketches closer in terms of appearance to the corresponding photographs. In addition, the significant increase in performance indicates the usefulness of employing image editing software to fine-tune details according to the eye-witnesses requests. However, the relatively low recognition rates at lower ranks indicates that Set B still poses a challenge for all algorithms, due to the deformations and shape exaggerations present. Algorithm Name Set A AuC Set B Eigenfaces [TP91] ± ± VGG-Face [PVZ15] ± ± HAOG [GS12] ± ± P-RS [KJ13] ± ± D-RS [KJ10, KJ13] ± ± P-RS+D-RS [KJ13] ± ± LGMS [GF16] ± ± Tab. 2: Means and standard deviations over 25 random train/test set-splits of the AuCs for the algorithms considered on both sets of the UoM-SGFS database

10 10 Christian Galea an Reuben A. Farrugia Algorithm Matching Rate (%) at Rank-N N=1 N=10 N=50 N=100 N=150 Eigenfaces [TP91] 2.03± ± ± ± ±1.74 VGG-Face [PVZ15] 11.57± ± ± ± ±2.85 HAOG [GS12] 15.92± ± ± ± ±2.05 P-RS [KJ13] 6.56± ± ± ± ±4.37 D-RS [KJ10, KJ13] 24.93± ± ± ± ±1.81 D-RS+P-RS [KJ13] 23.55± ± ± ± ±1.62 LGMS [GF16] 29.63± ± ± ± ±1.72 Tab. 3a: Means and standard deviations over 25 random train/test set-splits of the rank-retrieval rates for sketches in Set A. The performance of all algorithms when using the software-generated composites can also be compared to their performance on hand-drawn sketches in the CUFSF database. Although these results are omitted due to space constraints, it can be shown that similar trends in performance between algorithms are present, i.e. the Eigenfaces FRS performs worse than the inter-modality methods and the LGMS approach is statistically superior at the 95% confidence level using multi-comparison Analysis of Variance (ANOVA) to the other algorithms at lower ranks, while performance is comparable to D-RS and the fusion of P-RS and D-RS at higher ranks. However, all algorithms achieve lower rank-retrieval rates and AuCs when using the software-generated composites than when these algorithms are operated on hand-drawn sketches. For example, the LGMS method achieves between approximately 10% to 52% lower rank-retrieval rates on Set A and between approximately 4% to 35% lower rank-retrieval rates on Set B sketches, up to Rank-100. This is mainly due to three reasons: (i) the different sketch modalities (software-generated vs. hand-drawn), (ii) the lower number of software-generated composites available, and (iii) differences in the extent of deformations and exaggerations present in the CUFSF and UoM-SGFS databases. Also noteworthy is the performance of the VGG-Face algorithm, which achieved over 90% accuracy for unconstrained face recognition [PVZ15]. However, the performance is significantly degraded on the composite sketches, highlighting the challenges of not only the differences in modalities but also the deformations and exaggerations present in sketches. However, VGG-Face still outperforms P-RS on both sets of the UoM-SGFS database and achieved performance similar to HAOG at higher ranks on Set B. This is likely because the sketches generated using EFIT-V are quite photo-like, thereby facilitating the task of this state-of-the-art FRS that was designed to operate on face photographs. Nevertheless, its inferior performance compared to D-RS, D-RS + P-RS and LGMS indicate that the use of dedicated photo-sketch recognition algorithms is still the optimal choice. 5 Conclusions & Future Work A new database containing 600 EFIT-V-generated composite sketches of 300 subjects has been presented. This database is the largest publicly available dataset containing software-

11 A Large-Scale Software-Generated Face Composite Sketch Database 11 Algorithm Matching Rate (%) at Rank-N N=1 N=10 N=50 N=100 N=150 Eigenfaces [TP91] 3.04± ± ± ± ±1.70 VGG-Face [PVZ15] 17.41± ± ± ± ±2.58 HAOG [GS12] 26.93± ± ± ± ±1.55 P-RS [KJ13] 11.84± ± ± ± ±3.86 D-RS [KJ10, KJ13] 37.63± ± ± ± ±0.94 D-RS+P-RS [KJ13] 36.75± ± ± ± ±1.09 LGMS [GF16] 46.83± ± ± ± ±1.12 Tab. 3b: Means and standard deviations over 25 random train/test set-splits of the rank-retrieval rates for sketches in Set B. generated composite sketches, enabling more robust evaluation of face photo-sketch recognition algorithms to be performed. In addition, it is the only database containing sketches depicted in full colour rather than grayscale. Therefore, future work can include an investigation into the use of colour information for photo-sketch recognition and the use of physique information as a feature that can assist in recognition. References [Bh12] Bhatt, H. S.; Bharadwaj, S.; Singh, R.; Vatsa, M.: Memetically Optimized MCWLD for Matching Sketches With Digital Face Images. IEEE Transactions on Information Forensics and Security, 7(5): , Oct [Co] Corel PaintShop Pro, [FA] [Fr05] [Ge08] [GF15] [GF16] [Gr08] FACES 4.0, IQ Biometrix, Frowd, C. D.; Carson, D.; Ness, H.; McQuiston-Surrett, D.; Richardson, J.; Baldwin, H.; Hancock, P.: Contemporary composite techniques: The impact of a forensically-relevant target delay. Legal and Criminological Psychology, 10(1):63 81, George, Ben; Gibson, Stuart J.; Maylin, Matthew I.S.; Solomon, Christopher J.: EFIT-V -: Interactive Evolutionary Strategy for the Construction of Photo-realistic Facial Composites. In: Proceedings of the 10th Annual Conference on Genetic and Evolutionary Computation. GECCO 08, pp , Galea, Christian; Farrugia, Reuben A.: Fusion of Intra- and Inter-modality Algorithms for Face-Sketch Recognition. In (Azzopardi, George; Petkov, Nicolai, eds): Computer Analysis of Images and Patterns, volume 9257 of Lecture Notes in Computer Science, pp Springer International Publishing, Galea, Christian; Farrugia, Reuben A.: Face Photo-Sketch Recognition using Local and Global Texture Descriptors. In: 24th European Signal Processing Conference (EU- SIPCO). Budapest, Hungary, August Accepted for publication. Gross, R.; Matthews, I.; Cohn, J.; Kanade, T.; Baker, S.: Multi-PIE. In: 8th IEEE International Conference on Automatic Face Gesture Recognition. pp. 1 8, Sept 2008.

12 12 Christian Galea an Reuben A. Farrugia [GS12] [Ha13] [Id] [JA06] [JKP12] [KJ10] [KJ13] [Kl13] [Kl14] [KLJ11] [Mi14] Galoogahi, H.K.; Sim, T.: Inter-modality Face Sketch Recognition. In: 2012 IEEE International Conference on Multimedia and Expo (ICME). pp , July Han, Hu; Klare, B. F.; Bonnen, K.; Jain, A. K.: Matching Composite Sketches to Face Photos: A Component-Based Approach. IEEE Transactions on Information Forensics and Security, 8(1): , Jan Identi-Kit, Identi-Kit Solutions, Jones, C.; Abbott, A.L.: Color face recognition by hypercomplex Gabor analysis. In: 7th International Conference on Automatic Face and Gesture Recognition (FGR 2006). pp , April Jain, A. K.; Klare, B.; Park, Unsang: Face Matching and Retrieval in Forensics Applications. IEEE MultiMedia, 19(1):20 20, Jan Klare, B.; Jain, A. K.: Heterogeneous Face Recognition: Matching NIR to Visible Light Images. In: 20th International Conference on Pattern Recognition (ICPR). pp , Aug Klare, Brendan F.; Jain, Anil K.: Heterogeneous Face Recognition Using Kernel Prototype Similarities. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(6): , June Klum, S.; Han, H.; Jain, A. K.; Klare, B.: Sketch based face recognition: Forensic vs. composite sketches. In: International Conference on Biometrics (ICB). pp. 1 8, Klum, Scott J.; Han, Hu; Klare, Brendan; Jain, Anil K.: The FaceSketchID System: Matching Facial Composites to Mugshots. Technical Report MSU-CSE-14-6, Michigan State University, Klare, B.F.; Li, Zhifeng; Jain, A. K.: Matching Forensic Sketches to Mug Shot Photos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(3): , Mittal, P.; Jain, A.; Goswami, G.; Singh, R.; Vatsa, M.: Recognizing composite sketches with digital face images via SSD dictionary. In: IEEE International Joint Conference on Biometrics (IJCB). pp. 1 6, Sept [MTM06] Mcquiston, D.; Topp, L.; Malpass, R.: Use of facial composite systems in US law enforcement agencies. Psychology, Crime and Law, 12(5): , [MVS15] [PVZ15] [PWR98] [TP91] [TW04] Mittal, P.; Vatsa, M.; Singh, R.: Composite sketch recognition via deep network - a transfer learning approach. In: 2015 International Conference on Biometrics (ICB). pp , May Parkhi, O. M.; Vedaldi, A.; Zisserman, A.: Deep Face Recognition. In: British Machine Vision Conference Phillips, P. J.; Wechsler, H. amd Huang, J.; Rauss, P.: The FERET database and evaluation procedure for face recognition algorithms. Image and Vision Computing, 16: , Turk, M.; Pentland, A.: Eigenfaces for recognition. Journal of Cognitive Neuroscience, 3(1):71 86, Tang, X.; Wang, X.: Face sketch recognition. In: IEEE Transactions on Circuits and Systems for Video Technology. volume 14, pp , January 2004.

13 A Large-Scale Software-Generated Face Composite Sketch Database 13 [Vi] [Wa14] [WT09] [ZWT11] VisionMetric, About EFIT-V, Wang, Nannan; Tao, Dacheng; Gao, Xinbo; Li, Xuelong; Li, Jie: A Comprehensive Survey to Face Hallucination. International Journal of Computer Vision, 106(1):9 30, Wang, Xiaogang; Tang, Xiaoou: Face Photo-Sketch Synthesis and Recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 31(11): , Zhang, Wei; Wang, Xiaogang; Tang, Xiaoou: Coupled information-theoretic encoding for face photo-sketch recognition. In: 2011 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp , June 2011.