A Multimedia Application for Location-Based Semantic Retrieval of Tattoos

A Multimedia Application for Location-Based Semantic Retrieval of Tattoos Michael Martin, Xuan Xu, and Thirimachos Bourlai Lane Department of Computer Science and Electrical Engineering West Virginia University, Morgantown, WV, 26506 Mmarti40@mix.wvu.edu, Xuxu@mix.wvu.edu, Thirimachos.Bourlai@mail.wvu.edu Abstract The design and development of an automatic identity management system for the retrieval and matching of tattoo images is considered to be very important in the advancement of the investigative capabilities of forensics as well as law enforcement agencies. Conventional tattoo-based retrieval techniques are keyword-based, where content and body location meta-data are combined in an indistinguishable manner. However, the main drawback of these techniques is their inability to create search areas around manually annotated tattoo body location regions that can overcome user error in the meta-data or ambiguity in keywords. In this paper, we propose a novel method of using body location meta-data to perform semantic retrieval that significantly reduces error compared to keyword-based methods and reduces the number of matches needed to achieve high accuracy when performing tattoo-based human recognition. Our proposed approach is able to overcome the aforementioned challenges as well as other challenges that may be introduced when noise is present in the body location meta-data (e.g. possible tagging mistakes by the operators). 1. Introduction Tattoos are a soft biometric trait that have not been fully explored and have the potential to greatly benefit law enforcement when primary traits, such as face, iris, or fingerprints, may not be available or, when they are, their quality may be poor and cannot be used for human identification or to rule out suspects. While other soft biometric traits (ger, ethnicity, eye color etc.) are not sufficient unique to be used for identification, the detail present in tattoos gives them an advantage over more traditional soft biometric traits. However, due to the unconstrained dynamic nature of tattoos (i.e. variations in size, shape, color, content, body location, etc.), their usage in large-scale biometric systems can be more challenging when compared to the primary biometric traits. Thus, when only tattoo images are available, more complex computer vision and biometricrelated methods need to be developed in order to perform efficient matching and retrieval tasks that can assist in suspect identify verification scenarios. For this reason, there is an increased importance on the development of enhanced tattoo-based retrieval methods that are scalable and can improve the efficiency of indepent tattoo-based matching approaches. 1.1. Literature Review Previous research in tattoo-based human recognition was focused in segmentation, retrieval, and matching techniques. We have created a list (shown in Table 1) of the research found in the open literature that has been performed in the area of tattoo-based biometric recognition. In the table we refer to Constrained data as data collected by an expert using professional grade equipment with consistent background and controlled illumination conditions. Similarly, Unconstrained data is collected often using variable grade imaging sensors (e.g. cell phone cameras) with an inconsistent background. In this paper, all of the data used was of the Unconstrained type, which can often be much more challenging. What follows is a brief discussion on the segmentation, retrieval, and matching approaches used for tattoo-based recognition studies. Tattoo segmentation is very challenging due to the variability in the style of tattoos, their position on the human body, their shape and color, their content, level of completion and background interference. Several papers and techniques have been proposed specifically for overcoming some of the challenges associated with tattoo segmentation. Allen et al. [1] proposed a segmentation method combining bottom-up and top-down cues for the segmentation of tattoos. Using such techniques the authors demonstrated limited capability in isolating accurately the tattoo region. A more recent method of segmentation was proposed by Kim et al. [10] that uses a variety of skin and boundary techniques that achieves high accuracy when tested with the Tatt-C dataset. This work was exted in [9] to incorporate retrieval and matching techniques. An alternative to

Table 1. Tattoo Methods Reported in the Literature Problem domain Publication Data Source Style Allen et al. 2011 [1] www.gangink.com Unconstrained Duangphasuk et al. 2013 [2] Thai Criminal Records Division Constrained Huynh et al. 2014 [5] Custom Collected Constrained Segmentation Yi et al. 2015 [15] Tatt-C + Singapore Police Dept. Constrained + Unconstrained Kim et al. 2015 [7] Indiana Gang Network Unconstrained Kim et al. 2016 [10] Tatt-C Constrained Lee et al 2011 [11] Michigan State Police Dept. Constrained Manger et al. 2012 [12] www.tattoodesign.com + www.eviltattoo.com Unconstrained Jain et al. 2012 [6] Michigan State Police Dept. Constrained Matching and Retrieval Han et al. 2013 [4] www.gangink.com Unconstrained Kim et al. 2015 [8] Indiana State Police Dept. + www.eviltattoo.com Constrained + Unconstrained Kim et al. 2016 [9] Tatt-C Constrained Our Proposed www.checkoutmyink.com Unconstrained segmentation may be the use of Local Image Features. In the works of Jain et al. [11] [6] [4], segmentation was not performed directly, and instead, SIFT was used for the extraction of local image based tattoo features that can be used for recognition. In this work, we also use a similar tattoo feature extraction approach. Semantic retrieval has proven to be a useful tool in facebased biometric systems. In the work of Sridharan et al. [13], the authors proposed a multimedia application to operate their proposed semantic face retrieval approach. Their method uses both continuous (such as the location of certain facial features) and discrete features (such the presence of glasses on the subject) in combination with Bayesian learning, to reduce the gallery to a small number of face images where face-based matching techniques can be applied more efficiently in terms of speed and accuracy. The use of retrieval has also been used in tattoo-based biometrics having the same potential to positively impact system performance. Several methods have been proposed to use meta-data in an attempt to reduce the strain on performing matching. The most basic methods of performing retrieval are keywordbased, where a set of keywords are chosen from a discrete dictionary and, then, attempt to describe a tattoo s content or its location on the human body. For example, a common use of keywords is the ANSI/NIST-ITL1 2011 standard [14]. A tattoo collection and segmentation method was proposed in [5]. This approach used a full-body imaging system to capture tattoo, skin marks, androgenic hair, and veins. Using a joint detection algorithm, body locations of these features were extracted. Due to the limited tattoo datasets that are publicly available, most of the data used in previous studies were generated from online sources. The most commonly used online tattoo resource is Gang Ink website that contains images of gang-related tattoos from the Chicago area, i.e. 1,500 tattoo pictures on over 70 known gangs. Unfortunately, many of the images are of low quality and may contain challenging backgrounds. Some other popular online resources include Tattoo Design (www.tattoodesign.com) and Evil Tattoo (www.eviltattoo.com), which are online tattoo image galleries inted to help with tattoo design ideas. These online resources are inted to be art galleries, and thus, it is unlikely to find two tattoo images of the same subject. Another online tattoo resource Check Out My Ink (www.checkoutmyink.com), has never been used before as a source for tattoo-based biometric research. The advantage of this website is that it allows users to share pictures of their tattoos. This greatly increases the number of images of the same tattoo that can be found due to the likelihood of users to post multiple pictures of their tattoo to their profile. In this work, we took advantage of the variability but also the operational challenges offered by the aforementioned websites and generated challenging tattoo datasets that we used for the purpose of this study. This as well as our other contributions of our work are discussed below. 1.2. Contributions Our contributions for this work will be as follows: Large Tattoo Dataset - Using openly available tattoo images collected from Check Out My Ink we have created a dataset containing more than 1,000 distinct tattoos, with multiple images per tattoo. For each tattoo type, we randomly selected one image to generate a probe set of tattoo images. The remaining tattoo images were used to generate a gallery set. Additionally, meta-data is manually dictated for each gallery tattoo in order to facilitate our proposed tattoo retrieval process (the distribution of these labels are shown in Figure 1). Multimedia Application - We have designed and developed a multimedia application to allow for the reading, organization, and display of tattoo images from the probe and gallery sets. The C#.Net Framework application provides a foundation that was used to build our semantic-based tattoo retrieval method and matching schemes. This will allow a user (e.g. law enforcement officer) to select a probe tattoo image or a set of images and a gallery database, dictate meta-data for the probe tattoo, perform semantic retrieval, and then match the probe tattoo image to the retrieved gallery set. Semantic Retrieval Capability - Conventional keyword-based tattoo retrieval methods do not distinguish between tattoo body location information and tattoo content

Figure 1. Histogram showing the frequency of tattoo locations according to our labeled tattoo meta-data. The histogram is generated by using our tattoo database collected from a online source that consists of 27 body attributes of 1,012 distinct tattoos. descriptors, which cannot adjust for ambiguity in interpretation of the tattoo s location by the user or tattoos that span partial body regions. In order to address these limitations, we designed and developed a location based semantic retrieval method for the prioritization of the gallery set. Our method attempts to develop relations between body locations such that we can adjust for interpretation of the tattoo s location by the user. The method is very beneficial and manages to achieve 98.6% accuracy when 70% of the gallery set has been removed with semantic retrieval. The computational complexity of our proposed method is also efficient taking only 2.0 milliseconds to perform retrieval on a gallery of 1,000 tattoo subjects. To the best of our knowledge this is the first time such an approach is wrapped around a user frily application. Our approach is focused on body location based tattoo retrieval and it overcomes the aforementioned challenges due to ambiguity in body part labels. 2. Methodology We have created a software application to allow law enforcement operators to: read tattoo images and databases, select body region meta-data, perform matching prioritization, and, lastly, perform tattoo matching of a probe tattoos to a gallery enrolled tattoos. Our software application was developed using Visual Studio 2012 in C# and WPF (Windows Presentation Foundation). 2.1. Data Collection Tattoos are an emerging topic in biometrics, and thus, there is a lack of standardization of test data. Very recently, there was an effort by the National Institute of Standards and Technology (NIST) to create a tattoo recognition competition called the Tattoo Recognition Technology - Challenge (Tatt-C). In 2014 the first edition of a tattoo dataset was released by NIST to be used in Tatt-C. Although this dataset is likely to be very beneficial to the efforts of different researchers and practitioners to use tattoos as a biometric modality, its identification section is rather small compared to datasets used for other identification-based biometric studies, consisting of only 109 subjects (with at least two samples per subject). To the best of our knowledge there are no large-scale publicly available datasets of tattoo images that have been collected and used to assist the development of identification and retrieval algorithms when using soft biometricbased traits. For this reason we created our own large dataset of tattoo images from a set of online sources. One such valuable source is a social-media style tattoo website Check Out My Ink (see Table 1). This website allows users to create profiles, share their tattoos with other users, and offers many other functionalities including commenting, searching, etc. In order to gather a large number of tattoo images to allow for the development of retrieval and matching techniques that can be scalable, we searched through public user profiles from Check Out My Ink and can download their publicly available images. We were able to collect 124,086 images from 21,122 users over the span of a few days 1. In order for the collected images to be useful for retrieval and matching studies, the images were organized into subjects that contain matching images of the same tattoo. It was found that the majority of tattoo images collected did not have a pair image from the same individual to match. We were eventually able to manually organize a dataset of 1,012 tattoo pairs. The tattoo images were then manually cropped and organized into a gallery and probe set. The probe set contains exactly one image while the gallery is allowed to be an unconstrained number of samples to allow for instances where a single image perspective is not sufficient to capture the entirety of a tattoo. In addition, meta-data was generated for each tattoo pair that include a unique tattoo ID, the original profile username the tattoo was collected from, and body regions where the tattoo was located (discussed further in Section 2.3). All of the meta-data was stored in Microsoft Office Excel spreadsheet that can be read by our application. 1 The Terms of Service of Checkoutmyink.com prohibit us from directly sharing the images used in this study. Contact Checkoutmyink.com for permission to use the data and Dr. Thirimachos Bourlai for the public URL links to Checkoutmyink.com of the images used.

Figure 2. (a) A snapshot of the application developed is illustrated in this figure. On the left hand side, we can see the processing bar and all buttons available to the operator, including the probe, gallery, body area, prioritize, and match. Once matching has started the probe and gallery pair currently being matched is shown to the right. (b) Multimedia dialog allowing the user to input body part location meta-data for a tattoo. 2.2. Probe and Gallery Retrieval Interface In order to make our system convenient for law enforcement or forensic operators, we developed an easy-to-use interface. The tools allow a user to select a Probe tattoo set and a gallery set of tattoos. To allow for batch processing capabilities, the user can either select a single tattoo image from the probe set or a set of probe images by selecting a list of probe folder names and meta-data (where each folder contains at least one image of a tattoo). For simplicity, we stored all information about our probe and gallery sets, including meta-data, in a Microsoft Excel Spreadsheet. Each row of the spreadsheet is an instance of a different tattoo and stores all of the meta-data about the tattoo (including the tattoo ID, username, and body locations) in the columns corresponding to the particular row. The interface we designed (shown in Figure 2(a)) consists of three main windows. The middle and the right display windows show the probe (on the left) and the gallery (on the right) images currently being matched. The user controls are included on the left-most panel of the application and contains several buttons to: allow the reading of the probe and gallery images or databases, open the body location meta-data input window (shown in Figure 2b), run semantic prioritizing on the gallery database, and to begin matching the probe tattoos to the gallery tattoos. The body part meta-data selection window contains several checkboxes that have been organized into approximate body regions to the assist the user in easily locating specific regions. Multiple regions can be selected at any time and there is no constraint on what regions the user can chose at any point. Additionally, once the matching process has begun, a progress bar and cancel button is displayed so that the user can see the progress of matching the gallery to the probe and the matching execution at any point. Figure 3. Unidirectional Node Graph Representing the human body where each node is a body part and all nodes are connected by a weight of 1. 2.3. Semantic Retrieval Prioritizing To prioritize the database we chose to use the location of the tattoo as a preliminary indicator of how likely the tattoos are to match. We divided the human body into 27 regions (shown in Figure 3) in which tattoos could be located.

Choosing a subset of a gallery has been previously researched and has been used in schemes with other biometric modalities such as face and fingerprint [13]. The majority of these use keywords or labels input by the user to separate the gallery into subsets for matching. In our application, finding a subset of the database on labels alone (as used in other traditional biometric modality schemes) is not a viable option due to user interpretation of body regions and tattoos that can span multiple continuous body regions. Thus by creating a scheme of a semi-continuous relationship between labels, we manage to account for user errors when inputting the meta-data. We will demonstrate the advantages of our proposed scheme in terms of retrieval efficiency that also results in improving the overall matching capability of our tool when compared to the baseline matching approach. For this reason, we have created a new location-based semantic retrieval method that is based on representing the body as an undirected graph, where each node of the graph is a corresponding body part. Body regions that are adjacent are connected with a weight of 1, the resulting graph is shown in Figure 3. Using this graph, we can begin to determine the distance between any two given body regions. We define the distance between two body parts as the mincost-path between the respective nodes. 2.3.1 Min-Cost Body Part Match Probability For faster performance, a C# application was written to precompute the min-cost path between every two nodes in the graph. The min-cost path was then stored in a lookup table, which can then be accessed when performing semantic retrieval. Next, in order to solve the min-cost path between two nodes, we used the Floyd-Warshall algorithm. A brief description of this algorithm is shown in Algorithm 1. A more complete description of the Floyd-Warshall algorithm can be found in [3]. Algorithm 1: Floyd-Warshall Algorithm for i = 1 to N do for j = 1 to N do if there exists an edge from i to j then dist[0][i][j] = the cost of the edge from i to j else dist[0][i][j] = for k = 1 to N do for i = 1 to N do for j = 1 to N do dist[k][i][j] = min(dist[k-1][i][j], dist[k-1][i][k] + dist[k-1][k][j]) Where, N is the number of nodes and dist is a distance matrix to store the cost of the path throughout the iterations. Due to the relative simplicity of the node graph, the runtime of the Floyd-Warshall algorithm was very quick at only.667 milliseconds. A N N lookup table T is then constructed from the dist matrix such that the min path between a node i and a node j is T i,j. We then formulate a distance-based match probably P D : P D (X i, Y j ) = e ( T l(x i ),l(y j ) σ 1 ) Where X i is the i-th Probe tattoo, Y j is the j-th Gallery tattoo, l( ) are the body part labels and corresponding nodes from a tattoo s meta-data, and σ 1 is a scaling factor closely associated with the average min-cost path. If the tattoo spans multiple body part locations, and therefore contains more than one body part location label, the average of the min-cost paths is found such that the distance-based match probably P D is found as: P D (X i, Y j ) = e ( mean(t l(xi )m,l(y ) j )n σ ) 1 Where, l(x i ) m is the m-th label from the i-th probe metadata, and l(y j ) n is the n-th label from the j-th gallery metadata. 2.3.2 Label-Based Match Probability A label-based match probability P L is used to add robustness to our retrieval and increase match probability when there are multiple matching labels between the meta-data of the probe tattoo and a gallery tattoo such that l(x i ) m = l(y j ) n. We compute the label-based match probability P L as: P L (X i, Y j ) = e ( 27 ρ 1 +.25 ρ 2 σ ) 2 (3) Where ρ 1 is the sum of the matching labels between the probe tattoo and a gallery tattoo and is defined as: ρ 1 = 1 if l(x i ) m = l(y j ) n (4) m n Similarly, ρ 2 is the sum of non-matching labels between the probe tattoo and a gallery tattoo and is defined as: ρ 2 = 1 if l(x i ) m l(y j ) n (5) m n 2.3.3 Total Retrieval Match Likelihood Once the distance-based match probability P D and the label-based match probability P L have been computed, we can then combine them for the total match probability P total. We define P total as: P total (X i, Y j ) = ω 1 P D (X i, Y j ) + ω 2 P L (X i, Y j ) (6) (1) (2)

Where ω 1 is a weight assigned to the distance-based match probability P D, and ω 2 is a weight assigned to the label based match priority P L. The total match probability is then computed for every gallery tattoo. The results are ordered such that the highest match probability are matched first. 2.4. Tattoo Matching Tattoos can be quite difficult to match given their dynamic nature when compared to other more uniform biometric modalities. One method that has proven to be effective in the matching of tattoos is the use of local image features descriptor, namely the SIFT [11]. With SIFT we find matching features or keypoints between two images that contain similar content but it may also contain variations in lighting, change in 3D viewpoint, or noise. Some other commonly used local feature detectors include the SURF (Speeded Up Robust Features) and HOG (Histogram of Oriented Gradients). The use of these algorithms was explored in addition to SIFT, however, preliminary studies determined they exhibit overall lower performance. This is in agreement with previous research [11]. For these reasons we will limit our discussion to SIFT only in this work. The SIFT algorithm can only be used to extract keypoints from grayscale images. Thus, we first find the grayscale equivalent of the original color tattoo images. To enhance image content we then apply Contrast Limited Adaptive Histogram Equalization (CLACHE) to all images before extracting SIFT keypoints. We use a newly developed SIFT implementation developed by OpenCV (an open source computer vision and machine learning software library available for C++, C, Python, Java, and MAT- LAB at opencv.org ) in version 3.0 Alpha, which has been ported over for use in C#.NET Framework by EMGU (an open source third-party.net Framework wrapper for the OpenCV libraries). 3. Experiments and System Performance We conducted two experiments to 1) test the accuracy of our proposed tattoo-based semantic retrieval method and 2) evaluate the accuracy and performance benefits of using our semantic retrieval method in conjunction with a matching scheme. These experiments are described below. 3.1. Performance of Semantic Retrieval Firstly, proposed system is tested by examining the order of the gallery after prioritization (ordering the gallery from high to low priority) using body location meta-data. We show the distribution of the gallery with respect to the position of the correct gallery match corresponding to the probe using a Cumulative Match Characteristic (CMC) curve. Additionally, several other confounding factors are investigated to evaluate the performance of our system: The size of the dataset used: One factor that we consider is the scalability of our system. We tested our method on a set of subsets chosen at random from the complete tattoo dataset consisting of 1,012 tattoo images. Then, we analyzed at what percentage of the dataset the CMC curve reaches a particular level of accuracy (i.e. 1.0 or 100%). Variability in meta-data labels: The level or variability between the meta-data of the probe and gallery tattoo can significantly effect the performance of a meta-data based retrieval system. This is caused by bias introduced by ambiguous interpretation of the tattoo on the part of users both in the meta-data created on enrollment of the tattoo into the gallery set, and by the dictation of the probe tattoo s metadata. Additionally, errors in the meta-data creation can also contribute to variability between the probe and corresponding gallery tattoos meta-data. To test the effect of meta-data variability we first obtained the system performance when the meta-data of the probe set of tattoo images correctly match the meta-data of the gallery. Next we tested our system with probe meta-data labeled by an indepent individual that contains a certain amount of determinable variability from the gallery meta-data. To accomplish this, another expert in designing and developing tattoo-based biometric systems was asked to generate the indepently labeled set. Our experiments are designed to ensure the aforementioned factors are properly considered. In order to test our semantic retrieval method, we will consider the scenarios with gallery sizes of 250, 500, and 1,000 tattoos. The tattoos to form the subsets will be chosen at random from the total dataset of 1,012 tattoos. We also considered two scenarios where the probe meta-data labels are exactly matching that of the corresponding gallery meta-data (ideal scenario), and where the probe meta-data labels have been generated by another expert, indepent of the gallery meta-data labels (real-world scenario). The remainder of this section will discuss the results from two experiments testing: 1) the scalability of our system and 2) performance with meta-data variability. 3.1.1 Gallery Scalability Experiment Results To test the scalability of our proposed retrieval system, we performed a series of tests with various gallery sizes. The results from the tests with sizes of 250, 500, and 1,000 are shown in Figure 4(a). Additionally, we have analyzed the Rank for each test size at particular levels to determine what percentage of the top prioritized gallery is appropriate to use. In the worst case 73.6% of the gallery could be removed and retain 99.9% of the correct matches. The results show that a stable amount of the lowest scores in the gallery can be safely removed and our method s performance is indepent from the size of the gallery. In these

Figure 4. CMC curve performance of our Semantic Retrieval Method for (a) a gallery size of 250 tattoo, 500 tattoos, and 1,000 tattoos. (b) rank normalized to gallery percentage for a gallery size of 250 tattoos, 500 tattoos, and 1,000 tattoos. (c) Meta-data with the presence of a label noise of 14.03% Error (d) tattoo matching using Local Image Features between 60% of the gallery removed using Semantic Retrieval and the full size of the gallery (250 tattoos). experiments, where correctly labeled data was used, 73.6% of the gallery could be removed before matching and still achieve a performance of 99.9%, invariant of the gallery size. This can be better visualized when the ranks of the CMC curve are normalized to be a percentage of the gallery size as shown in Figure 4(b). Additionally, we performed a one-way ANOVA test to confirm that our conclusions are statistically significant. 3.1.2 Meta-data Variability Experiment Results. In this meta-data variability experiment we compare a set of indepently labeled meta-data (similar to what we would expect when the system is deployed in a Real-World Scenario) to correctly labeled meta-data that contains no errors. The labeled meta-data used for the gallery contained 1,614 labels while the indepently labeled meta-data contained 1,661 labels for 1,012 tattoos. Two types of meta-data label errors can exist with Type 1 Error being when the gallery meta-data contains a label that is not present in the probe meta-data and Type 2 Error being when the probe meta-data contains a label that is not present in the gallery meta-data. When the gallery meta-data was compared to the indepently labeled meta-data we found that there were 93 instances of Type 1 Error (missing probe meta-data label) and 140 instances of Type 2 Error (extra probe meta-data label). We calculate that 14.03% of our probe meta-data is erroneous by: error = T ype 1 Error + T ype 2 Error Number of P robe Labels Semantic Retrieval was then performed for the indepently labeled and correctly labeled meta-data on the entire dataset of 1,012 tattoos. The results (shown in Figure 4(c)) show that even with 14.03% error in the meta-data labels, our semantic retrieval methods is still able to achieve high performance. At Rank 300, our probability of identification is 98.62% for the meta-data with errors, and 99.9% for the meta-data with no errors. At this meta-data error level we would still be able to remove 70% of the gallery and still (7)

retain an accuracy of 98.62% prior to matching. 3.2. Performance of Matching Tests To assess the impact of our proposed Semantic Retrieval method in a matching scenario, we performed an identification experiment in which we match a set of 250 probe and gallery tattoos. SIFT features were used to create a similarity score between probe and gallery tattoo images. We removed 60% of the gallery using our proposed semantic retrieval and compared the results with full gallery dataset of 250 tattoos. We see a significant performance increase in matching accuracy (shown in Figure 4(d)) and matching time. The performance increase is tested to be statistically significant using a Wilcoxon signed rank test at the 5% confidence interval. The computational performance of our system takes an average of 2.0 milliseconds to prioritize and order a gallery of 1,012 tattoos when testing on a desktop PC running 64-bit Windows 7 and containing 12.0 GB of RAM, Intel Core i7 CPU 950 (3.07 GHz), and a NVIDIA GeForce GTXZ 570. 4. Conclusion In this paper, we presented an application for the retrieval and matching of probe and gallery tattoo sets that is indepent from image quality and able to adjust for error in meta-data. Furthermore, we have proposed a novel approach for the prioritization of retrieved tattoos from a large database based solely on body location meta-data in order to quickly determine the most likely match. Through developing relations between human body parts, we have created a method of adjusting for noise and ambiguity in tattoo meta-data that is not possible with traditional keyword label based retrieval schemes [6]. By using only meta-data, the performance of our system method cannot be impacted by poor image quality or low resolution. Additionally, this allows us to achieve a very fast system performance time (2.0 milliseconds for a gallery of 1,012 tattoos) that is not currently possible when dealing with image data. Lastly, our method could be used in conjugation with existing contentbased retrieval methods [11] to extensively improve the results with an insignificant additional system performance time increase. Acknowledgment We acknowledge that this effort is partly supported by West Virginia University. The opinions, findings, and conclusions or recommations expressed in this publication are those of the authors and do not necessarily reflect the views of our sponsors. The identification of any commercial product or trade name does not imply orsement or recommation by the authors and/or West Virginia University. References [1] J. D. Allen, N. Zhao, J. Yuan, and X. Liu. Unsupervised Tattoo Segmentation Combining Bottom-Up and Top- Down Cues. In SPIE Defense, Security, and Sensing, pages 80630L 80630L. International Society for Optics and Photonics, 2011. [2] P. Duangphasuk and W. Kurutach. Tattoo Skin Detection and Segmentation Using Image Negative Method. In Communications and Information Technologies (ISCIT), 13th International Symposium on, pages 354 359, Sept 2013. [3] R. W. Floyd. Algorithm 97: Shortest path. Communications of the ACM, 5(6):344 348, June 1962. [4] H. Han and A. Jain. Tattoo Based Identification: Sketch to Image Matching. In Biometrics (ICB), International Conference on, pages 1 8, June 2013. [5] N. Q. Huynh, X. Xu, A. W. K. Kong, and S. Subbiah. A Preliminary Report on a Full-Body Imaging System for Effectively Collecting and Processing Biometric Traits of Prisoners. In 2014 IEEE Symposium on Computational Intelligence in Biometrics and Identity Management (CIBIM), pages 167 174. IEEE, 2014. [6] A. Jain, R. Jin, and J.-E. Lee. Tattoo Image Matching and Retrieval. IEEE Computer, 45(5):93 96, May 2012. [7] J. Kim, A. Parra, H. Li, and E. J. Delp. Efficient Graph- Cut Tattoo Segmentation. In SPIE/IS&T Electronic Imaging, pages 94100H 94100H. International Society for Optics and Photonics, 2015. [8] J. Kim, A. Parra, J. Yue, H. Li, and E. J. Delp. Robust Local and Global Shape Context for Tattoo Image Matching. In Image Processing (ICIP), 2015 IEEE International Conference on, pages 2194 2198. IEEE, 2015. [9] J. Kim, J. Yue, and E. J. Delp. Tattoo Image Retrieval for Region of Interest. In Technologies for Homeland Security (HST), 2016 IEEE International Symposium on. IEEE, 2016. [10] J. Kim, J. Yue, J. Ribera, E. J. Delp, and L. Huffman. Automatic and Manual Tattoo Localization. In Technologies for Homeland Security (HST), 2016 IEEE International Symposium on, pages 000 000. IEEE, 2016. [11] J.-E. Lee, R. Jin, A. Jain, and W. Tong. Image Retrieval in Forensics: Tattoo Image Database Application. MultiMedia, IEEE, 19(1):40 49, Jan 2012. [12] D. Manger. Large-Scale Tattoo Image Retrieval. In Computer and Robot Vision (CRV), Ninth Conference on, pages 454 459, May 2012. [13] K. Sridharan, S. Nayak, S. Chikkerur, and V. Govindaraju. A Probabilistic Approach to Semantic Face Retrieval System. In Audio- and Video-Based Biometric Person Authentication, volume 3546, pages 977 986, 2005. [14] B. J. Wing. ANSI NIST-ITL Data Format for the Interchange of Fingerprint, Facial and Other Biometric Information, 2013. [15] H. Yi, P. Yu, X. Xu, and A. W. K. Kong. The Impact of Tattoo Segmentation on the Performance of Tattoo Matching. In 2015 IEEE International WIE Conference on Electrical and Computer Engineering (WIECON-ECE), pages 43 46. IEEE, 2015.