Shoeprint image retrieval and crime scene shoeprint image linking by using convolutional neural network and normalized cross correlation

A shoeprint image retrieval process aims to identify and match images of shoeprints found at crime scenes with shoeprint images from a known reference database. It is a challenging problem in the forensic discipline of footwear analysis because a shoeprint found at the crime scene is often imperfect. Recovered shoeprints may be partial, distorted, left on surfaces that do not mark easily, or perhaps come from shoes that do not transfer marks easily. In this study, we present a shoeprint retrieval method by using a convolutional neural network (CNN) and normalized cross-correlation (NCC). A pre-trained CNN was used to extract features from the pre-processed shoeprint images. We then employed NCC to compute a similarity score based on the extracted image features. We achieved a retrieval accuracy of 82% in our experiments, where a “successful” retrieval means that the ground truth image was returned in the top 1% of returned images. We also extend our shoeprint retrieval method to the problem of linking shoeprints recovered from crime scenes. This new method can provide a linkage between two crime scenes if the two recovered shoeprints originated from the same shoe. This new method achieved a retrieval accuracy of 88.99% in the top 20% of returned images.     

The Use of Scale‐Invariance Feature Transform Approach to Recognize and Retrieve Incomplete Shoeprints

Shoeprints left at the crime scene provide valuable information in criminal investigation due to the distinctive patterns in the sole. Those shoeprints are often incomplete and noisy. In this study, scale‐invariance feature transform is proposed and evaluated for recognition and retrieval of partial and noisy shoeprint images. The proposed method first constructs different scale spaces to detect local extrema in the underlying shoeprint images. Those local extrema are considered as useful key points in the image. Next, the features of those key points are extracted to represent their local patterns around key points. Then, the system computes the cross‐correlation between the query image and each shoeprint image in the database. Experimental results show that full‐size prints and prints from the toe area perform best among all shoeprints. Furthermore, this system also demonstrates its robustness against noise because there is a very slight difference in comparison between original shoeprints and noisy shoeprints.     

The interpretation of shoeprint comparison class correspondences

The underlying principles involved in the interpretation of shoeprint comparisons have become a topical subject due to criticisms in the 2009 National Academy of Science (NAS) report on forensic sciences [1]. Difficulties in the application and understanding of these principles were also highlighted in a recent court ruling [2–5]. We report here a survey that may inform some aspects of this interpretation and discuss the implications of findings from this survey in the light of that court ruling and more importantly the NAS report.Five hundred shoeprints taken from student volunteers in Auckland, New Zealand were compared against each other for the presence of any pattern correspondences. Comparisons were undertaken of the full outsole and of smaller portions of the more common patterns.Of the 500 shoe impressions collected 488 (97.6%) were ultimately represented only once in the survey. The greatest number of corresponding patterns was for the most common brand of shoe (Converse Chuck Taylor All Star) and occurred in 3 of 500 observations. No instances of an imitation brand matching the authentic brand were found. Smaller sections of the common patterns showed a greater number of corresponding prints. However, the greatest number of matching partial patterns was again for the most common brand of shoe (Converse Chuck Taylor All Star) and occurred in 29 of 500 observations.We conclude that pattern match alone is of considerable evidential value even when the print is partial.     

足迹信息的计算机管理与查询

目的　根据国内足迹计算机管理系统的现状和需求 ,研制了适合我国刑侦工作特点的足迹管理与查询系统。方法　通过建立鞋厂鞋样、现场足迹和犯罪嫌疑人足迹数据库 ,利用足迹分析与比对软件进行管理和查询。结果　该软件完全能实现足迹管理计算机化 ,降低技术人员的劳动强度 ,提高足迹信息的使用率。结论　该软件使用方便快捷 ,适合推广使用     

Geographical variation of shoeprint comparison class correspondences

The underlying principles involved in the interpretation of shoeprint comparisons have become a topical subject due to criticisms in the 2009 National Academy of Science (NAS) report on forensic sciences[1]. Difficulties in the application and understanding of these principles were also highlighted in a recent court ruling [2–5] and subsequent discussion of the ruling. We report here a survey that may inform some aspects of this interpretation and discuss the implications of findings from this survey in the light of that court ruling and more importantly the NAS report.1,511 shoeprints were taken from student volunteers in Auckland, Wellington and Dunedin, New Zealand. 500 shoeprints were sampled from student volunteers at Australian universities. 100 from each of the University of Technology in Sydney, University of Queensland in Brisbane, University of Newcastle, Charles Sturt University in Bathurst and University of Canberra, Australia. These cities are distributed along the east coast of Australia.The shoeprints, taken from each country, were compared against each other for the presence of any pattern correspondences However shoeprints have not been compared between countries.In all locations the pattern of some common and many rare outsole patterns was repeated, with Converse Chuck Taylor All Stars and Vans Canvas Era common in all locations.     

Robust Shoeprint Retrieval Method Based on Local‐to‐Global Feature Matching for Real Crime Scenes

In this study, an automatic and robust crime scene shoeprint retrieval method is proposed. As most shoeprints left at crime scenes are randomly partial and noisy, crime scene shoeprint retrieval is a challenging task. To handle partial, noisy shoeprint images, we employ denoising deep belief network (DBN) to extract local features and use spatial pyramid matching (SPM) to obtain a local‐to‐global matching score. In this study, 536 query shoeprint images from crime scenes and a large scale database containing 34,768 shoeprint images are used to evaluate the retrieval performance. Experimental results show that the proposed method outperforms other state‐of‐the‐art methods in terms of retrieval accuracy, feature dimension, and retrieval speed. The proposed method achieves a cumulative match score (CMS) of 65.67% at top 10 which is 5.60% higher than the second best performing method.     

A novel technique for automatic shoeprint image retrieval

Recent developments in forensic science have resulted in large numbers of scene of crime images being collected for recording and analysis. Shoeprint images are no exception. In fact, these have recently been of great interest to police and forensic scientists as footwear evidence is now treated in the same manner as fingerprint and DNA evidence. Traditional approaches to shoeprint representations attempt to classify shoeprint images based on a number of possible patterns. Such approaches are difficult to implement in an automatic fashion without the intervention of a forensic specialist. This paper presents a robust algorithm for shoeprint matching based on Hu's moment invariants. It is shown that decreasing the resolution of images does not have a significant effect on the performance of the algorithm. It is also shown that the optimal performance of the proposed system is attained for images rotated by any angle.     

Developing a Spatial-Temporal Method for the Geographic Investigation of Shoeprint Evidence

Abstract:  This article examines the potential of a spatial-temporal method for analysis of forensic shoeprint data. The large volume of shoeprint evidence recovered at crime scenes results in varied success in matching a print to a known shoe type and subsequently linking sets of matched prints to suspected offenders. Unlike DNA and fingerprint data, a major challenge is to reduce the uncertainty in linking sets of matched shoeprints to a suspected serial offender. Shoeprint data for 2004 were imported from the Greater London Metropolitan Area Bigfoot database into a geographic information system, and a spatial-temporal algorithm developed for this project. The results show that by using distance and time constraints interactively, the number of candidate shoeprints that can implicate one or few suspects can be substantially reduced. It concludes that the use of space-time and other ancillary information within a geographic information system can be quite helpful for forensic investigation.     

Determining Shoe Length from Partial Shoeprints

The length of a shoe has significant value in reflecting characteristics of the owner, and thereby, it can help in tracking suspects in criminal cases. However, the shoeprints left at a crime scene are often incomplete, resulting in difficulties in assessing shoe length. To find a way to estimate the shoe length from a partial shoeprint, 109 shoes with different sizes and general patterns were collected, and their prints were lifted using magnetic powder. Four feature points were defined on a shoeprint, and the longest distance between the feature points was defined as shoe length. Using linear, quadratic, and cubic regression analyses, a total of 15 equations were obtained between the shoe length and the other distances between the feature points. Out of these, the five most accurate equations were selected as the optimal equation. The verified test, including another 18 pairs of shoes, showed an average error of equations between 0.591 cm and 0.732 cm. The equations were also applied in two practical cases, resulting in good accuracy. The study demonstrates that shoe length could be determined from partial shoeprints through the proven equations.     

The role and duties of the shoeprint/toolmark examiner in forensic laboratories

Shoeprint and toolmark examinations have been carried out for forensic purposes for a great number of years. According to literature, the first case of footwear identification dates back to the year 1786. This particular case was worked out by a local police constable. Who then, are the persons working in crime laboratories and specialising in shoeprint and toolmark examinations today? Are they responsible for some other kind of duties, and if ‘yes’, what kind of duties? To find answers to these questions, a questionnaire was prepared and sent to 78 forensic laboratories, 62 laboratories in 27 European countries and 16 laboratories in six countries outside Europe. Forty-three laboratories (55%) in 23 countries; 35 laboratories in 19 European countries and eight laboratories in four countries outside Europe returned the questionnaire. The conducted study shows that the shoeprint and toolmark examiners are performing a great variety of tasks besides shoeprint and toolmark examinations, and also that the tasks vary a lot from one laboratory to another.     

Reference ballistic imaging database performance

Ballistic imaging databases allow law enforcement to link recovered cartridge cases to other crime scenes and to firearms. The success of these databases has led many to propose that all firearms in circulation be entered into a reference ballistic image database (RBID). To assess the performance of an RBID, we fired 4200 cartridge cases from 600 9 mm Para Sig Sauer model P226 series pistols. Each pistol fired two Remington cartridges, one of which was imaged in the RBID, and five additional cartridges, consisting of Federal, Speer, Winchester, Wolf, and CCI brands. Randomly selected samples from the second series of Remington cartridge cases and from the five additional brands were then correlated against the RBID. Of the 32 cartridges of the same make correlated against the RBID, 72% ranked in the top 10 positions. Likewise, of the 160 cartridges of the five different brands correlated against the database, 21% ranked in the top 10 positions. Generally, the ranking position increased as the size of the RBID increased. We obtained similar results when we expanded the RBID to include firearms with the same class characteristics for breech face marks, firing pin impressions, and extractor marks. The results of our six queries against the RBID indicate that a reference ballistics image database of new guns is currently fraught with too many difficulties to be an effective and efficient law enforcement tool.     

Reference ballistic imaging database performance

Ballistic imaging databases allow law enforcement to link recovered cartridge cases to other crime scenes and to firearms. The success of these databases has led many to propose that all firearms in circulation be entered into a reference ballistic image database (RBID). To assess the performance of an RBID, we fired 4200 cartridge cases from 600 9mm Para Sig Sauer model P226 series pistols. Each pistol fired two Remington cartridges, one of which was imaged in the RBID, and five additional cartridges, consisting of Federal, Speer, Winchester, Wolf, and CCI brands. Randomly selected samples from the second series of Remington cartridge cases and from the five additional brands were then correlated against the RBID. Of the 32 cartridges of the same make correlated against the RBID, 72% ranked in the top 10 positions. Likewise, of the 160 cartridges of the five different brands correlated against the database, 21% ranked in the top 10 positions. Generally, the ranking position increased as the size of the RBID increased. We obtained similar results when we expanded the RBID to include firearms with the same class characteristics for breech face marks, firing pin impressions, and extractor marks. The results of our six queries against the RBID indicate that a reference ballistics image database of new guns is currently fraught with too many difficulties to be an effective and efficient law enforcement tool.     

Computer controlled scanner for laser enhanced photography

The use of a scanner to control the photography of fingerprints, shoeprints and other trace contacts which fluoresce under irradiation from an argon ion laser is described. This scanner has been built to cover areas varying in size from a single fingerprint to a shoeprint and provides for variable speed, step size and angle of radiation which are all very important in this highly specialised area of photography.     

Removing interfering contaminations from gelatin lifters

Gelatin lifter is widely used for recovering shoeprints from crime scenes. Dusty shoeprints removed from paper with loose fibers, cloth or plasterboard, might be concealed by the detached fibers. A novel technique to clean the gelatin lifter from the interfering contaminations, using adhesive lifters, was developed. The adhesive lifter is applied directly on the surface of the gelatin lifter, and is removed instantly. The adhesive lifter removes the upper layer of the attached material on the gelatin lifter. After removing the concealing material from the hidden imprint, the quality of the visualized imprint is improved. The cleaning process can be applied as many times as needed and the optimum number is different for each substrate. The small dust particles comprising the shoeprint remained attached to the sticky side of the gelatin lifter even though the adhesion force applied by the adhesive lifter is great. Repeating the procedure too many times could harm the quality of the print severely, yet the advantages of this method are greater than the risks of relocation of image quality reduction.     

Shoe wear patterns and pressure distribution under feet and shoes, determined by image analysis.

A method is outlined for extracting contours of the pressure exerted by feet on the ground, either directly or through the soles of shoes, and of the wear patterns on shoe soles. The method is applied to two types of image: shoeprint images made in the laboratory, and pedobarography images recording the distribution of pressure under the foot during a footstep. The first procedure in extracting the information from images involves low pass filtering in the Fourier domain, performed by a microcomputer and frame-grabbing system. The boundaries of the regions in the resultant images are then extracted from the filtered images to display the shape of constant pressure areas and of the general shoe-wear.     

A new retrieval system for a database of 3D facial images

A new retrieval system for a 3D facial image database was designed and its reliability was experimentally examined. This system has two steps, firstly to automatically adjust the orientation of all 3D facial images in a database to that of the 2D facial image of a target person, and then to identify the facial image of the target person from the adjusted 3D facial images in the database using a graph-matching method. From the experimental study [M. Yoshino, K. Imaizumi, T. Tanijiri, J.G. Clement, Automatic adjustment of facial orientation in 3D face image database, Jpn. J. Sci. Tech. Iden. 8 (2003) 41-47], it is concluded that the software developed for the first step will be applicable to the automatic adjustment of facial orientation in the 3D facial image database. In 28 out of 110 sets (25.5%), the 3D image of the target person was chosen as the best match (from a database of 132 3D facial images) according to the similarity of the facial image characteristics based on the graph matching. The 3D facial image of the target person was ranked in the top of 10 of the database in 75 out of 110 sets (68.2%). These results suggest that this system is inadequate for the identification level, but may be feasible for screening method in a small database. It will be necessary to further pursue the possibility of realization of a facial image retrieval system for a large database such as suspects' facial images in future.     

Automated detection of regions of interest in cartridge case images using deep learning

This paper explores a deep-learning approach to evaluate the position of circular delimiters in cartridge case images. These delimiters define two regions of interest (ROI), corresponding to the breech face and the firing pin impressions, and are placed manually or by an image-processing algorithm. This positioning bears a significant impact on the performance of the image-matching algorithms for firearm identification, and an automated evaluation method would be beneficial to any computerized system. Our contribution consists in optimizing and training U-Net segmentation models from digital images of cartridge cases, intending to locate ROIs automatically. For the experiments, we used high-resolution 2D images from 1195 samples of cartridge cases fired by different 9MM firearms. Our results show that the segmentation models, trained on augmented data sets, exhibit a performance of 95.6% IoU (Intersection over Union) and 99.3% DC (Dice Coefficient) with a loss of 0.014 for the breech face images; and a performance of 95.9% IoU and 99.5% DC with a loss of 0.011 for the firing pin images. We observed that the natural shapes of predicted circles reduce the performance of segmentation models compared with perfect circles on ground truth masks suggesting that our method provide a more accurate segmentation of the real ROI shape. In practice, we believe that these results could be useful for firearms identification. In future work, the predictions may be used to evaluate the quality of delimiters on specimens in a database, or they could determine the region of interest on a cartridge case image.     

Lifting shoeprints using gelatin lifters and a hydraulic press

A method for lifting two-dimensional dust footwear marks on rough or porous surfaces, such as cardboard or cloth, using a hydraulic press, was examined. It was found that exerting pressure on the lifter by the press usually improves the quality of the results. When the shoeprints were on rough or soft surfaces, the prints transferred to the gelatin lifters were better than those obtained by the conventional method. In other cases, using the press did not improve the results but was much simpler to apply. Based on the results of this study, the hydraulic press/gelatin lifter method (the "press method") is used at the authors' laboratory, depending on the surface from which the shoeprint is to be lifted. It is the authors' intention to apply the method to other surfaces after finding the optimal pressure for surfaces with loose fibers.