# Perceptual Colour-based Geolocation of Human Trafficking Images for Digital Forensic Investigation

## Full Text

This Markdown is an automated text extraction for search and reading tools. The [hosted PDF](PerceptualColour-BasedImageGeolocation.pdf) is the authoritative publication source.

Source PDF: [PerceptualColour-BasedImageGeolocation.pdf](PerceptualColour-BasedImageGeolocation.pdf)
Source SHA-256: `dc7ba2d87fac981f7d8969d4bcae3bede8756111ee8060551d499d0be68c4151`
Generated with `pdftotext -nopgbrk`; formatting and reading order may differ from the PDF.

---

Perceptual Colour-based Geolocation of Human
Trafficking Images for Digital Forensic Investigation
Jessica Herrmann∗ , Opeyemi Bamigbade† , John Sheppard† , Mark Scanlon∗
∗ Forensics and Security Research Group, School of Computer Science, Univeristy College Dublin, Ireland.

herrmann.jessica@protonmail.com, mark.scanlon@ucd.ie
† Department of Computing and Mathematics, South East Technological University, Waterford, Ireland.

opeyemi.bamigbade@postgrad.wit.ie, john.sheppard@setu.ie

Abstract—This paper investigates the effectiveness of colourbased descriptors in Content-Based Image Retrieval (CBIR) and
examines the impact of various parameters on image matching
accuracy. The aim is to improve image retrieval methods to
support digital forensic investigators in human trafficking cases.
Colour values are used as key components to describe specific
image characteristics, and the technique is evaluated on the
Hotels-50K dataset. The method achieved a Top-50 accuracy
of over 95%, enabling efficient data triage and significantly
reducing the volume of images to be examined. Using 2 colour
descriptors is found to optimise the balance between information
richness and dimensionality reduction. Performance is further
improved by optimised image selection, reducing false-positive
rates, and increasing robustness. The approach demonstrates
potential in advancing image analysis tools in human trafficking
investigations and other contexts, opening new avenues for using
colour values in crime detection and image data analysis. Future
research may refine the Euclidean distance method used in the
image similarities measure by introducing weighted distance
measurements to reduce the impact of common colour values,
and investigate lighting and saturation effects.
Index Terms—Content-Based Image Retrieval (CBIR), Colour
Descriptors, Computer Vision, Human Trafficking, Artificial
Intelligence, Multimedia Geolocation

I. I NTRODUCTION
Human trafficking, a complex and serious crime, affects millions of people worldwide, transcending age, gender, and background, and causes severe personal, communal, and societal
harm. It involves the illegal trade of individuals through deception, violence, extortion, or exploitation, leading to various
forms of abuse, including forced labour, sexual exploitation,
and organ trafficking. The European Union employs “Directive
2011/36/EU” to combat this crime, setting rules for offences
and sanctions, prioritising prevention, and victim protection
with a gender perspective [1].
European data collection on trafficking progress occurs
biannually, revealing deficits in data and technology-assisted
investigative methods. Identification of trafficking victims typically relies on location analysis, phone numbers, and textual
content, while image analysis, especially involving image
content, is underused or limited to exact matches. Stylianou
et al. [2] underscored the importance of images in fighting
human trafficking and the potential of Content-Based Image
Retrieval (CBIR). CBIR uses visual characteristics to analyse
and find similarities in images, employing advanced technologies such as neural networks. Despite practical challenges

in data collection and processing, the creation of specialised
datasets, e.g., Hotels-50K [3] and Hotel-ID [4], has driven
significant research progress in image matching for potential
crime locations. Similar hotel images are often posted by
Europol’s “Stop Child Abuse – Trace an Object” project to
crowdsource the location1 . In many cases, the section of the
images that contain the victim is hidden to protect privacy and
comply with legal regulations.
The increasing volume of data underscores the need for AIbased solutions, particularly for digital forensic laboratories
that handle large datasets [5]. However, artificial intelligence
(AI) technology remains in its early stages, with significant
gaps between research on analysis methods and practical
applications for law enforcement [6]. These tools must not
only produce valid results, but also comply with legal, ethical,
and practical standards [7].
The ethical considerations surrounding AI have been formally recognised at the EU level with the enactment of the
AI Act, which came into force in August 2024 [8]. This
comprehensive legislation aims to ensure that the benefits of
AI are realized while adhering strictly to ethical principles and
safety standards. The AI Act establishes a unified regulatory
framework that promotes the development of secure and
trustworthy AI systems and regulates their responsible use
across the private and public sectors.
There is a growing trend within the digital forensic community toward approaches aided by AI [6, 9], indicating that
intelligently automated digital evidence processing is gaining
in popularity as a technique [10]. In an environment of increasing data volumes, CBIR shows potential for improving image
analysis during digital investigation. Specifically, the use of
colour as a key feature in CBIR [11]. By incorporating CBIR
into investigations, law enforcement can better analyse photos
and videos, revealing patterns linked to human trafficking that
might be missed by traditional methods.
This paper contributes to CBIR research by deeply exploring
the role of colour in content-based image matching. Through
empirical analysis using the Hotels-50K dataset, the study
highlights the effectiveness of colour features in distinguishing
and retrieving relevant images. These findings have practical
implications for visual content analysis, particularly in human
1 https://www.europol.europa.eu/stopchildabuse

trafficking investigations, offering insights for the design and
implementation of CBIR systems that leverage colour information.
A. Research Questions
This research evaluates a colour matching procedure for
hotel image identification within the Hotels-50K dataset, focussing on its accuracy in assigning images to the correct
hotels based on extracted colour values. The study examines
the optimal number of colour values for maximising matching
accuracy and precision, with the aim of assessing the procedure’s performance and providing insights for improving
image-based hotel identification.
The research focusses on the following questions:
How effective and accurate can a matching method based
on colour values identify hotel rooms in a large dataset
of images from different sources?
• What influence does the number of extracted colour
values have on the accuracy and precision of the colour
descriptors in the matching procedure?
• What influence can colour values as hard descriptors have
on reducing the search field in the identification of hotel
rooms?

•

A hierarchical analysis of the Top-K accuracy at a 95%
threshold is intended to achieve data triage, effectively narrowing the search field and enabling a more targeted image
analysis. In this context, ’hard descriptors’ refer to specific and
well-defined features that provide consistent information about
an image’s color composition, allowing for precise matching
and identification.
II. L ITERATURE R EVIEW
A. Hotel Image Datasets
Indoor scene recognition presents unique challenges, and
existing methods developed for outdoor scenes, such as landmark recognition, are not well suited to distinguish between
similar images [3]. In 2015, Stylianou et al. [2] addressed
these specific challenges with the main objective of supporting
investigations and prosecutions in the context of criminal
activities related to sex trafficking. Here, identification of hotel
rooms served as a key link to the locations of the victims [2].
For this purpose, the researchers created a database of hotel
room images used for marketing and travel sales, e.g., via the
Expedia Affiliate Network API, and developed a smartphone
crowd-sourcing app called “TraffickCam”, through which travellers could upload their own images of hotel rooms [2].
From the combined approaches, the Hotels-50K dataset
was released in 2019, which consists of a total of 1,027,871
images from 50,000 unique hotels around the world [3].
However, travel website images, which are often professionally
taken under ideal conditions, are often visually very different
from investigation images. In contrast, crowd-sourced images,
although less frequent, have more similarities to images from
genuine queries [3].

B. Related Work
In 2015, Stylianou et al. [2] compared SIFT matching with
Convolutional Neural Networks (CNNs) for hotel room image
matching. SIFT performed well in exact matches, but faced
limitations due to a small dataset and lack of real-world variation. In 2019, Stylianou et al. [12] improved results with the
ResNet-50 model, enhancing hotel room image recognition.
In 2020, Xuan et al. [13] introduced “easy positive” mining
to generate embeddings, aligning similar images and distancing dissimilar ones. A 2021 study expanded this by optimising
hard negative triplets in deep metric learning, acknowledging
their importance where distance metrics fail to capture semantic similarity. Xuan et al. [14] introduced a novel loss function
and a “triplet diagram” to improve optimisation.
Subsequent research focused on hotel recognition, developing a loss function adapted for better hotel representation [15].
In 2021, Kamath et al. [4] combined Convolutional Neural
Networks (CNNs) with Graph Neural Networks (GNNs) for
a more comprehensive image analysis, while a study in 2022
by Black et al. [16] focused on interpretability, introducing
visualisation methods to identify image similarity in police
investigations.
Saikia et al. [17] highlighted the value of colour-based
image search for its rich information and invariance to pixel
shifts and rotations, also emphasising the importance of texture
features in complex scenarios. Their algorithm combined local
and global scene content to create robust descriptors for colour,
texture, and perspective changes, using CNN features from
various colour spaces.
Compared to existing methods that aim for exact matches,
this research focusses specifically on investigating a colourbased approach that aims to identify optimal parameters for
hard descriptors. The objective is to use colour descriptors
to provide effective data triage, which is particularly valuable
for police practice. In many forensic scenarios, effective data
triage is essential, as more complex models are time- and
resource-intensive. By reducing the image volume, the use
of complex models can be focused specifically on the most
relevant images.
C. Content-Based Image Retrieval
As image collections grow, CBIR plays a crucial role in the
efficient classification and retrieval of digital images by evaluating visual attributes like colour, texture, and shape, making it
ideal for searching large image databases. Its applications span
fields such as medical imaging, geoinformation, and biodiversity research [18, 19, 20]. However, CBIR techniques must be
tailored to each field’s unique needs. In forensic applications,
CBIR methods often encounter limitations due to factors such
as noise, lighting variations, and fluctuating image quality,
which impair their accuracy. Techniques like SIFT or CNNs,
which perform well under controlled conditions, tend to lose
reliability in real-world investigations and frequently result in
false positives. Colour values have proven to be more robust
against variations in lighting and image quality, providing a
method to capture image similarities more accurately even in

demanding forensic scenarios. For example, medical images
are often high-resolution 3D, while surveillance images are
low-resolution and noisy. Accuracy requirements also vary,
with fields such as medical imaging and forensic analysis
demanding high precision, particularly in legal contexts.
In law enforcement, ensuring interpretability and reliable
documentation is vital [5]. Legal proceedings, especially the
preservation of evidence, introduce challenges that differ between legal systems. The Artificial Intelligence Act of the EU
recognises the importance of AI in law enforcement and its
potential impact on fundamental rights [21]. Combining handcrafted descriptors with deep learning approaches can enhance
CBIR performance, offering flexibility, lower computational
needs, and improved interpretability, making them valuable
supplements to CNN-extracted features.
1) CBIR Global and Local Features: CBIR feature extraction methods are generally divided into global and local
approaches. Global methods extract features from the entire
image, while local methods focus on smaller sections, making
them more robust to geometric deformations and lighting
changes.
Global methods rely on statistical properties such as average
hue, colour variance, and features such as mean, standard
deviation, or variance of colour values. Although efficient,
this approach may miss finer image details. Histograms, commonly used to assess colour distribution, are robust to pixel
shifts and rotation, focusing on relative colour distribution
rather than absolute values [17]. This robustness addresses the
challenges encountered in previous scene-based approaches
using the Hotel-50K dataset [4]. Histograms can be created for
individual colour channels or combined into a single channel.
Local methods, on the other hand, capture colour information in specific image regions. Colour textures describe
characteristic features of different areas, offering greater accuracy and discrimination. Despite the success of colour
histograms and shape descriptors in certain applications, they
face both theoretical and practical issues, particularly the curse
of dimensionality [22].
III. M ETHODOLOGY
This study follows a two-step approach to optimize and
evaluate image matching parameters. First, empirical observation is conducted using a subset of the Hotels-50K dataset to
identify optimal colour parameters for the matching procedure.
Various parameter configurations are tested, with the selection
guided by systematic analysis.
Next, the identified colour parameters are applied in a CBIR
experiment. The performance of the matching procedure is
evaluated by analysing the impact of the selected parameters
on retrieval accuracy, as detailed in Section IV. The results are
documented for each hotel in the dataset, providing insight into
how the choices of colour parameters influence the quality of
the image matching.
The empirical study will test the following hypothesis:
• The number of extracted colour values significantly impacts the accuracy of colour descriptors in the matching

procedure, with the assumption that increasing the number of values improves matching accuracy.
• Matching performance is directly correlated with data
quality, with the assumption that higher-quality data significantly enhances matching outcomes.
• Colour values as hard descriptors are hypothesised to
effectively narrow the search space, thereby restricting
potential matches and improving search efficiency.
The results of the matching will be evaluated and compared
with the hypotheses. This will encompass the precision of the
matching and any deviations from the anticipated values.
IV. E MPIRICAL D ESIGN
This analysis outlines the CBIR process and establishes the
basis for the planned experiments. Insights from this chapter
inform the systematic variation of different parameters in later
experiments. The detailed analysis and results’ presentation
will serve as the groundwork for subsequent analyses and
optimisations within this work.
A. Content Based Image Retrieval Process
For any problem in CBIR, the solution starts with image
analysis and feature definition. This paper utilises colour features extracted from images in the dataset subset to facilitate
image matching. Colour descriptors are used as the primary
feature for representation. These descriptors are integral to
limiting the search space and improving the precision of the
matching procedure.
The performance of CBIR is evaluated using the top-k
retrieval metric, where the system returns the k most similar images ranked by colour similarity. The top-k accuracy
indicates the proportion of relevant matches among the first
k results, serving as a key measure of the effectiveness of
retrieval and the impact of the chosen colour parameters on
system performance [22].
B. Dataset Configurations
The data_source level of the dataset contains two
folders. The travel_website folder, which contains a collection of images from hotel websites, and the traffickcam
folder, which contains user-contributed images from TraffickCam. Index CSV files are available for the identification of
the individual images. The chain_id describes the hotel
chain, so that the respective subfolder with the corresponding
hotel_id can be assigned to a certain chain.
The hotels examined represent diverse hotel chains,
showcasing a range of characteristics. The number of
available images within the travel_website and
traffickcam folders exhibited considerable variation. In
travel_website folders, images ranged from 9 to 24,
while traffickcam folders contained 1 to 8 images. As
anticipated, travel_website images exhibited higher
quality and more professional presentation compared to those
in traffickcam folders, as can be seen in Figure 1. The
latter often featured foreign objects and hotel guests in the
pictures.

results, while K-Means showed slight deviations. Thus, the
more robust Median-Cut was selected for further testing.

Fig. 1. On the left side: Image from the TraffickCam folder of hotel 4824,
On the right side: Three images from the travel websites folder of hotel 4824

Empirical observation was carried out with a “5-hotels”
subset consisting of five randomly selected hotel folders.
C. Image Preparation
Following the examination of the dataset, the initial step
in CBIR involves image preparation. A potential approach
is to transform colour representation into alternative colour
spaces, e.g., RGB to HSV or HLS. These spaces separate the
brightness and saturation from the pure colour value, allowing
for colour similarity assessment based on independent colour
properties, unaffected by brightness or saturation variations.
An investigation into whether considering brightness or saturation as individual values may lead to insignificant differences
in brightness that affect similarity ranking is deferred. The
primary focus now is on the number of extracted colour values
and their impact on matching precision. As a result, images
are retained in their original RGB colour space.
D. Colour Palette Extraction Technique
The colour value extraction was extracted using Pylette2 ,
which extracts colour palettes from images using various
clustering methods. In addition to RGB, HLS and HSV colour
modes, Pylette supports colour quantisation using the K-Means
and Median-Cut algorithms, both of which can be used to
cluster similar data points. The purpose of clustering is to
identify and consolidate patterns in the data. K-Means is a
general clustering algorithm that is used in various fields of
application to organise data into predefined clusters. Each data
point is assigned to the nearest cluster centre. K-Means is
widely used in data analysis, machine learning, and colour
quantisation.
Hu et al. [23] compared these methods in 2013, concluding
that K-Means offers better palette quality and computational
efficiency, although it is more susceptible to variations due to
initialisation [23]. Tests on the “5-hotels” subset confirmed
this, with the Median-Cut algorithm producing consistent

Fig. 3. Hotel 398. On the left: Image 6263111.jpg from the travel website
folder, on the right: Colour palette containing 5, 10 and 20 values.

The number of colour values to be extracted can be set
manually in Pylette. Initially, 5, 10 and 20 colour values were
extracted per image.
Figures 2 and 3 suggest that specific colour details in
palettes become distinctly visible with an extraction of at
least 20 colour values or more. For example, in Figure 2
the pink of the cushion becomes prominently visible only
with a higher number of extracted colour values, and the blue
tones are reproduced more faithfully. In contrast, the second
example image in Figure 3 exhibits barely perceptible differences compared to palettes with fewer extracted colour values,
despite increased extractions. This palette primarily comprises
various brown tones, which could explain the minimal visual
distinctions in the created colour palettes, even with a higher
number of extracted colour values.
A 2D scatter plot was used to visualise the extracted
colour values, similar to histograms in representing the colour
distribution but plotting individual colour values as points
in a coordinate system. This allows for detecting shifts or
deviations in colour values.
With Matplotlib3 , the extracted RGB values were used to
create the scatter plots. NumPy4 was used for array manipulation, and the RGB values were scaled to the range [0, 1] for
consistency. The scatter plots depicted the R and G values on
the corresponding axes within this range.
The scatter plots shown in Figure 4, depicting 5, 10,
and 20 extracted colour values, reveal interesting patterns
regarding the variety and distribution of colours. With 20
extracted values, the scatter plots exhibit a larger dispersion,
indicating greater variation and colour diversity. This broadens
the representation of image content. Conversely, scatter plots
with only five extracted colour values display limited colour
variety, especially challenging for images within similar colour
ranges. These plots may not capture subtle colour differences,
3 https://matplotlib.org/

2 https://github.com/qTipTip/Pylette

4 https://numpy.org/

Fig. 2. Hotel 451. On the left: Image 6000659.jpg from the travel website folder, and on the right: Colour palette containing 5, 10 and 20 values.

Fig. 4. Hotel: 451. On the left : Scatter Plot with 5 values extracted per
image, in the middle: Scatter Plot with 10 values extracted per image, on the
right side: Scatter Plot with 20 values extracted per image.

hindering the detection of nuances in images with analogous
colour palettes.
In scatter plots with more extracted colour values, clustering
or concentration in specific colour spectra is noticeable, implying the frequent occurrence or dominance of particular colour
ranges. But these clusters can exacerbate the “curse of dimensionality”, which shortens distances in higher-dimensional
space, making it harder to distinguish similar images. In
summary, the choice of the number of extracted colour values
requires careful consideration to balance the distinguishability
of similar images and the representativeness of the palette.
E. Visual Feature matching
To visually compare individual plots, a three-dimensional
plot was generated using the colour palette of Hotel 451’s
traffickcam folder, with all data points in red. Subsequently, the travel_website folders of different hotels from the “5 hotels” subset were displayed in separate
three-dimensional plots, using the colour green for better
differentiation. To facilitate comparability, data points from
both sets of plots were merged into a combined plot. This
step aims to illustrate the visual differences and similarities between the colour palettes of the traffickcam and
travel_website data sources.
The combined scatter plots reveal significant differences
in the data-point distributions. Despite potential data quality limitations in the traffickcam data, a close correlation can be observed between the outer data points of the
travel_website palette from hotel folder 451 and the
matching plot of the same hotel folder. This is particularly
evident in edge areas, where data points from hotel folder 451
lack nearby counterparts in the middle and right comparison
plots. However, data points in the hotel folder show little distinctiveness in densely populated areas due to agglomeration.
These suggest some consistency and correlation between the
colour values in the different datasets.
F. Calculation of the Smallest Distance between Two Plots
Euclidean distance; a commonly adopted metric for evaluating colour similarity, is used to calculate the similarity
between the colour vectors of the matching image and those in
the image series. It measures the geometric distance between
the colour points in three-dimensional space, where smaller
distances indicate greater similarity, as expressed with the
formula below.
p
Distance = (x2 − x1 )2 + (y2 − y1 )2 + (z2 − z1 )2 (1)

As outlined above, certain frequently occurring colours may
be poor identifiers, while rarer shades are more distinctive. The
weighted distance could reduce clustering effects caused by a
larger number of colour values, helping to distinguish between
outer data points and potential false positives.
The Euclidean distance considers the coordinates (x, y, z)
of the colour values in the three-colour channels (e.g., red,
green, blue) for two data points. For each point in the matching
image, it identifies the closest point from the respective hotel
folder’s image series. This calculation generates a numerical
value that quantifies the similarity of the colour values. A
lower value indicates higher similarity, while a higher value
signifies a greater distance between the colour values.
V. E XPERIMENTAL D ESIGN
A. Dataset Configuration
In the course of the experiment, three additional subsets
were added to the existing dataset “minimum requirements”,
to ensure that the results are robust and not limited to specific
properties or patterns of a single dataset. Each of these subsets
is made up of three individual datasets, each of which, in turn,
comprises 100 different hotel folders.
The datasets, namely “unclean” dataset, “clean” dataset
and “three beds” dataset, are explained in detail below. The
“unclean” datasets include images from the traffickcam
and travel_website folders. These folders were taken
unchanged from the “minimum requirements” subset, resulting
in significant variance in the number of images per folder.
In total, the “three beds” datasets contain 6,562 images from
travel_website and 1,281 from traffickcam.
The “clean” datasets were created through a step-by-step
process. First, images from the traffickcam folders were
removed. The remaining datasets were then cleaned by eliminating images taken outside hotels, images not clearly showing
hotel rooms, and, as shown in Figure 6, pictures of individual
objects or people.

Fig. 6. Sample images that were removed for cleaning.

The remaining images from the travel_website data
were further processed. In a “leave one out” procedure, an
image from each hotel folder was selected for matching,
ensuring that at least one other image in the folder showed
the same setting. Ultimately, 5,384 images remained from the
travel_website dataset, with 300 used for matching. This
approach aimed to maintain the comparability and relevance
of the selected images, considering the diversity of content.
The “three beds” datasets were developed from the “clean”
datasets to optimise image matching. Four images were selected from each hotel folder in the travel_website
data, exclusively depicting beds with similar colouring. Three
images were retained in the folder, while the fourth was set
aside for comparison, as shown in Figure 7.

Fig. 5. Left side: Scatter Plot comparing colour values from hotel 451’s photos from TraffickCam (red) vs. the travel website (green). Middle and right:
Scatter Plot comparing colour values from hotel 451’s photos from TraffickCam (red) vs. hotel 398’s and hotel 611’s photos from the travel website (green),
respectively.

depends on the objective, with critical applications that require
higher thresholds, while others may accept lower ones.

Fig. 7. Example from the three beds datasets. On the left side is the retrieval
image, and on the right side the three images that remain in the Travel
Websites folder.

TABLE I
U NCLEAN DATASETS AVERAGE T OP -K ACCURACY E UCLIDEAN
DISTANCES

B. Method Validation
The developed method is evaluated through supervised
validation against known classes/labels. The primary objective is to verify the accuracy of the method in correctly
matching images to the corresponding hotels. The validation
process involves several steps. First, descriptors are created
for each dataset, with colour palettes containing 1, 2, 5, 10,
and 20 colour values stored in separate CSV files. Then,
precise distance measurements between these colour palettes
are conducted. To ensure the stability of the results, supervised
validation is performed in three separate runs, with each run
using different hotel folders to account for random variations
in the test set. Performance is assessed using Top-K accuracy,
providing an authentic evaluation of the test runs. The focus of
this research is not on maximising precision, but on analysing
the suitability of colour values as descriptors for a data triage.

Table I shows that the first test run with “unclean” datasets
produced suboptimal results for all the extracted colours.
Although the top 100 accuracy met the threshold 95%, the
top 50 accuracy was approx. 50%, indicating that the matching
image was found in the top 50% of hotel images in only half
of the cases. This was largely due to the presence of out-ofscene images, whose colours often differed significantly from
actual hotel room images. The blue spectrum, identified in previous studies as a high-quality identifier, contributed to false
positives due to the inclusion of tourist-related images. These
results confirmed the hypothesis. To resolve this, “clean”
datasets were created by removing out-of-scene images.

VI. R ESULTS
For each test run, tables and diagrams were created to
illustrate the achieved Top-K accuracies in percentages. The
initial test run used the “unclean” datasets, with direct Euclidean distance as the comparison metric. These datasets
included images from both the travel_website and
traffickcam folders, with a minimum of 8 images in
the travel_website folders and at least one in the
traffickcam folder. The datasets reflected the original
Hotels-50K dataset, including both hotel room images and
unrelated scenes. A satisfactory Top-K accuracy was set at
95%, aligning with realistic thresholds. Threshold selection

TABLE II
C LEAN DATASETS AVERAGE T OP -K ACCURACY E UCLIDEAN DISTANCES

The “clean” datasets show significant improvement, as illustrated in Table II. In almost all cases, the 95% threshold was
achieved for Top-70 accuracy. The best results were obtained
using 2 and 20 descriptors, with both predicting with 96% reliability that the matching image is in the top 70%. Interestingly,
2 descriptors performed slightly better. When comparing 2 and
20 descriptors, it is possible that beyond a certain number of
values, concentration centres develop, exacerbating the “curse

of dimensionality.” Using 2 descriptors may provide enough
information for effective matching while minimising false
positives, suggesting a balance between information retention
and dimensionality. The superior performance of 2 descriptors
compared to 5 may be due to more complex dynamics, as more
descriptors do not always lead to better performance. However,
a single descriptor per image appears insufficient.

TABLE III
R ESULTS - THREE BEDS DATASETS AVERAGE T OP -K ACCURACY
E UCLIDEAN DISTANCES

The results from the “three beds” datasets, as shown in
Table III, confirm previous findings, with a top-50 accuracy
exceeding 95% using 2 descriptors. Using 10 descriptors
yielded slightly lower results, with a top-60 accuracy of
96%. Other descriptor values performed significantly worse,
though they still exceeded the top-70 accuracy threshold,
indicating acceptable outcomes. The consistent success of 2
descriptors suggests that more generalised descriptors may
be better suited for this application. With the achieved Top50 accuracy of more than 95% using two descriptors in the
“three beds” dataset, the search field is significantly reduced.
This data triage allows approximately half of the images to be
excluded, thereby decreasing investigative effort and focussing
on relevant matches within the remaining images.
A. Discussion
This study thoroughly investigated the effectiveness of
colour-based descriptors for CBIR, analysing parameters such
as the number of descriptors and distance metrics. The findings
contribute to the development of advanced image search
systems for the digital investigation of human trafficking,
highlighting both strengths and limitations. The research identified the optimal number of descriptors, with 2 offering a
balance between information content and dimension reduction,
enabling efficient image retrieval with high Top-K accuracy.
Furthermore, selective image sorting and dataset sampling improved the robustness of the method, reducing false positives
and improving its practical applicability in various scenarios.
The use of colour descriptors to narrow the search field
enables efficient data triage, significantly reducing the number
of images requiring analysis. This method can be integrated as
a preliminary filter in forensic workflows, making the analysis
more targeted and resource-efficient. Following this narrowing,
texture-based methods could then be applied specifically to the
most relevant images.
Figure 8 provides a clear distinction between the worst
results of the “unclean” datasets, on the left, and the optimal
results of the “three beds” datasets, on the right. The black
diagonal line highlights the threshold that must be reached
in order for the method to have a positive effect over random
(i.e., result lines higher than this demonstrate the benefit of the

Fig. 8. Diagram on the left side: Top-K accuracy results unclean dataset 3,
diagram on the right side: Top-K accuracy results for three beds dataset 2,
both with the Euclidean distance

proposed approach). The chart on the left of Figure 8 shows
that even in the less successful test configurations, at least
neutral results were achieved. On the other hand, the chart on
the right (with the optimal results) illustrates the considerable
added value that can be achieved not only by applying the
proposed method, but also by cleaning up the datasets. Here,
in particular, the study showed that the method is significantly
influenced by the quality of the input images. Images with very
similar colour spectra can affect the accuracy of the matching
and lead to false positives. The same applies to scenes of
foreign images, which have to be sorted out in advance.
VII. C ONCLUSION AND F UTURE W ORK
This research demonstrates that the colour values of the
images significantly contribute to the identification of hotel
rooms in human trafficking and CSAM investigations. Our
findings show that colour serves as a reliable descriptor under
specific conditions, with promising applications for improving
the quality and accuracy of digital investigations. Future work
could focus on refining the quantisation process and exploring
advanced colour extraction techniques to increase robustness.
Although colour alone cannot fully address the complexities
of contextual image analysis, continuous research such as [24]
has demonstrated the value of combining colour with texture
features. Addressing limitations of the Euclidean distance
method, especially in “centres of concentration,” along with
histogram-based weighting, could further refine the colourbased analysis, reducing false positives when used in conjunction with texture data.
Furthermore, this colour-based approach could aid in filtering irrelevant images in CSAM, reducing the size of the
dataset, and improving detailed analysis. Since hash databases
are limited by their inability to identify dynamically generated
images, minimal colour values and edge detection techniques
may improve the classification and filtering of such files. In
conclusion, colour descriptors offer a valuable and flexible tool
for image recognition, search, and classification, capable of
narrowing search fields and providing significant benefits in
tackling complex investigative challenges.
R EFERENCES
[1] “Regulation (COD) No 426/2022 of the European
Parliament and of the Council of 19 December 2022 on

preventing and combating trafficking in human beings
and protecting its victims,” 2022,
https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=
CELEX (Accessed on September 30, 2024).
[2] A. Stylianou, A. Norling-Ruggles, R. Souvenir, and
R. Pless, “Indexing open imagery to create tools to fight
sex trafficking,” in Proceedings of the IEEE Applied
Imagery Pattern Recognition Workshop (AIPR), Washington, DC, USA, 2015, pp. 1–6.
[3] A. Stylianou, H. Xuan, M. Shende, J. Brandt, R. Souvenir, and R. Pless, “Hotels-50k: A global hotel recognition dataset,” Proceedings of the AAAI Conference on
Artificial Intelligence, vol. 33, no. 1, pp. 726–733, 2019.
[4] R. Kamath, G. Rolwes, S. Black, and A. Stylianou, “The
2021 Hotel-ID to Combat Human Trafficking Competition Dataset,” arXiv preprint arXiv:2106.05746, 2021.
[5] X. Du, C. Hargreaves, J. Sheppard, F. Anda,
A. Sayakkara, N.-A. Le-Khac, and M. Scanlon,
“SoK: Exploring the State of the Art and the Future
Potential of Artificial Intelligence in Digital Forensic
Investigation,” in The 13th International Workshop on
Digital Forensics (WSDF), held at the 15th International
Conference on Availability, Reliability and Security
(ARES). ACM, 08 2020.
[6] C. Hargreaves, F. Breitinger, L. Dowthwaite, H. Webb,
and M. Scanlon, “DFPulse: The 2024 Digital Forensic
Practitioner Survey,” Available at SSRN 4954821, 2024.
[7] M. Scanlon, F. Breitinger, C. Hargreaves, J.-N. Hilgert,
and J. Sheppard, “ChatGPT for digital forensic investigation: The good, the bad, and the unknown,” Forensic
Science International: Digital Investigation, vol. 46, p.
301609, 2023.
[8] European Parliament and Council of the European
Union, “Regulation (EU) 2024/1689 of the European
Parliament and of the Council of 13 June 2024
laying down harmonised rules on artificial intelligence,” https://eur-lex.europa.eu/legal-content/EN/TXT/
?uri=CELEX%3A32024R1689, 2024, official Journal of
the European Union, L 1689, published on July 12, 2024,
accessed on November 2, 2024.
[9] F. Breitinger, J.-N. Hilgert, C. Hargreaves, J. Sheppard,
R. Overdorf, and M. Scanlon, “DFRWS EU 10-Year
Review and Future Directions in Digital Forensic Research,” Forensic Science International: Digital Investigation, vol. 48, p. 301685, 03 2024.
[10] X. Du and M. Scanlon, “Methodology for the automated metadata-based classification of incriminating
digital forensic artefacts,” in Proceedings of the 14th
International Conference on Availability, Reliability and
Security, ser. ARES ’19. New York, NY, USA: Association for Computing Machinery, 2019.
[11] O. Bamigbade, J. Sheppard, and M. Scanlon, “Computer
Vision for Multimedia Geolocation in Human Trafficking
Investigation: A Systematic Literature Review,” CoRR,
vol. abs/2402.15448, 2024.
[12] A. Stylianou, R. Souvenir, and R. Pless, “TraffickCam:

Explainable Image Matching For Sex Trafficking Investigations,” in Proceedings of AAAI Fall Symposium Series
(AAAI FSS-19): Artificial Intelligence in Government and
Public Sector, Arlington, Virginia, USA, 2019.
[13] H. Xuan, A. Stylianou, and R. Pless, “Improved embeddings with easy positive triplet mining,” in Proceedings
of the IEEE/CVF Winter Conference on Applications of
Computer Vision (WACV), Snowmass, CO, USA, 2020,
pp. 2463–2471.
[14] H. Xuan, A. Stylianou, X. Liu, and R. Pless, Hard
Negative Examples are Hard, but Useful, ser. Computer
Vision – ECCV 2020. Cham: Springer International
Publishing, 2020, vol. 12359, pp. 126–142.
[15] B. Tseytlin and I. Makarov, “Hotel recognition via latent
image embeddings,” in Advances in Computational Intelligence, I. Rojas, G. Joya, and A. Català, Eds. Cham:
Springer International Publishing, 2021, pp. 293–305.
[16] S. Black, A. Stylianou, R. Pless, and R. Souvenir, “Visualizing paired image similarity in transformer networks,”
in Proceedings of the IEEE/CVF Winter Conference on
Applications of Computer Vision (WACV), January 2022,
pp. 3164–3173.
[17] S. Saikia et al., “Image feature representation using deep
learning for instance search and scene recognition,” Ph.D.
dissertation, Universidad de León, León, Spain, 2021.
[18] N. A. M. Zin, R. Yusof, S. A. Lashari, A. Mustapha,
N. Senan, and R. Ibrahim, “Content-based image retrieval
in medical domain: A review,” Journal of Physics: Conference Series, vol. 1019, no. 1, p. 012044, June 2018.
[19] R. Padhy, L. Samantaray, S. K. Dash, and J. Mishra,
“Classification of high-resolution satellite image with
content based image retrieval and local binary pattern,”
in International Conference on Innovations in Intelligent
Computing and Communications. Springer, 2022, pp.
409–416.
[20] H. Cui, N. Giebink, J. Starr, D. Longert, B. Ford, and
E. Léveillé-Bourret, “From noisy data to useful color
palettes: One step in making biodiversity data fair,” vol.
13971, 2023, pp. 469–481.
[21] “Framework (COD) No 106/2021 of the European
Parliament with recommendations to the Commission
of 20 October 2020 on ethical aspects of artificial intelligence, robotics and related technologies,”
2020, https://oeil.secure.europarl.europa.eu/oeil/popups/
printsummary.pdf (Accessed on September 30, 2024).
[22] H. Aboaisha, “The optimisation of elementary and integrative content-based image retrieval techniques,” PhD
thesis, University of Huddersfield, Huddersfield, United
Kingdom, 2015.
[23] Y.-C. Hu, M.-G. Lee, and P. Tsai, “Colour palette
generation schemes for colour image quantization,” The
Imaging Science Journal, vol. 57, no. 1, pp. 46–59, 2009.
[24] F. A. Alghamdi, “An effective hybrid framework based
on combination of color and texture features for contentbased image retrieval,” Arabian Journal for Science and
Engineering, vol. 49, no. 3, pp. 3575–3591, 2024.