# Plug to place: Indoor multimedia geolocation from electrical sockets for digital investigation

## Full Text

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

Source PDF: [PlugToPlace-IndoorMultimediaGeolocation.pdf](PlugToPlace-IndoorMultimediaGeolocation.pdf)
Source SHA-256: `14854e8cba55dd491898b399f6365ac7d0e1f1321f3f2cd3945aebde8bdccae9`
Generated with `pdftotext -nopgbrk`; formatting and reading order may differ from the PDF.

---

Forensic Science International: Digital Investigation 56 (2026) 302056

Contents lists available at ScienceDirect

Forensic Science International: Digital Investigation
journal homepage: www.elsevier.com/locate/fsidi

DFRWS EU 2026 - Selected Papers from the 13th Annual Digital Forensics Research Conference Europe

Plug to place: Indoor multimedia geolocation from electrical sockets for
digital investigation
Kanwal Aftab a,* , Graham Adams b , Mark Scanlon a
a
b

Forensics and Security Research Group, School of Computer Science, University College Dublin, Belfield, Dublin, 4, Ireland
School of Engineering, Case Western Reserve University, Cleveland, OH, United States

A R T I C L E I N F O

A B S T R A C T

Keywords:
Multimedia Geolocation
Computer Vision
Hotels-50K
Indoor
Multimedia forensics
Human trafficking

Computer vision is a rapidly evolving field, giving rise to powerful new tools and techniques in digital forensic
investigation, and shows great promise for novel digital forensic applications. One such application, indoor
multimedia geolocation, has the potential to become a crucial aid for law enforcement in the fight against human
trafficking, child exploitation, and other serious crimes. While outdoor multimedia geolocation has been widely
explored, its indoor counterpart remains underdeveloped due to challenges such as similar room layouts,
frequent renovations, visual ambiguity, indoor lighting variability, unreliable GPS signals, and limited datasets in
sensitive domains.
This paper introduces a pipeline that uses electrical sockets as consistent indoor markers for geolocation, since
plug socket types are standardised by country or region. The three-stage deep learning pipeline detects plug
sockets (YOLOv11, mAP@0.5 = 0.843), classifies them into one of 12 plug socket types (Xception, accuracy =
0.912), and maps the detected socket types to countries (accuracy = 0.96 at >90 % threshold confidence). To
address data scarcity, two dedicated datasets were created: socket detection dataset of 2328 annotated images
expanded to 4074 through augmentation, and a classification dataset of 3187 images across 12 plug socket
classes. The pipeline was evaluated on the Hotels-50K dataset, focusing on the TraffickCam subset of crowdsourced hotel images, which capture real-world conditions such as poor lighting and amateur angles. This
dataset provides a more realistic evaluation than using professional, well-lit, often wide-angle images from travel
websites. This framework demonstrates a practical step toward real-world digital forensic applications. The code,
trained models, and the data for this paper are available open source.

1. Introduction
Human trafficking is a severe global crime affecting millions across
all ages, genders, and backgrounds, causing deep personal, community,
and societal harm (Walby and Francis, 2025). It entails the illegal trade
of people through deception, violence, or exploitation, resulting in
forced labour, sexual abuse, and organ trafficking (Dimas et al., 2022).
Given its severe and long-lasting impact, the fight against human traf­
ficking is explicitly prioritised under 3 of the United Nations Sustainable
Development Goals (SDGs) (United Nations Office on Drugs and Crime,
2025). In addition, child sexual exploitation material (CSEM) investi­
gation is one of the most common case types encountered in digital fo­
rensics laboratories within law enforcement agencies worldwide
(Hargreaves et al., 2024).
With rapid technological advancement, the rise of feature-rich

smartphones, expanded storage capabilities, widespread internet ac­
cess, and the growing influence of social media, nearly every facet of
modern life has become digital (Kumar et al., 2024). However, modern
technologies are increasingly exploited by offenders to facilitate serious
crimes, underscoring the critical importance of digital forensics
(Fähndrich et al., 2023). As the volume of digital data continues to
surge, the analysis and interpretation of digital evidence have become
indispensable to modern investigations (Khadka et al., 2024). The
application of artificial intelligence (AI) to digital forensic investigations
is still very much in its infancy (Breitinger et al., 2024), but nonetheless,
practitioners have already identified image/media classification as
having the most potential for the future use of AI in their investigations
(Hargreaves et al., 2024). In parallel, hotel recognition has become a
common need for human trafficking investigations, as hotels are
frequently used as intermediary stopover locations during the trafficking

* Corresponding author.
E-mail address: kanwal.aftab@ucdconnect.ie (K. Aftab).
https://doi.org/10.1016/j.fsidi.2026.302056
Available online 24 March 2026
2666-2817/© 2026 The Author(s). Published by Elsevier Ltd on behalf of DFRWS. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).

K. Aftab et al.

Forensic Science International: Digital Investigation 56 (2026) 302056

of victims (Kamath et al., 2021). Online human trafficking advertise­
ments or the interception of organised crime gang's electronic commu­
nications are often the starting points for many human trafficking
investigations. Identifying hotel rooms from these photographs is an
extremely arduous task. Indeed, police agencies often resort to crowd­
sourcing their locations. For example, hotel rooms often feature in the
Trace an Object projects run by Europol1 or the Australian Centre to
Counter Child Exploitation,2 asking for the general public's help in
identifying the hotels during investigations of cases involving child
sexual exploitation material. In the DFPulse 2024 survey (Hargreaves
et al., 2024), digital forensic practitioners identified image/media
classification and CSAM investigation as two of the main areas where
artificial intelligence has the potential to assist in their future cases.
In terms of automated geolocation, indoor geolocation specifically
faces significant challenges. GPS is unreliable indoors. Alternative in­
door sensors are relatively costly and have limitated range and accuracy
(Hromadová et al., 2021), and image metadata is frequently stripped
during online or instant messaging sharing – making it difficult to trace
an image's origin (Bamigbade et al., 2024). While outdoor environments
typically provide clear geolocation cues, such as landmarks and infra­
structure (Zhang et al., 2021; Brejcha and Čadík, 2017; Luo et al., 2011),
indoor settings are far more complex. Similar layouts, recurring furni­
ture, and inconsistent lighting make it challenging to reliably differen­
tiate between locations (Bhavanasi and Stylianou, 2023). Amid these
challenges, it is therefore crucial to identify consistent and distinctive
cues for indoor geolocation. One such feature of indoor environments is
the presence of electrical sockets, which can serve as distinctive and
geographically informative visual markers. Each country adheres to
standardised socket designs governed by national or regional electrical
regulations (International Electrotechnical Commission, 2015). Plug
and socket designs vary considerably in shape, grounding, voltage, and
frequency. This study focusses solely on their visual characteristics
(McGregor, 2025). Using Computer Vision (CV), detecting and classi­
fying these plug sockets can provide a reliable cue for narrowing down
the search space, as illustrated in Fig. 1.

2. Related work
To address the challenge of indoor multimedia geolocation, specif­
ically hotel recognition, researchers have employed a variety of tech­
niques, ranging from hand-crafted feature extraction, image
classification to neural networks for automated feature extraction, as
well as more advanced approaches such as object-based similarity,
image embeddings, and semantic scene understanding. Regardless of the
methodology employed, the fundamental building block remains CV.
Broadly, researchers approach multimedia geolocation in two ways.
When a large and representative reference dataset is available, the task
is formulated as a Content based Image Retrieval (CBIR) problem; in the
absence of such a dataset, geolocation instead relies on universal visual
cues, i.e., landmarks or sun angle in outdoor environments, to infer
location directly from the image itself (Hays and Efros, 2008; Kakar and
Sudha, 2012).
In this context, high-quality data remains essential. However,
obtaining such data can be particularly challenging in sensitive sce­
narios, such as human trafficking investigations. Recognising these
ethical and operational challenges, the European Union formalised AI
regulations through the AI Act (Gstrein et al., 2024; European Union,
2024), aiming to maximise the benefits of AI while enforcing rigorous
ethical and safety standards.
In the context of data collection and processing, the creation of
specialised datasets has been significant. The Hotels-50K dataset
(Stylianou et al., 2019) contains over one million labelled images from
50,000 hotels worldwide, sourced from both travel websites and the
Traffick Cam mobile application, and is designed specifically for hotel
recognition research. The Hotel-ID dataset (Kamath et al., 2021) pro­
vides a similarly large-scale resource for the same domain. Both datasets
serve as benchmarks for evaluating recognition approaches.
Deep learning models are often black-box systems, and their complex
structures make it difficult to interpret how visual inputs lead to an
output, obscuring the specific cues that drive similarity judgments. To
address this, Black et al. (2022) introduced a paired image similarity
visualisation technique tailored for Transformer architectures. Their
method enables the inspection of attention maps to reveal which regions
contribute most to a given similarity score. When comparing ResNet and
Vision Transformer (ViT) on datasets such as Hotels-50K, Clean Google
LandmarksV2, and Stanford Online Products, the study found that ViT's
attention focused on meaningful cues, such as bed covers, runners, and
tiled shower walls. These results emphasise the need for explainable
embeddings in visual retrieval.
Similarly, Wazzan et al. (2024) explored how context affects object
matching, finding that a moderate amount of context improves anno­
tation efficiency and retrieval accuracy, while excessive context com­
plicates recognition in ambiguous scenes.
Beyond these pixel-level and embedding-based methods, some
research focussed on object-centric retrieval. This approach represents
images as groups of distinct objects rather than just overall textures,
aiming to bridge the gap between simple image features and what
people actually see. For instance, Kim et al. (2003) introduced a method
to detect and segment the main object in an image, as subjects are often
centred in photos. The authors used relevance feedback on these seg­
ments to improve retrieval, though their reliance on colour made it less
effective in busy scenes. Later, Pradhan et al. (2016) used visual saliency
maps to identify object regions and create feature vectors for retrieval.
This method worked well for images with a clear main object but was
less effective for complex scenes.
Furthermore, Bhavanasi and Stylianou (2023) argued that hotel
recognition cannot be treated as a standard image-classification task due
to the extremely large number of hotel classes. To address this, the au­
thors proposed an object-centric ensemble approach in which separate
models are trained for different object types. The system employed a
Hybrid Vision Transformer combining a ResNet-26 backbone for feature
extraction with a ViT classifier across 3110 classes. This hybrid design

1.1. Contribution of this work
This paper makes the following key contributions:
• Dataset Creation and Release. Two novel CV datasets have been
curated and publicly shared: (i) a socket detection dataset for iden­
tifying sockets in indoor room images, and (ii) a socket classification
dataset containing 12 socket categories. These datasets provide
valuable benchmarks for advancing research in fine-grained indoor
object recognition.
• Algorithmic Pipeline for Detection and Classification. A comparative
study of state-of-the-art detection models and Convolutional Neural
Network (CNN) classifiers has been conducted. Based on this anal­
ysis, a pipeline was designed to select the most effective combination
of detection and classification models for socket recognition.
• Evaluation of the Proposed Approach on Real-World Data. The
approach is evaluated on the Hotels-50K dataset, specifically the
TraffickCam subset, demonstrating its practical utility for law
enforcement by narrowing down search spaces in real-world human
trafficking investigations. Furthermore, this work aims to lay the
foundation for a universal socket detector, enabling broader appli­
cations beyond the law enforcement/investigative domains.
All code, trained models, and data created as part of this paper are
made available open source and can be accessed from https://github.
com/markscanlonucd/plugtoplace.
1
2

https://www.europol.europa.eu/stopchildabuse.
https://www.accce.gov.au/what-we-do/trace-an-object.
2

K. Aftab et al.

Forensic Science International: Digital Investigation 56 (2026) 302056

Fig. 1. Worldwide plug type distribution map

delivered superior performance, achieving nearly 80 % top-1 accuracy
on the Hotel-ID 2022 dataset, compared with only 60 % using a con­
ventional full-image method. This shows the value of pinpointing
distinctive features, such as furniture, fixtures, or wall patterns. Addi­
tionally, semi-automated labelling pipelines (Wazzan et al., 2024)
demonstrate that using less context speeds up labelling without hurting
performance, suggesting that focusing on main objects and a moderate
amount of context best balances interpretability and efficiency.
Besides these semantic methods, feature-level CBIR focusses on
capturing broad visual qualities such as colour, texture, and layout (Cao
et al., 2020). Among these attributes, colour is a dominant and widely
used image characteristic, as it is both robust, computationally efficient,
and independent of orientation and image resolution, making it highly
versatile (Shamoi et al., 2022; Bakair et al., 2017). While RGB is com­
mon, colour spaces like CIELAB, Munsell, and fuzzy-based models often
yield more meaningful results (Shamoi et al., 2022). Using a mix of
colour spaces, such as RGB, YCbCr, and Lab*, can also boost precision
(Sangeetha et al., 2022). Specific to hotel room identification, Herrmann
et al. (2024) evaluated CBIR systems using colour features on
Hotels-50K, achieving over 95 % Top-50 accuracy with just two de­
scriptors, thus supporting faster and more reliable investigative work­
flows. More recently, Bamigbade et al. (2025) combined major colour
palettes and simple histograms with deep metric learning and classifi­
cation to raise top-20 retrieval accuracy by 17 %. These findings show
that blending handcrafted features with deep embeddings can improve
both clarity and results.
Even with sophisticated methods, systematic reviews expose key
gaps in using CV for social good. Dimas et al. (2022) note that Opera­
tions Research and Analytics efforts for anti-human trafficking mostly
target sex trafficking and prosecution, with less attention to labour
trafficking, prevention, or victim protection. Similarly, Bamigbade et al.
(2024) highlighted the benefits of using CV based geolocation in in­
vestigations, calling for more data types and clearer deep learning
models to extract useful information.
Overall, these studies show a shift in hotel recognition research: from
data-driven deep metric learning to more comprehensible, objectfocused, and feature-blended CBIR systems. This change not only

improves technical performance but also supports the growing demands
for ethical, transparent, and socially responsible AI in sensitive
investigations.
3. Experimental setup and analysis
The CV methodology proposed as part of this paper is a three-tiered
process, as detailed in Fig. 2. In the first stage, YOLO object detection
identifies and localises electric sockets in an image, generating cropped
regions of interest (ROI). In the second stage, a CNN classifies each
cropped socket ROI by type. Finally, in the third stage, the detected
socket type is mapped to potential countries, which narrows the list of
likely locations and supports law enforcement investigations. Sections 4
to 6 provide a detailed discussion of each stage, including dataset
preparation for every stage, the methodologies used, and a comparative
analysis of different algorithms along with their results.
4. Stage 1: socket detection
Stage 1 of the pipeline focuses on detecting electrical sockets in
image to infer indoor geographical location. Electrical sockets are
consistent and recognisable markers, as each country or region uses
specific socket types defined by distinct pin configurations. These fea­
tures are visually distinctive in indoor environments such as hotel
rooms. A CV-based object detection model is applied to automatically
identify socket instances and their positions within an image. The
detected socket type is then used to constrain the possible geographical
location, supporting subsequent investigative analysis.
4.1. Dataset preparation
High-quality and diverse datasets are critical in CV, as both the
quantity and quality of training data directly affect model accuracy and
generalisation (Joshi et al., 2024). For this study, a total of 2328 socket
images were compiled, of which 1525 were obtained from publicly
available Roboflow (Roboflow (2025) socket datasets licenced under CC
4.0, while the remaining 803 images were cropped from hotel room
3

K. Aftab et al.

Forensic Science International: Digital Investigation 56 (2026) 302056

Fig. 2. Architecture of the proposed three-stage pipeline: (1) Socket detection, (2) Socket type classification, and (3) Geolocation.

scenes in the Hotels-50K dataset. This original dataset (Dataset A) was
partitioned into training (70 %, 1629 images), validation (20 %, 455
images), and test (10 %, 244 images) subsets. All images were annotated
with bounding boxes using the Roboflow platform and manually clas­
sified into two categories: class 0 (NA) for switchboards and class 1 for
sockets. Although the initial focus was exclusively on sockets, the
frequent misclassification of visually similar switchboard buttons
necessitated the inclusion of the NA class to better differentiate
non-target elements from actual sockets. This confusion was heavily
compounded by the small sizes of the objects, their low resolution in
cropped images, and the similar colour and high concentration of but­
tons and sockets on the switchboards.
To enhance dataset diversity and improve model generalisation, data
augmentation was applied exclusively to the training set, doubling its
size from 1629 to 3258 images. Augmentation techniques included
random cropping (0–20 %), rotation (− 15◦ to +15◦ ), grayscale con­
version (15 % of images), hue adjustment (− 24◦ to +24◦ ), and bright­
ness adjustment (− 19 % to +19 %). After additional images were added,
the dataset resulted in training (80 %, 3258 images), validation (10 %,
409 images), and test (10 %, 407 images) subsets, for a total of 4074
images. The validation and test sets were not augmented; however, an
additional 117 images cropped from the Hotels-50K dataset were added
to ensure an unbiased evaluation of model performance.

prioritised, rendering single-stage object detectors a more practical op­
tion. The YOLO family was chosen for its combination of high speed and
competitive accuracy (Wang et al., 2023), which is suitable for socket
detection as a first-stage step before potential refinement. All YOLO
models used in this study are initialised with pre-trained weights and
subsequently fine-tuned for socket detection, allowing reliable perfor­
mance with limited task-specific training data.
Although newer YOLO versions introduce architectural improve­
ments, the latest release is not always stable (Jegham et al., 2025). A
comparative analysis was then conducted. The experimental outcomes,
as outlined in Table 1, demonstrate the performance of each model
across diverse metrics, starting with YOLOv8 as a benchmark, followed
by YOLOv11 and YOLOv12. Each release provides multiple variants
(nano, small, medium, large). This study predominantly focused on
small and medium models, and in one case, the large model was also
considered to balance computational cost with accuracy. Hyper­
parameter tuning, particularly the number of training epochs, was
evaluated to minimise underfitting and overfitting, and the effect of data
augmentation on model performance was also assessed.
4.3. Performance matrix
Mean Average Precision (mAP) is a standard metric in object
detection that summarises how accurately a model detects and localises
objects. It is based on precision and recall, using Intersection over Union
(IoU) to measure the overlap between predicted and ground truth
bounding boxes. mAP@0.5 computes the average precision at an IoU
threshold of 0.5, meaning a detection is considered correct if at least 50
% of the predicted box overlaps the ground truth, providing a measure of
overall detection success. mAP@0.5:0.95 averages performance across
multiple IoU thresholds from 0.5 to 0.95, evaluating the model's ability
to both detect objects and precisely localise them. High mAP@0.5:0.95
indicates strong localisation accuracy, while mAP@0.5 primarily re­
flects general detection performance. These metrics together provide a
robust assessment of the model's effectiveness in identifying and
locating objects in images.

4.2. Methodology
Recent advances in CV have positioned object detection as a funda­
mental task, achieving performance that, in some cases, rivals or even
surpasses human capabilities (Borji and Itti, 2014). A wide range of
state-of-the-art algorithms are available for object detection, and the
choice of an appropriate model for custom datasets depends heavily on
the specific requirements of the application. Object detection algorithms
are broadly classified as two-stage or single-stage detectors. Two-stage
models are more accurate but slower, while single-stage models offer
faster inference with only a slight accuracy trade-off (Mohammed,
2025). To handle large volumes of images efficiently, speed was
4

K. Aftab et al.

Forensic Science International: Digital Investigation 56 (2026) 302056

Table 1
Test results for all classes across models. Bold indicates the best value per column.
Version

Type

Epochs

Precision (P)

Recall (R)

mAP@0.5

mAP@0.5:0.95

YOLOv8
YOLOv8
YOLOv8
YOLOv11
YOLOv11
YOLOv11
YOLOv12

s
s
m
s
m
l
s

180
250
set = 300 (early stopping at 233)
180
250
set = 250 (early stopping at 234)
180

0.843
0.834
0.883
0.847
0.846
0.851
0.720

0.709
0.709
0.750
0.734
0.791
0.764
0.650

0.793
0.790
0.809
0.804
0.832
0.806
0.748

0.529
0.524
0.540
0.540
0.539
0.534
0.483

4.4. Comparative analysis and results

Table 2
Comparison of YOLOv11m variant with data augmentation and k fold cross
validation on and Dataset B. Bold indicates the best value per dataset.

This study evaluated the performance of various YOLO models for
socket detection using Dataset A, the original dataset, in the first phase,
and subsequently assessed the effect of data augmentation in the second
phase on the augmented data, Dataset B. All models were trained with a
batch size of 16 and an input resolution of 640 pixels. The optimiser was
set to AdamW, which automatically tuned hyperparameters and over­
rode the default learning rate and momentum values, resulting in an
effective learning rate of 0.001667 with momentum fixed at 0.9.
The experiment began with YOLOv8, starting with the small variant
(YOLOv8s) to assess the effect of increasing the number of training
epochs from 180 to 250. The focus then shifted to the medium variant
(YOLOv8m) to evaluate the trade-off between model capacity and per­
formance. Although the number of epochs was initially set to 300,
training was stopped early at 233 epochs due to the early stopping
mechanism designed to prevent overfitting.
The analysis then progressed to YOLOv11, beginning with the small
variant (YOLOv11s) trained for 180 epochs, followed by the medium
variant (YOLOv11m) trained for 250 epochs. For the large variant
(YOLOv11l), training was scheduled for 250 epochs but was stopped
early at 234 epochs due to early stopping. Finally, YOLOv12 was eval­
uated; however, it achieved comparatively lower performance metrics.
Specifically, YOLOv12 underperformed, achieving mAP@0.5 of
0.748 and mAP@0.5:0.95 of 0.483, with a precision of 0.720 and a recall
of 0.650. These results indicate weaker socket detection performance,
highlighting that cutting-edge models do not always outperform more
established versions such as YOLOv8 and YOLOv11. Newly released
models are often unstable and improve over time; therefore, it is
generally advisable to allow them to mature before deployment in CV
applications (Jiang and Zhong, 2025).
Validation accuracy should not be solely relied upon, as it may
indicate overfitting on the training data and not necessarily reflect
performance on unseen test data (Santos and Papa, 2022). Therefore,
test accuracy is considered more decisive than validation accuracy. On
the test set, YOLOv8m (early stopping at 233 epochs) achieves the
highest precision (0.883), while both YOLOv8m and YOLOv11s (180
epochs) achieve the highest mAP@0.5:0.95 (0.540). This indicates that
these models are highly accurate in correctly detecting sockets, though
their recall varies (0.750 for YOLOv8m, 0.734 for YOLOv11s). In
contrast, YOLOv11m (250 epochs) attains the highest recall (0.791) and
the highest mAP@0.5 (0.832), indicating strong overall detection
coverage and good localisation.
To further assess the impact of data augmentation, YOLOv11m was
trained on Dataset B with augmentation, while YOLOv11s was trained
using both augmentation and 5-fold cross-validation. Table 2 summa­
rises the results. Adding only augmentation slightly altered perfor­
mance, with precision decreasing from 0.846 to 0.783, recall increasing
from 0.748 to 0.765, mAP@0.5 decreasing from 0.832 to 0.766, and
mAP@0.5:0.95 decreasing from 0.539 to 0.485. These results suggest
that augmentation improves recall at a modest cost to precision and
overall localisation accuracy.
In contrast, combining K-fold cross-validation with augmentation
further enhances model robustness, reducing variance between folds
and providing more reliable generalisation to unseen images.

Dataset

Setting

Precision
(P)

Recall
(R)

mAP@0.5

mAP
@0.5:0.95

Test
(Dataset
B)

Yolov11m
without Aug

0.805

0.748

0.792

0.498

Yolov11m
with Aug
Yolov11s
with Aug

0.783

0.765

0.766

0.485

0.8675

0.7990

0.8427

0.5771

​
K fold cross
validation

Specifically, average precision increased from 0.847 to 0.872, and recall
increased from 0.734 to 0.756, demonstrating that K-fold training
effectively mitigates overfitting on smaller datasets and improves
overall detection performance. Based on this analysis, YOLOv11s with
augmentation and K-fold cross-validation was identified as the best
performer in Stage One socket detection. Figs. 3 and 4 show the visual
socket detection results of this best-performing model.
5. Stage 2: socket type classification
Stage 2 of the pipeline focuses on classifying detected electric sockets
into region-specific socket types based on their visual characteristics.
Although 14 domestic socket types are internationally recognised, this
stage considers only visual distinctions relevant for image-based classi­
fication. As illustrated in Figs. 5 and 1, socket designs exhibit strong
regional variation, enabling geographical constraints to be inferred from
socket appearance.
In this experiment, CV is used to classify socket types, since the focus
is exclusively on visual characteristics, the study considers 12 classes
instead of the full set of 14. Types D and M have been merged into a
single class (DM) because, despite differences in pin size, their layouts
are visually indistinguishable. Type M closely resembles Type D but
features larger pins. Similarly, Types J and N are merged into a single
class (JN). Both sockets use three round pins with nearly identical
configurations, differing only in the precise offset of the earth pin – with
Type J, it is offset by 5 mm, and with Type N, it is offset by 3 mm
(International Electrotechnical Commission, 2015). This small struc­
tural difference makes them electrically incompatible, but the variation
was deemed too subtle to be reliably distinguished visually. All other
socket types present clear visual differences and are, therefore, treated
as separate classes. In addition, a noise class was introduced to exclude
non-socket objects, such as light switches, thermostats, or low-quality
regions of interest detected by YOLO, thereby further enhancing
accuracy.
5.1. Dataset preparation
The dataset for socket type classification was constructed from two
primary sources. The first source was publicly available datasets from
Roboflow, consisting of plug socket images licenced under CC 4.0. These
images were downloaded, cleaned, and merged into a consistent
5

K. Aftab et al.

Forensic Science International: Digital Investigation 56 (2026) 302056

Fig. 3. YOLOv11m detection results on room images (a–b), showing bounding boxes for socket classes.

Fig. 4. YOLOv11m detection results on bathroom images (a–b), showing bounding boxes for socket classes.

Fig. 5. Plug and socket types from type A to type N (McGregor, 2025).

collection. The second source was the Hotels-50K dataset, from which
socket regions were automatically detected, cropped, and assigned to
their respective classes. After processing, the final dataset comprised 12
socket type classes, with the number of images per class summarised in
Table 3. In total, the dataset contains 3187 images, which were

partitioned into training, validation, and test sets using a 70:15:15 split.
5.2. Methodology
A transfer learning approach for multi-class image classification
6

K. Aftab et al.

Forensic Science International: Digital Investigation 56 (2026) 302056

Table 3
Socket types with image counts and corresponding country usage count.

Table 5
Performance summary of Xception models.

Socket

Images

Countries

Model

Accuracy

Precision

Recall

F1

A
B
C
D/M
E
F
G
H
I
J/N
K
L

192
302
305
300
304
303
304
138
268
262
291
222

46
28
65
12/9
24
35
32
1
11
5/4
6
9

Xception
Xception (5-Fold CV)

0.854
0.877

0.878
0.894

0.853
0.884

0.855
0.881

Total

3187

–

70:15:15.
Therefore, Xception with 5 cross-validation was identified as the best
performer in stage two socket type classification.
6. Stage 3: Inferring geolocation through socket Type–Country
mapping
The final stage of the proposed pipeline focusses on using socket
detection results to infer the geolocation of hotel rooms, highlighting the
practical applicability of the proposed approach.

using five state-of-the-art CNN architectures was implemented: VGG16,
InceptionV3, Xception, ResNet50, and ResNet101. The dataset, con­
sisting of 3187 images across 12 classes (socket types), was split into
training, validation, and test sets, with 2224 images for training, 473
images for validation, and 490 images for testing. Images were pre­
processed and augmented with normalisation and horizontal flipping to
improve generalisation. Each model was initialised with ImageNet pretrained weights, and the convolutional base was frozen to leverage prelearned features while training a new classification head for the target
classes. Models were trained independently using categorical crossentropy loss and the Adam optimiser.

6.1. Test dataset preparation
The test dataset for this experiment was derived from the Hotels-50K
TraffickCam dataset, available from its official GitHub repository. To
prepare the data, the dataset was obtained by modifying the download
script to preserve the original image resolution. This was a crucial step
because the target object, electric sockets, is often small within the
larger image, and resizing could lead to a loss of detail essential for
accurate detection.
The Hotels-50K dataset comprises two subcategories: TraffickCam
and travel website images. TraffickCam consists of crowd-sourced
photos submitted by travellers worldwide, reflecting real-world condi­
tions. After retaining and downloading the images in their original
resolution of 1024 × 768, the dataset provides images in this resolution,
with sockets typically appearing in small regions of approximately 130
× 87 pixels.
In contrast, travel website images are professionally captured under
ideal lighting and angles, often with colour correction and photo editing.
These images are also provided at a lower resolution (350 × 233),
making the sockets barely visible. As a result, they were excluded from
testing. The TraffickCam subset, by depicting more realistic, nonprofessionally taken and/or edited photos, is considered to provide a
better representation of the conditions encountered in practical in­
vestigations and was therefore used exclusively for evaluation.
To establish a ground truth for the experiment, the dataset was
processed to associate each image with its corresponding country. This
involved a multistep process: Merging Metadata: the various CSV files
from the original dataset were consolidated to create a unified file
containing image IDs, hotel IDs, and geographic coordinates (latitude
and longitude). Geolocation: Using the geopy.geocoders library, the
geographic coordinates were converted into country names. Stand­
ardisation: To ensure consistency, the country names were standardised
using the pycountry library, as the raw geolocation output sometimes
returned names in native languages. Illegal characters were removed,
and spaces were replaced with underscores to create valid directory
names.
The final dataset was restructured into a clean directory, with images
organised into subfolders named after their respective countries. This
structure, along with a consolidated CSV file containing all relevant
metadata, streamlined the subsequent country-specific analysis. The
reason images were arranged in folders instead of being directly taken
and compared from the CSV file was to facilitate visual inspection,
ensuring that the code functioned correctly and that each country was
accurately represented with its respective socket type.

5.3. Evaluation and comparative analysis
The performance of the model was evaluated using standard metrics,
including accuracy, precision, recall, F1-score, and a confusion matrix to
provide detailed insights into class-wise predictions. Among the evalu­
ated models, VGG16, InceptionV3, and Xception achieved the highest
accuracies. Xception attained the best overall performance with an ac­
curacy of 91.22 %, consistently demonstrating high precision, recall,
and F1-scores across all 12 classes, as summarised in Table 4. In com­
parison, VGG16 achieved 82.65 % accuracy, while InceptionV3 reached
89.80 %, highlighting the performance differences among the individual
models.
To further validate the robustness of Xception, the impact of K-fold
cross-validation was evaluated by modifying the dataset. A new Noise
class, consisting of 304 images representing non-socket objects poten­
tially missed by Step 1 YOLO, was added to the original 3187 images,
resulting in a total of 3491 images across 13 classes (12 socket types +
Noise). These images were organised into 13 folders for training and
validation in a 5-fold cross-validation setup, while an additional 175
unseen images, spanning all classes, were reserved for testing. This setup
allowed us to assess whether cross-validation improves model
generalisation.
As summarised in Table 5, the 5-fold cross-validation strategy
increased accuracy from 85.4 % to 87.7 %, precision from 87.8 % to
89.4 %, recall from 85.3 % to 88.4 %, and F1-score from 85.5 % to 88.1
%. These results demonstrate that K-fold cross-validation provided a
modest but consistent improvement in performance over the single-run
Xception model. Note that the test dataset differs from the split of
Table 4
Performance summary of different models.
Model

Accuracy

Precision

Recall

F1-score

VGG16
InceptionV3
Xception
ResNet50
ResNet101

0.827
0.898
0.912
0.492
0.443

0.846
0.907
0.914
0.599
0.634

0.816
0.900
0.910
0.477
0.429

0.819
0.901
0.911
0.466
0.433

6.2. Data analysis, evaluation, and results
The model's performance was evaluated by assessing its ability to
7

K. Aftab et al.

Forensic Science International: Digital Investigation 56 (2026) 302056

detect a plug socket in an image and then classify it to determine its
corresponding country. For each image in which a socket was success­
fully detected and classified, it was checked whether the predicted plug
type and the image's actual country formed a valid pair according to this
mapping. A correct match was assigned a score of 1, an incorrect match
received a score of − 1, and the noise class was assigned a neutral score of
0. These scores were used to calculate key performance metrics, such as
the confusion matrix, precision, recall, and F1 score. This approach
provided a comprehensive assessment of the model's accuracy. Finally,
the results were compiled into a detailed summary report. They were
presented with graphical visualisations, such as bar charts, to provide a
clear and intuitive interpretation of the model's overall performance.
A total of 44,630 TraffickCam images were processed through the
algorithmic pipeline. In the first stage, YOLO detected 3759 potential
sockets. To enhance detection accuracy and eliminate false positives, a
second-stage classifier was employed to identify and remove noise.
Specifically, instances where non-socket objects (e.g., switchboards)
were incorrectly detected as sockets in the first stage were classified as
noise. This step identified 1393 noisy detections, leaving 2366 valid
sockets. These valid detections were subsequently passed to the socket
classification stage, where only those with a confidence threshold above
70 % were retained. The results are summarised in Table 6.
When considering different socket classification confidence thresh­
olds, the performance varies. The confidence threshold represents the
model's predicted probability that a socket belongs to the predicted
class. Without applying any socket confidence threshold, 1967 pre­
dictions were correct, and 399 were incorrect, resulting in an accuracy
of 83.08 %. At a threshold above 70 %, 1595 predictions were correct,
and 146 were incorrect, resulting in an accuracy of 91.61 %. Increasing
the threshold above 80 % reduced the number of correct detections to
1,421, while incorrect detections decreased to 95, yielding an improved
accuracy of 93.73 %. At the highest threshold of above 90 %, correct
detections further decreased to 1,167, with only 45 incorrect pre­
dictions, resulting in the highest accuracy of 96.29 %. These results
illustrate the trade-off between confidence and accuracy: lower thresh­
olds capture more sockets but result in more false positives, whereas
higher thresholds reduce errors at the cost of missing some detections, as
shown in Fig. 6.

Fig. 6. Country prediction accuracy by varying socket type confidence
threshold values.

technical limitations of the current detection and classification pipeline,
although some inherent limitations remain when atypical or uncommon
socket types are present. Improvements in detection accuracy, higherresolution imagery, and enhanced classification models could reduce
such errors.
8. Conclusion
This paper presents a universal pipeline for socket detection and
classification, motivated by its potential to enhance geolocation capa­
bilities in investigative contexts where conventional cues, such as met­
adata, outdoor landmarks, and sensor data, are unavailable. By curating
two new datasets, benchmarking multiple detection and classification
models, and evaluating performance on the Hotels-50K TraffickCam
dataset, this study demonstrates both the feasibility and the challenges
associated with socket-based indoor geolocation. Despite challenges
such as small object size and low-resolution imagery, the results
demonstrate strong detection accuracy and validate the concept's prac­
tical potential, enabling investigators to automate geolocation across
large volumes of digital evidence and convert it into actionable
intelligence.
Beyond technical performance, this work contributes to the
emerging field of AI-driven multimedia forensics, where visual scene
elements are utilised to support digital investigations. Socket detection
offers a unique, region-specific forensic cue that can aid law enforce­
ment agencies in narrowing search regions, corroborating other forms of
evidence, and identifying the possible origins of illicit or traffickingrelated imagery. The proposed framework, therefore, represents an
important step towards scalable, context-aware forensic tools that
bridge the gap between CV and real-world investigative practice.
Future work will focus on grouping visually similar scenes based on
distinctive feature representations, enabling indirect geolocation infer­
ence and the propagation of cues from one image across an entire image
set. Such clustering can reveal patterns in large datasets and link un­
known or unlabelled images to geographically known locations. Further
integration of multimodal cues, including architectural styles, interior
designs, or visible textual information, could strengthen geolocation
when electrical fixtures are unavailable. Human inspection of repre­
sentative subsets could also help verify socket presence and detection
accuracy, refine confidence thresholds, and assess the pipeline's appli­
cability to diverse, real-world images. Together, these strategies would
make the system more scalable and practical for forensic and investi­
gative applications, bridging the gap between AI-based contextual
geolocation and digital investigations. to cluster visually similar scenes
and strengthen geolocation inference when sockets are not visible. By
combining socket-based detection with broader contextual cues, the

7. Discussion
The proposed pipeline has been developed with generalisability in
mind, aiming to detect electric sockets in images beyond the Hotels-50K
dataset. As a universal visual cue, sockets have the potential to support
geolocation in any indoor image. However, socket detection is inher­
ently challenging due to their small size and the frequent low resolution
of imagery. Despite these challenges, the method performs strongly
when sockets are present, with geolocation inference achieving 96.29 %
accuracy. The reported precision of 0.85 demonstrates effective detec­
tion, but only around 10 % of TraffickCam images contain visible
sockets, which explains why applying high confidence thresholds re­
duces useable images to about 2 % of the dataset. Misidentification of
countries is primarily due to the YOLO-based detection stage identifying
non-socket objects, such as switchboards or other sources of visual noise,
as potential sockets. In stage two, these detections are assigned the
NOISE class, resulting in a incorrect socket score. Contributing factors
include low image resolution, occlusions, and challenging lighting
conditions. These issues indicate that most misclassifications are
Table 6
Classifier threshold analysis for socket detection.
Class Confidence

Correct

Wrong

Total

Accuracy (%)

≥ 70%
≥ 80%
≥ 90%

1595
1421
1167

146
95
45

1741
1516
1212

91.61
93.73
96.29

8

K. Aftab et al.

Forensic Science International: Digital Investigation 56 (2026) 302056

framework could evolve into a robust and general-purpose forensic
system capable of supporting a wide range of investigative and hu­
manitarian applications.

Jegham, N., Koh, C.Y., Abdelatti, M., Hendawi, A., 2025. YOLO evolution: a
comprehensive benchmark and architectural review of YOLOv12, YOLO11, and their
previous versions. URL: https://arxiv.org/abs/2411.00201. arXiv:2411.00201.
Jiang, T., Zhong, Y., 2025. Odverse33: is the new YOLO version always better? A multi
domain benchmark from YOLO v5 to v11. CoRR abs/2502.14314. https://doi.org/
10.48550/ARXIV.2502.14314.
Joshi, S., Jain, A., Payani, A., Mirzasoleiman, B., 2024. Data-efficient contrastive
language-image pretraining: prioritizing data quality over quantity. In: Dasgupta, S.,
Mandt, S., Li, Y. (Eds.), Proceedings of the 27th International Conference on
Artificial Intelligence and Statistics. PMLR, pp. 1000–1008. URL: https
://proceedings.mlr.press/v238/joshi24a.html.
Kakar, P., Sudha, N., 2012. Authenticating image metadata elements using geolocation
information and sun direction estimation. In: 2012 IEEE International Conference on
Multimedia and Expo, pp. 236–241. https://doi.org/10.1109/ICME.2012.82.
Kamath, R., Rolwes, G., Black, S., Stylianou, A., 2021. The 2021 Hotel-ID to combat
human trafficking competition dataset. URL: https://arxiv.org/abs/2106.05746.
Khadka, S., Singh, R., Begum, M., Muhammad, O., Naeem, B., Chaeikar, S.S., 2024.
A review on computer crimes and digital forensics. In: 2024 IEEE Consumer Life
Tech (ICLT), pp. 1–4. https://doi.org/10.1109/ICLT63507.2024.11038544.
Kim, S., Park, S., Kim, M., 2003. Central object extraction for object-based image
retrieval. In: Bakker, E.M., Lew, M.S., Huang, T.S., Sebe, N., Zhou, X.S. (Eds.), Image
and Video Retrieval. Springer Berlin Heidelberg, Berlin, Heidelberg, pp. 39–49. URL:
https://link.springer.com/chapter/10.1007/3-540-45113-7_5#citeas.
Kumar, S., Verma, A.K., Mirza, A., 2024. Digital Revolution, Artificial Intelligence, and
Ethical Challenges. Springer Nature Singapore, Singapore, pp. 161–177. https://doi.
org/10.1007/978-981-97-5656-8_11.
Luo, J., Joshi, D., Yu, J., Gallagher, A., 2011. Geotagging in multimedia and computer
vision—a survey. Multimed. Tool. Appl. 51, 187–211. https://doi.org/10.1007/
s11042-010-0623-y.
McGregor, C.H., 2025. Plug & socket types around the world. URL: https://www.world
standards.eu/electricity/plugs-and-sockets/. (Accessed 18 September 2025).
Mohammed, S.Y., 2025. Architecture review: two-stage and one-stage object detection.
Franklin Open 12, 100322. https://doi.org/10.1016/j.fraope.2025.100322. URL:
https://www.sciencedirect.com/science/article/pii/S2773186325001100.
Pradhan, J., Pal, A.K., Banka, H., 2016. A prominent object region detection based
approach for CBIR application. In: 2016 Fourth International Conference on Parallel,
Distributed and Grid Computing (PDGC), pp. 447–452. https://doi.org/10.1109/
PDGC.2016.7913237.
Roboflow, 2025. Roboflow: computer vision tools for developers and enterprises. URL:
https://roboflow.com/. (Accessed 18 September 2025).
Sangeetha, S.K.B., Mathivanan, S.K., Pandi, T., Arivu selvan, K., Jayagopal, P., Teshite
Dalu, G., 2022. An enhanced triadic color scheme for content-based image retrieval.
Math. Probl Eng. 2022, 5736630. https://doi.org/10.1155/2022/5736630. URL:
https://onlinelibrary.wiley.com/doi/abs/10.1155/2022/5736630.
Santos, C.F.G.D., Papa, J.a.P., 2022. Avoiding overfitting: a survey on regularization
methods for convolutional neural networks. ACM Comput. Surv. 54. https://doi.org/
10.1145/3510413.
Shamoi, P., Sansyzbayev, D., Abiley, N., 2022. Comparative overview of color models for
content-based image retrieval. In: 2022 International Conference on Smart
Information Systems and Technologies (SIST), pp. 1–6. https://doi.org/10.1109/
SIST54437.2022.9945709.
Stylianou, A., Xuan, H., Shende, M., Brandt, J., Souvenir, R., Pless, R., 2019. Hotels-50K:
a global hotel recognition dataset. In: Proceedings of the Thirty-Third AAAI
Conference on Artificial Intelligence and Thirty-First Innovative Applications of
Artificial Intelligence Conference and Ninth AAAI Symposium on Educational
Advances in Artificial Intelligence. AAAI Press, pp. 726–733. https://doi.org/
10.1609/aaai.v33i01.3301726.
United Nations Office on Drugs and Crime, 2025. Human trafficking and the SDGs.
https://www.unodc.org/unodc/human-trafficking/sdgs.html.
Walby, S., Francis, B., 2025. Improving the estimate of trafficking in human beings and
modern slavery by integrating data from ILO/walk Free/IOM and UNODC. Soc.
Indic. Res. 176, 669–693. https://doi.org/10.1007/s11205-024-03474-w.
Wang, X., Li, H., Yue, X., Meng, L., 2023. A comprehensive survey on object detection
YOLO. In: The 5th International Symposium on Advanced Technologies and
Applications in the Internet of Things (ATAIT 2023), pp. 77–89.
Wazzan, A., Ahmad, I., Macneil, S., Souvenir, R., 2024. Context or clutter? Efficiently
matching objects across scenes. In: Proceedings of the 2024 International Conference
on Multimedia Retrieval. Association for Computing Machinery, New York, NY, USA,
pp. 404–413. https://doi.org/10.1145/3652583.3658090.
Zhang, X., Wang, L., Su, Y., 2021. Visual place recognition: a survey from deep learning
perspective. Pattern Recogn. 113, 107760. https://doi.org/10.1016/j.
patcog.2020.107760. URL: https://www.sciencedirect.com/science/article/pii/S00
3132032030563X.

References
Bakair, A.S., Mohamed, T.M., Sadek, R.A., 2017. Enhancing the capability of color based
CBIR by voting of color histograms. In: 2017 27th International Conference on
Computer Theory and Applications (ICCTA), pp. 4–10. https://doi.org/10.1109/
ICCTA43079.2017.9497165.
Bamigbade, O., Scanlon, M., Sheppard, J., 2025. Improving image embeddings with
colour features in indoor scene geolocation. IEEE Access 13, 79860–79870. https://
doi.org/10.1109/ACCESS.2025.3564496.
Bamigbade, O., Sheppard, J., Scanlon, M., 2024. Computer vision for multimedia
geolocation in human trafficking investigation: a systematic literature review. URL:
https://arxiv.org/abs/2402.15448.
Bhavanasi, S.S., Stylianou, A., 2023. Hotel recognition using object ensembles. In: 2023
IEEE Applied Imagery Pattern Recognition Workshop (AIPR), pp. 1–8. https://doi.
org/10.1109/AIPR60534.2023.10440661.
Black, S., Stylianou, A., Pless, R., Souvenir, R., 2022. Visualizing paired image similarity
in transformer networks. In: 2022 IEEE/CVF Winter Conference on Applications of
Computer Vision (WACV), pp. 1534–1543. https://doi.org/10.1109/
WACV51458.2022.00160.
Borji, A., Itti, L., 2014. Human vs. computer in scene and object recognition. In: 2014
IEEE Conference on Computer Vision and Pattern Recognition, pp. 113–120. https://
doi.org/10.1109/CVPR.2014.22.
Breitinger, F., Hilgert, J.N., Hargreaves, C., Sheppard, J., Overdorf, R., Scanlon, M.,
2024. DFRWS EU 10-year review and future directions in digital forensic research.
Forensic Sci. Int.: Digit. Invest. 48, 301685. https://doi.org/10.1016/j.
fsidi.2023.301685. https://www.sciencedirect.com/science/article/pii/S2666281
723002044. DFRWS EU 2024 - Selected Papers from the 11th Annual Digital
Forensics Research Conference Europe.
Brejcha, J., Čadík, M., 2017. State-of-the-art in visual geo-localization. Pattern Anal.
Appl. 20, 613–637. https://doi.org/10.1007/s10044-017-0611-1.
Cao, B., Araujo, A., Sim, J., 2020. Unifying deep local and global features for image
search. In: Computer Vision – ECCV 2020: 16Th European Conference, Glasgow, UK,
August 23–28, 2020, Proceedings, Part XX. Springer-Verlag, Berlin, Heidelberg,
pp. 726–743. https://doi.org/10.1007/978-3-030-58565-5_43.
Dimas, G.L., Konrad, R.A., Lee Maass, K., Trapp, A.C., 2022. Operations research and
analytics to combat human trafficking: a systematic review of academic literature.
PLoS One 17, 1–24. https://doi.org/10.1371/journal.pone.0273708.
European Union, 2024. Regulation (EU) 2024/1689 of the European Parliament and of
the Council of 13 June 2024 laying down harmonised rules on artificial intelligence
and amending Regulations (EC) No 300/2008, (EU) No 167/2013, (EU) No 168/
2013, (EU) 2018/858, (EU) 2018/1139 and (EU) 2019/2144 and Directives 2014/
90/EU, (EU) 2016/797 and (EU) 2020/1828 (Artificial Intelligence Act). Official
Journal of the European Union, L2024/1689. URL: https://eur-lex.europa.eu
/legal-content/EN/TXT/?uri=CELEX%3A32024R1689. (Accessed 16 December
2025).
Fähndrich, J., Honekamp, W., Povalej, R., Rittelmeier, H., Berner, S., Labudde, D., 2023.
Digital forensics and strong AI: a structured literature review. Forensic Sci. Int.:
Digit. Invest. 46, 301617. https://doi.org/10.1016/j.fsidi.2023.301617. URL:
https://www.sciencedirect.com/science/article/pii/S2666281723001294.
Gstrein, O., Haleem, N., Zwitter, A., 2024. General-purpose AI regulation and the
European Union AI Act. Internet Policy Review 13, 1–26. https://doi.org/10.14763/
2024.3.1790.
Hargreaves, C., Breitinger, F., Dowthwaite, L., Webb, H., Scanlon, M., 2024. DFPulse: the
2024 digital forensic practitioner survey. Forensic Sci. Int.: Digit. Invest. 51, 301844.
https://doi.org/10.1016/j.fsidi.2024.301844. URL: https://www.sciencedirect.co
m/science/article/pii/S2666281724001719.
Hays, J., Efros, A.A., 2008. IM2GPS: estimating geographic information from a single
image. In: 2008 IEEE Conference on Computer Vision and Pattern Recognition,
pp. 1–8. https://doi.org/10.1109/CVPR.2008.4587784.
Herrmann, J., Bamigbade, O., Sheppard, J., Scanlon, M., 2024. Perceptual colour-based
geolocation of human trafficking images for digital forensic investigation. In: 2024
Cyber Research Conference - Ireland (Cyber-RCI). IEEE, pp. 1–8. https://doi.org/
10.1109/Cyber-RCI60769.2024.10941203.
Hromadová, V., Machaj, J., Brída, P., 2021. Impact of user orientation on indoor
localization based on wi-fi. Transp. Res. Procedia 55, 882–889. https://doi.org/
10.1016/j.trpro.2021.07.056. https://www.sciencedirect.com/science/article/pii/
S2352146521004609, 14th International Scientific Conference on Sustainable,
Modern and Safe Transport.
International Electrotechnical Commission, 2015. Plugs and socket-outlets for domestic
and similar general use standardized in member countries of IEC. Technical Report.
URL: https://webstore.iec.ch/publication/23628.

9