Small Patterns Detection in Historical Digitised Manuscripts Using
Very Few Annotated Examples
Hussein Mohammed
and Mahdi Jampour
Cluster of Excellence, Understanding Written Artefacts, Universit
at Hamburg, Hamburg, Germany
Pattern Detection, Deep Learning, Historical Manuscripts, Datasets.
Historical manuscripts can be challenging for computer vision tasks such as writer identification, style clas-
sification and layout analysis due to the degradation of the artefacts themselves and the poor quality of dig-
itization, thereby limiting the scope of analysis. However, recent advances in machine learning have shown
promising results in enabling the analysis of vast amounts of data from digitised manuscripts. Nevertheless,
the task of detecting patterns in these manuscripts is further complicated by the lack of annotations and the
small size of many patterns, which can be smaller than 0.1% of the image size. In this study, we propose to ex-
plore the possibility of detecting small patterns in digitised manuscripts using only a few annotated examples.
We also propose three detection datasets featuring three types of patterns commonly found in manuscripts:
words, seals, and drawings. Furthermore, we employed two state-of-the-art deep learning models on these
novel datasets: the FASTER ResNet and the EfficientDet, along with our general approach for standard eval-
uations as a baseline for these datasets.
Object detection is a task in computer vision that in-
volves locating and identifying objects within an im-
age or video. There have been several advances in
object detection in recent years. One of the main
areas of progress has been in the development of
deep learning-based approaches, which have achieved
state-of-the-art results on a number of benchmarks.
The use of visual-pattern detection in manuscript
research is crucial for addressing various research
queries. This technology enables scholars to effi-
ciently explore digitised manuscripts, locating rele-
vant images through specific patterns. It enhances
searchability for textual content and visual elements
like seals and drawings. Even when Handwriting
Recognition (HTR) is viable, the patterns tied to re-
search questions may pertain to specific visual styles
within the handwriting itself, such as that of a partic-
ular scribe.
Detecting visual patterns in historical manuscripts
presents challenges distinct from object detection
tasks, where objects typically have clear boundaries.
Unlike standard benchmarks with well-defined ob-
jects like animals or vehicles, patterns in manuscripts
may lack distinct boundaries, making detection more
challenging. The annotation process for such datasets
is often more time-consuming due to unclear pattern
boundaries, posing an additional challenge in pattern
detection for historical manuscripts.
Applying deep learning models to object detec-
tion demands extensive training on labelled examples,
specifying object locations and classes in each image.
This requirement, vital yet costly, poses challenges,
particularly for datasets needing specialized annota-
tion. Researchers employ data augmentation to max-
imize annotated data use, but substantial examples
per class remain essential, particularly for manuscript
patterns differing significantly from those in standard
Digitised manuscript annotation typically requires
expert supervision, often from relevant research
fields, yet even with this, some annotations are sub-
jective. Obtaining annotations for more than a few
examples per pattern is challenging and sometimes
impossible. Manuscript images often feature scripts
understood by only a few humanities experts, mak-
ing context-dependent patterns challenging for non-
experts. Additionally, images may suffer degradation
due to poor manuscript preservation or writing sup-
port nature, further complicating the annotation pro-
Mohammed, H. and Jampour, M.
Small Patterns Detection in Historical Digitised Manuscripts Using Very Few Annotated Examples.
DOI: 10.5220/0012269500003654
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 13th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2024), pages 605-612
ISBN: 978-989-758-684-2; ISSN: 2184-4313
Proceedings Copyright © 2024 by SCITEPRESS Science and Technology Publications, Lda.
Figure 1: Examples of detection by the proposed approach using three novel datasets of digitised historical manuscripts are
shown. The detected patterns represent three common types found in manuscripts: seals, drawings, and words. The word
examples have been enlarged for improved visibility.
Finally, patterns in digitised manuscripts often oc-
cupy a small image area, as illustrated in Fig. 1. This
poses a known challenge in computer vision, where
small objects may appear blurry or pixelated due to
limited model input resolution, hindering accurate de-
tection. Additionally, small objects may lack distinc-
tive features, making them harder to identify, compli-
cating the task of accurately distinguishing them from
their surroundings for detection algorithms.
In this research, we present three fully annotated
detection datasets featuring three types of patterns
commonly found in manuscripts: words, seals, and
drawings. We employed two state-of-the-art deep
learning models on these novel datasets: a two-stage
detector and a single-stage detector. Finally, we pro-
pose a general approach to improve detection perfor-
mance on all datasets and evaluate it using standard
object detection metrics to serve as a baseline for fu-
ture studies.
There are two main approaches for object detection:
two-stage and single-stage (Liu et al., 2020; Zaidi
et al., 2022; Jiao et al., 2019). Two-stage approaches
first identify regions of the image that are likely to
contain objects, and then classify those objects and
refine their locations. Examples include the Faster R-
CNN (Ren et al., 2017) and the Region-based Fully
Convolutional Network (R-FCN) (Dai et al., 2016).
Single-stage approaches, on the other hand, aim to
identify and classify objects in a single step, without
first identifying regions that are likely to contain ob-
jects (Ren et al., 2015). Examples include the You
Only Look Once (YOLO) (Redmon et al., 2016) and
EfficientDet (Tan et al., 2020). Single-stage algo-
rithms tend to be faster than two-stage algorithms, but
may have lower accuracy.
The concept of using machine learning to auto-
matically detect patterns in manuscript images has
been around for at least a decade (Yarlagadda et al.,
2011), but progress has been limited due to the lack of
standard and publicly available datasets with ground-
truth annotations. Additionally, the reliance of state-
of-the-art methods on annotated training data has hin-
dered the progress.
However, several pattern detection methods have
been proposed to detect symbols, logos, and other
types of patterns found in documents (Mohammed
et al., 2021; Le et al., 2014; Wiggers et al., 2019).
Some of these methods have been specifically de-
signed to detect patterns in historical documents
and manuscripts (
Ubeda et al., 2020; En et al.,
2016b), and optimized for certain types of patterns
and manuscripts. More recently, a general training-
ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods
Figure 2: (a) Example images and patterns from the SAM dataset. (b) An example from each of the selected seals. The most
complete and clear instances are selected in this figure for better visibility.
free approach has been proposed by Mohammed et
al.(Mohammed et al., 2021) for detecting patterns in
manuscript images. The authors of this work argued
that using a training-free approach can eliminate the
problem of annotations availability and provide state-
of-the-art results. While this approach may be use-
ful for many scholars in manuscript research, it has
two major drawbacks: first, performance can only be
slightly enhanced by adding more examples per pat-
tern, due to the lack of a training phase. Second, the
hand-crafted features used in this research may not be
useful in detecting some types of visual patterns.
There are several benchmark datasets that are
commonly used for evaluating object detection al-
gorithms. One widely used dataset is the PASCAL
Visual Object Classes (VOC) dataset (Everingham
et al., 2010), which consists of images annotated with
bounding boxes around objects of 20 different classes.
Another popular dataset for object detection is the Mi-
crosoft Common Objects in Context (COCO) dataset
(Lin et al., 2014), which consists of images annotated
with bounding boxes around objects of 90 different
These datasets are typically not relevant for pat-
tern detection in manuscript research, as the annotated
objects are everyday items such as cars, planes, and
animals. On the other hand, two challenging datasets
for pattern detection in manuscripts have been pub-
lished in the past few years: the AMADI LontarSet
dataset (Burie et al., 2016), which consists of hand-
writing on palm leaves for word spotting, and the Do-
cExplore dataset (En et al., 2016a), which consists
of medieval manuscripts for pattern detection. De-
spite being valuable contributions, the first dataset is
highly unbalanced, very specific, and some queries
are merely letters or other visual marks. In addition,
the annotation is provided as part of the file name.
The second dataset does not include any annotation.
Therefore, there is a significant demand for datasets,
especially historical data, intended for pattern detec-
The primary motivation for creating these three
datasets is to investigate the possibility of detecting
medium to small patterns in digitised manuscripts us-
ing a small number of annotated examples while both
small patterns and a low number of annotated data are
open challenges. To this end, the datasets were chosen
to represent different types of typical patterns found
in manuscripts. All of the datasets are annotated us-
ing the Pascal VOC format and saved as XML files.
All datasets are split into training, validation, and test
sets; however, one can alter the splits based on the
requirements of individual experiments.
The distribution of pattern instances per class in
the training subset is kept balanced as much as pos-
sible in order to focus on the main research question
and to make interpretation of results easier. Further-
more, the resolution of images in all datasets is kept
high enough to preserve the visual features of small
patterns. Finally, all images are saved in ”.jpg” for-
mat to standardise any required image processing.
The main challenges in all of the datasets pre-
sented in this work are the extremely limited number
of training samples (down to only three examples) per
pattern and the small size of many instances compared
to the image size. In addition, each dataset poses a dif-
ferent set of challenges, such as fading, low contrast,
arbitrary orientation, interclass similarities, and etc.
3.1 Dataset of Seals in Arabic
Manuscripts (SAM)
A dataset of seals in Arabic manuscripts has been cre-
ated from the publicly available images of the “Staats-
bibliothek zu Berlin” in (van Lit, 2020). Sample im-
ages of different seals are presented in Fig. 2a. Only
seals with a minimum of 4 occurrences in different
images have been selected, resulting in 8 different
seals and 77 images in total. The complete statis-
Small Patterns Detection in Historical Digitised Manuscripts Using Very Few Annotated Examples
(a) Example images from DMM dataset with various chal-
lenges such as fading, orientation.
(b) One example from each of the selected drawings in the
DMM dataset.
Figure 3: Example images and patterns in the DMM
tics are provided in Table 1. One example from each
of the selected seals is presented in Fig. 2b. The
SAM dataset is made publicly available in a research
data repository (Mohammed, 2023b) under the Cre-
ative Commons license. As can be seen from the pre-
sented examples, the main challenges for patterns in
this dataset include their small size, fading, low con-
trast, arbitrary orientation, and interclass similarities.
3.2 Dataset of Drawings in Medieval
Manuscripts (DMM)
A subset of 124 images has been selected from the
DocExplore images (En et al., 2016a) in order to
(a) Example images.
(b) One example from each of the selected words. Complete
and clear instances are selected in this figure for better visi-
Figure 4: Example images and patterns from the WPM
create a detection dataset of drawings in medieval
manuscripts. Since the original dataset has been pub-
lished without providing any annotations, we selected
and annotated 8 different patterns in the subset, which
resulted in a total of 268 annotated instances. The
complete statistics are provided in Table 2. Sample
images of different drawing are presented in Fig. 3a,
and one example from each of the selected patterns is
presented in Fig. 3b. The DMM dataset is made pub-
licly available (Mohammed, 2023a) under the Cre-
ative Commons license. Some of the main challenges
for patterns in this dataset include their small size, as
well as the colour and scale variance of different in-
stances for the same pattern.
3.3 Dataset of Words in Palm-Leaf
Manuscripts (WPM)
A dataset of words from colophons found in palm-
leaf manuscripts hailing from Tamil Nadu (a state
in India) has been created from images provided by
Centre for the Study of Manuscript Cultures (CSMC)
for the manuscripts belonging to the Staats- und Uni-
atsbibliothek (SUB) Hamburg, and images pro-
vided by the Biblioth
eque nationale de France (BnF),
the library of the
Ecole franc¸aise d’Extr
eme Orient
(EFEO) in Pondicherry and the Cambridge University
Library for their manuscript collections. All images
ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods
and patterns are selected and annotated by Giovanni
Ciotti from the CSMC within the scope of the activ-
ities of the Palm-Leaf Manuscript Profiling Initiative
(PLMPI). A total of 10 words have been selected and
annotated in 69 images. The complete statistics are
provided in Table 3. Sample images from the WPM
dataset are presented in Fig. 4a, and one example from
each of the selected patterns is presented in Fig. 4b.
The WPM dataset is made publicly available in a re-
search data repository (Mohammed and Ciotti, 2023)
under the Creative Commons license.
Table 1: Number of pattern instances in each subset within
the SAM dataset.
Pattern / No. of Train Validate Test
Seal 1 3 1 7
Seal 2 3 1 2
Seal 3 3 1 30
Seal 4 3 1 1
Seal 5 3 1 7
Seal 6 4 3 3
Seal 7 3 1 0
Seal 8 3 1 0
Total 85
Table 2: Number of pattern instances in each subset within
the DMM dataset.
Pattern / No. of Train Validate Test
Corner Diamond 16 2 96
Letter A 4 1 9
Letter BP 4 1 7
Pine cone 16 1 12
Letter T 4 9 26
Letter D 4 1 34
Statue 8 2 2
Coat Shield 4 1 1
Total 115
Table 3: Number of pattern instances in each subset within
the WPM dataset.
Pattern / No. of Train Validate Test
Varsa 8 1 2
Samvatsa 4 3 3
Yeluti 8 1 1
Srikosa 10 1 2
Karakrta 8 1 1
Svahasta-likhita 5 3 2
Naksatra 6 2 3
Subhadin-attil 5 1 5
Kutu 11 1 3
Pillai 6 2 6
Total 265
The WPM dataset presents a unique set of chal-
lenges in addition to all the aforementioned issues
mentioned in the other two datasets. The patterns in
this dataset are extremely small compared to image
size. If we scale down the images to fit the input size
of the detection models, most of the annotated pat-
terns will be represented by only few pixels with no
meaningful visual features.
Furthermore, the annotated patterns themselves
have no distinctive visual features to define them as
objects with clear boundaries. Most of the visual fea-
tures in each of these patterns exist also in other parts
of the images (e.g. other words). In addition, the
boundaries of these patterns can only be accurately
detected after correctly classifying the patterns, be-
cause they are merely defined by the spacial relations
between the visual features of these patterns (e.g. let-
ters sequence). Moreover, the patterns in this dataset
are handwritten words by different scribes on palm-
leaves. Therefore, the handwriting style can differ
greatly between different instances of the same pat-
tern (word), and the texture of the writing support
(leaf) can differ significantly as well.
Two state-of-the-art models are employed to execute
and evaluate the suggested approach on the three
datasets introduced in this study. The initial model is
Faster R-CNN (Ren et al., 2017), which exemplifies
the two-stage methodology. In its first stage, Faster
R-CNN employs a region proposal network (RPN) to
produce a collection of region proposals, i.e., poten-
tial object locations. In the second stage, these region
proposals are passed through a classifier to determine
the class and location of the objects within the image.
The use of a two-stage approach allows for greater ac-
curacy and efficiency in object detection compared to
single-stage approaches. Faster R-CNN also incorpo-
rates a ResNet (He et al., 2016) architecture, which
utilizes skip connections and batch normalization to
improve the accuracy and efficiency of the model. the
ResNet50 variant will be used for the rest of this work,
as incremental gains are not the focus of this research.
The second model is EfficientDet (Tan et al.,
2020), which represents the single-stage approach.
This model performs object detection in a single
stage using a single neural network. This allows for
faster inference times and a simpler overall architec-
ture. Additionally, EfficientDet utilizes a weighted bi-
directional feature pyramid network (BiFPN) to effi-
ciently combine multi-scale feature maps, leading to
improved performance on small objects.
Small Patterns Detection in Historical Digitised Manuscripts Using Very Few Annotated Examples
Table 4: Detection results of Faster R-CNN and Efficient-
Det models using transfer learning, fine tuning and data
augmentation on the SAM, DMM and WPM datasets.
Model Metric SAM DMM WPM
Faster COCO mAP 0.84 0.56 0.0
R-CNN mAP@0.5 0.99 0.97 0.0
ResNet Recall@1 0.79 0.47 0.0
Efficient- COCO mAP 0.77 0.53 0.0
DetD1 mAP@0.5 0.97 0.86 0.0
Recall@1 0.75 0.42 0.0
The standard parameter values mentioned in the
original publications of the corresponding models are
used in all our experiments. However, the number of
training steps is fixed at 10 thousand in order to make
the results of different experiments comparable and to
speed up the training phase for all experiments. The
details of all used parameters and configurations for
both models are published in (Mohammed, 2023c) as
public research data.
As an evaluation metric, we used the coco mAP
metric which is the average value of the calculated
mAPs at IoU thresholds ranging from 0.5 to 0.95 with
a step of 0.05. In addition, we provided other metrics
in our base results such as mAP at 0.5 and 0.7, and
recall rate.
4.1 Learning from Few Examples
Transfer learning, a valuable technique for limited
annotated datasets (Li et al., 2020), enhances ob-
ject detection performance, particularly with few an-
notated images (Talukdar et al., 2018). This study
employs transfer learning to leverage insights from
a larger dataset, improving pattern detection across
three datasets. Fine-tuning of pre-trained models is
necessary due to dissimilarities between patterns in
these datasets and standard benchmarks. The mod-
els, initially trained on the COCO 2017 dataset with
200,000+ images and 250,000+ annotated objects
(Lin et al., 2014), were trained on 640x640 pixel res-
olution images. During fine-tuning, images in smaller
datasets were resized to 640 pixels on the smaller di-
mension while maintaining the aspect ratio.
Data augmentation enhances model performance
by increasing training data. This study employs ba-
sic augmentations—random jpeg quality, contrast and
brightness adjustment, and random black patches. For
the SAM dataset, 90-degree rotation and vertical flip
augmentations are included due to variant pattern ori-
entations. Table 4 displays performance metrics for
both models across the three datasets, incorporat-
ing the mentioned techniques. Results indicate the
FASTER R-CNN model’s superior performance on
Figure 5: An illustration of the proposed image tiling. The
used example in this illustration is the upper part of an
image from the SAM dataset. Each image is split into
640x640 sub-images, and then the corresponding annota-
tions are mapped properly into their new position within
each tile. The tiles are overlapped by 25% in order to avoid
missing the patterns located at the borders between the tiles.
the SAM and DMM datasets. Subsequent experi-
ments focus solely on the FASTER R-CNN model.
However, both models struggle with the WPM
dataset, primarily due to handwritten words lacking
distinct visual features against a background of simi-
lar words. These words also lack clear boundaries and
distinctiveness, making object recognition challeng-
ing. Additionally, there’s significant intra-class varia-
tion between instances of the same pattern, stemming
from differences in handwriting styles among differ-
ent scribes. Furthermore, images in the WPM dataset
are large, with selected patterns occupying less than
0.1% of the image. Scaling down during training re-
sults in annotated patterns represented by only a few
pixels. While a larger model input could address this
issue, it comes with a substantial increase in compu-
tational cost.
4.2 Detecting Small Patterns
Detecting small objects in images poses challenges
for deep learning models (Tian et al., 2018) due to
several factors. Small objects often have fewer pixels,
providing less visual information for the model to ex-
tract useful features. Furthermore, these objects may
be easily occluded or concealed by other elements in
the scene, complicating detection. The complexity
of shapes and features in small objects adds another
challenge for the model to accurately recognize and
classify them.
Image tiling technique is one of the approaches
used to help improving the performance of detecting
small objects in which an input image is divided into
a grid of smaller tiles or patches (Ozge Unel et al.,
2019). Each tile is then processed independently by a
machine learning model, which generates a prediction
for the presence or absence of small objects within
the tile. The predictions from the individual tiles can
ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods
Table 5: The impact of image tiling on the detection perfor-
mance of the FASTER ResNet50 model.
Without COCO mAP 0.84 0.56 0.0
image mAP@0.5 0.99 0.97 0.0
tiling Recall@1 0.79 0.47 0.0
With COCO mAP 0.84 0.66 0.10
image mAP@0.5 1.00 1.00 0.15
tiling Recall@1 0.86 0.61 0.11
then be combined to generate a final prediction for the
entire image.
The key benefit of the image tiling technique is
its ability to enable a machine learning model to con-
centrate on smaller, more manageable regions of an
image instead of processing the entire image simulta-
neously. This is particularly advantageous for small
object detection, enabling the model to focus on ar-
eas more likely to contain small objects and make
more accurate predictions. As part of this approach,
an image tiling mechanism has been implemented to
enhance the performance of small pattern detection
across all three datasets.
Every image in the datasets undergoes division
into sub-images sized 640x640. These tiles have a
25% overlap with adjacent tiles to incorporate fea-
tures from pattern borders effectively. To accommo-
date this, annotations of the original image are re-
calculated by shifting coordinates, ensuring accurate
placement in relevant tiles. Refer to Fig. 5 for an illus-
tration of the proposed image tiling. Proper resizing
of images precedes the tiling process to include image
borders in the tiles.
At the end of each row and column, tiles extend
beyond the image boundary due to the fixed tile size.
Therefore, we shift these tiles inwards so that they
are contained within the image boundary. As a result,
these tiles have a larger overlap area with their pre-
ceding tiles. The overlap increment is directly pro-
portional to the number of shifted pixels.
Concerning annotated patterns on the borders, in-
clusion occurs only when the overlap between the an-
notation and the tile is 50% or more. This ensures that
all annotated patterns are incorporated into at least
one tile, given their size is not comparable to the tile
size—a condition met in all three datasets. The im-
pact of image tiling on detection performance is evi-
dent in Table 5. The proposed image tiling technique
markedly improves detection performance across all
datasets, including the WPM dataset.
Improved detection results on the WPM dataset
are anticipated by incorporating advanced augmen-
Table 6: The base results for the SAM, DMM and WPM
datasets obtained using the FASTER R-CNN ResNet50
R-CNN val/test val/test val/test
COCO mAP 0.84 / 0.78 0.66 / 0.64 0.10 / 0.09
mAP@0.5 1.00 / 0.87 1.00 / 0.89 0.15 / 0.13
Recall@1 0.86 / 0.77 0.61 / 0.56 0.11 / 0.10
tations, a larger model-input size, more training
steps, and applying the dropout technique. However,
achieving high performance on handwritten patterns
requires further research. Consequently, we encour-
age researchers in the community to explore these
possibilities. The base results on both validation and
test sets are outlined in Table 6.
We explored detecting small patterns in digitised
manuscripts with limited annotated examples per pat-
tern. Three detection datasets were created and anno-
tated, featuring words, seals, and drawings commonly
found in manuscripts. Challenges included limited
training samples, small instance size, fading, and in-
terclass similarities. Two deep learning models were
tested, namely the FASTER ResNet and the Efficient-
Det, and detection performance was reported using
COCO metrics. A general approach was proposed to
serve as a baseline for these datasets, utilizing stan-
dard techniques and image tiling. While improve-
ments were made, performance on the WPM dataset
remained poor due to factors such as lack of saliency
and intra-class variations. Therefore, further research
is required to enhance the performance on such types
The research for this work was funded by the
Deutsche Forschungsgemeinschaft (DFG, German
Research Foundation) under Germany’s Excellence
Strategy EXC 2176 ‘Understanding Written Arte-
facts: Material, Interaction and Transmission in
Manuscript Cultures’, project no. 390893796. The
research was conducted within the scope of the Cen-
tre for the Study of Manuscript Cultures (CSMC) at
at Hamburg.
In addition, we would like to thank Giovanni
Ciotti for providing, selecting, and annotating all the
images and patterns in the WPM dataset, and Aneta
Small Patterns Detection in Historical Digitised Manuscripts Using Very Few Annotated Examples
Yotova for annotating the SAM and DMM datasets
and preparing the ”Number of instances” tables.
Burie, J.-C., Coustaty, M., Hadi, S., Kesiman, M. W. A.,
Ogier, J.-M., Paulus, E., Sok, K., Sunarya, I. M. G.,
and Valy, D. (2016). ICFHR competition on the anal-
ysis of handwritten text in images of balinese palm
leaf manuscripts. In 15th International Conference
on Frontiers in Handwriting Recognition, pages 596–
Dai, J., Li, Y., He, K., and Sun, J. (2016). R-fcn: Object de-
tection via region-based fully convolutional networks.
Advances in neural information processing systems,
En, S., Nicolas, S., Petitjean, C., Jurie, F., and Heutte,
L. (2016a). New public dataset for spotting patterns
in medieval document images. Journal of Electronic
Imaging, 26(1):1 – 15.
En, S., Petitjean, C., Nicolas, S., and Heutte, L. (2016b).
A scalable pattern spotting system for historical docu-
ments. Pattern Recognition, 54:149 – 161.
Everingham, M., Van Gool, L., Williams, C. K., Winn, J.,
and Zisserman, A. (2010). The pascal visual object
classes (voc) challenge. In International Conference
on Computer Vision, pages 404–417.
He, K., Zhang, X., Ren, S., and Sun, J. (2016). Deep resid-
ual learning for image recognition. In Proceedings of
the IEEE Conference on Computer Vision and Pattern
Recognition, pages 770–778.
Jiao, L., Zhang, F., Liu, F., Yang, S., Li, L., Feng, Z., and
Qu, R. (2019). A survey of deep learning-based object
detection. IEEE access, 7:128837–128868.
Le, V. P., Nayef, N., Visani, M., Ogier, J.-M., and De Tran,
C. (2014). Document retrieval based on logo spotting
using key-point matching. In 2014 22nd international
conference on pattern recognition, pages 3056–3061.
Li, X., Grandvalet, Y., Davoine, F., Cheng, J., Cui, Y.,
Zhang, H., Belongie, S., Tsai, Y.-H., and Yang, M.-
H. (2020). Transfer learning in computer vision tasks:
Remember where you come from. Image and Vision
Computing, 93:103853.
Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P.,
Ramanan, D., Doll
ar, P., and Zitnick, C. L. (2014).
Microsoft coco: Common objects in context. In Euro-
pean Conference on Computer Vision, pages 740–755.
Liu, L., Ouyang, W., Wang, X., Fieguth, P., Chen, J., Liu,
X., and Pietik
ainen, M. (2020). Deep learning for
generic object detection: A survey. Int. journal of
computer vision, 128(2):261–318.
Mohammed, H. (2023a). Dataset of drawings in medieval
manuscripts (dmm).
Mohammed, H. (2023b). Dataset of seals in arabic
manuscripts (sam).
Mohammed, H. (2023c). Model Parameters of FASTER
ResNet and EfficientDet Model Parameters of
FASTER ResNet and EfficientDet.
Mohammed, H. and Ciotti, G. (2023). Dataset of words in
palm-leaf manuscripts (wpm).
Mohammed, H., M
argner, V., and Ciotti, G. (2021).
Learning-free pattern detection for manuscript re-
search. International Journal on Document Analysis
and Recognition (IJDAR), 24(3):167–179.
Ozge Unel, F., Ozkalayci, B. O., and Cigla, C. (2019). The
power of tiling for small object detection. In Proceed-
ings of the IEEE/CVF Conference on Computer Vision
and Pattern Recognition Workshops, pages 0–0.
Redmon, J., Divvala, S., Girshick, R., and Farhadi, A.
(2016). You only look once: Unified, real-time object
detection. In Proceedings of the IEEE conference on
computer vision and pattern recognition, pages 779–
Ren, S., He, K., Girshick, R., and Sun, J. (2017). Faster r-
cnn: Towards real-time object detection with region
proposal networks. IEEE Transactions on Pattern
Analysis & Machine Intelligence, 39(06):1137–1149.
Talukdar, J., Gupta, S., Rajpura, P. S., and Hegde, R. S.
(2018). Transfer learning for object detection using
state-of-the-art deep neural networks. In 2018 5th In-
ternational Conference on Signal Processing and In-
tegrated Networks (SPIN), pages 78–83.
Tan, M., Pang, R., and Le, Q. V. (2020). Efficientdet: Scal-
able and efficient object detection. In Proceedings
of the IEEE/CVF conference on computer vision and
pattern recognition, pages 10778–10787.
Tian, Y., Li, B., Chen, C., Fu, Y., and Huang, Q. (2018).
Tiny object detection in dense crowds. In Proceed-
ings of the European Conference on Computer Vision
(ECCV), pages 497–513.
van Lit, L. C. (2020). Seals from the staatsbibliothek zu
berlin and their automated detection.
Wiggers, K. L., Britto, A. S., Heutte, L., Koerich, A. L.,
and Oliveira, L. S. (2019). Image retrieval and pat-
tern spotting using siamese neural network. In 2019
International Joint Conference on Neural Networks
(IJCNN), pages 1–8. IEEE.
Yarlagadda, P., Monroy, A., Carque, B., and Ommer, B.
(2011). Recognition and analysis of objects in me-
dieval images. In Koch, R. and Huang, F., editors,
Computer Vision ACCV 2010 Workshops, pages
296–305, Berlin, Heidelberg. Springer Berlin Heidel-
Zaidi, S. S. A., Ansari, M. S., Aslam, A., Kanwal, N., As-
ghar, M., and Lee, B. (2022). A survey of modern
deep learning based object detection models. Digital
Signal Processing, page 103514.
Ubeda, I., Saavedra, J. M., Nicolas, S., Petitjean, C., and
Heutte, L. (2020). Improving pattern spotting in his-
torical documents using feature pyramid networks.
Pattern Recognition Letters, 131:398 – 404.
ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods