=Paper=
{{Paper
|id=Vol-3740/paper-152
|storemode=property
|title=Identifying Training Data "Fingerprints" Using Border Enhancing Image Processing Methods and Their Ensemble
|pdfUrl=https://ceur-ws.org/Vol-3740/paper-152.pdf
|volume=Vol-3740
|authors=Koki Inoue,Tetsuya Asakawa,Kazuki Shimizu,Kei Nomura,Masaki Aono
|dblpUrl=https://dblp.org/rec/conf/clef/InoueASNA24
}}
==Identifying Training Data "Fingerprints" Using Border Enhancing Image Processing Methods and Their Ensemble==
https://ceur-ws.org/Vol-3740/paper-152.pdf
Identifying Training Data "Fingerprints" Using Border
Enhancing Image Processing Methods and Their Ensemble
Notebook for the inouekokiteam Lab at CLEF 2024
Koki Inoue1,* , Tetsuya Asakawa1 , Kazuki Shimizu2 , Kei Nomura2 and Masaki Aono1
1
Toyohashi University of Technology, 1-1 Hibarigaoka, Tempaku-cho, Toyohashi, Aichi, 441-8580, Japan
2
Toyohashi Heart Center, 21-1Gobutori, Ohyamacho, Toyohashi, Aichi, 441-8071, Japan
Abstract
This paper describes our approach to the Identify training data "fingerprints" task of ImageCLEFmedical GANs
2024. In Task 1, the goal is to detect "fingerprints" within the synthetic biomedical image data to determine
which real images were used in training to produce the generated images. The proposed method uses image
processing as a preprocessing step, and a pre-trained model, Resnet-152, is used for training. We also integrated
the predictions of each model. As a result, the model with histogram equalization was able to outperform the
baseline model trained without preprocessing by 66.6%. The model with prediction integration achieved 63.1%.
Keywords
Image Processing, Integrated the Predictions, Histogram Equalization
1. Introduction
ImageCLEF has been held as part of CLEF since 2003, and ImageCLEF2024 [1] approaches different areas,
including ImageCLEFmedical GANs 2024 [2]. In Task 1 (Task to identify training data "fingerprints"),
the goal is to detect "fingerprints" within the synthetic biomedical image data to determine which real
images were used in training to produce the generated images. We are participating as a member of
the inouekokiteam and are challenging this task. This paper describes the approach used to determine
which images were used in training to create the generative model.
2. ImageCLEF 2024 Dataset
This section describes the dataset for the Identify training data "fingerprints" task of ImageCLEFmedical
GANs 2024 [2]. This task uses two generative models. The dataset contains images used to train each
model, images not used for training, and images generated by the models.
2.1. Development Dataset
The first generative model consists of 200 images annotated as used/not used for training image
generation and 10k generative images generated by model 1. The second generative model consists
of 6k images annotated as used/not used for training image generation, and 10k generative images
generated by model 2.
CLEF 2024: Conference and Labs of the Evaluation Forum, September 09–12, 2024, Grenoble, France
*
Corresponding author.
$ inoue.koki.we@tut.jp (K. Inoue); asakawa.tetsuya.um@tut.jp (T. Asakawa); shimizu@heart-center.or.jp (K. Shimizu);
kein312@gmail.com (K. Nomura); masaki.aono.ss@tut.jp (M. Aono)
0009-0003-9600-1939 (K. Inoue); 0000-0002-8345-7094 (T. Asakawa); 0009-0000-3448-7986 (K. Shimizu);
0000-0003-2838-7844 (K. Nomura); 0000-0003-1383-1076 (M. Aono)
© 2024 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR
ceur-ws.org
Workshop ISSN 1613-0073
Proceedings
2.2. Test Dataset
The test dataset contains two CSV files and two folders, and does not specify which set of images was
used to train the generative model. The ratio of generated to real images is not identical. The first
folder contains 7200 generated images and 4000 real images. The second folder contains 5000 generated
images and 4000 real images.
3. Proposed Method
In this section, we describe our approach to the task of identifying the training data "fingerprints" of
ImageCLEFmed GANs 2024 [2]. We have observed that the color boundaries of the generated images
are often unclear. Therefore, we propose a method that captures the boundary sharpness using a set of
OpenCV [3] image processing functions as preprocessing for both training and prediction. We also
propose a method to integrate the prediction results of each training model into a single result. The
image processing methods used are shown below.
• Binarization
• Histogram Equalization
• Laplacian Process
• Contrast Adjustment
We also propose a method to integrate the predictions of each training model into a single prediction.
A total of five models are used: one model trained without image processing and four models trained
with the image processing described above. The integration procedure is described below.
• Take a majority vote of the five models’ forecasts and make an integrated forecast.
• If the predictions of all five models are not in agreement, a negative result is assumed.
4. Preprocessing by Image Processing
This section describes the image processing preprocessing performed on the development and test
datasets.
4.1. Binarization
Binarization was performed in preprocessing using OpenCV [3]. The image was loaded as grayscale,
and Otsu binarization [4] was performed. It uses the threshold that maximizes the separation between
classes.
4.2. Histogram Equalization
We describe the preprocessing histogram equalization performed using OpenCV [3]. The images were
loaded as grayscale and subjected to histogram equalization. This is a process that transforms the
density so that the histogram of pixel values is uniform throughout.
4.3. Laplacian Process
Laplacian processing was performed using OpenCV [3]. The images were loaded as grayscale and pro-
cessed with a Laplacian filter. It detects edges where the difference in pixel values changes significantly.
Figure 1: Flow of model Predictions
4.4. Contrast Adjustment
Contrast adjustment was performed in preprocessing using OpenCV [3]. The images were loaded as
grayscale images and the contrast was adjusted. It was adjusted with 𝛼=1.5 and 𝛽=0. v′ is the output
pixel value and v is the input pixel value.
v′ = 𝛼 × v + 𝛽 (1)
5. Train
In this section, we describe the training of the model. A pre-trained model from Resnet-152[5] was
used for training. As training data, we used 3100 images each from generated_1 and generated_2 in
the development dataset, for a total of 6200 images as generated, and all images from not_used_1,
used_1, not_used_2, and used_2 as real. A total of 6200 images were considered REAL. In addition to
preprocessing by image processing, random horizontal flipping was applied to the training images.
6. Prediction
In this section, we describe the prediction using the model described in the previous section and the
integration of the prediction results. Test dataset preprocesses the models for prediction by image
processing according to the model used. A total of five models are used for the prediction, one trained
without image processing and four trained with different image processing methods.
6.1. Model Predictions
The prediction for each model is described in the following section. The detailed flow is shown in Figure
1. For the prediction of a trained model without image processing, no image processing is applied to
test dataset. For the trained model with image processing, the same image processing was applied to
test dataset to make predictions.
6.2. Integration of Prediction
We describe the integration of the predictions, using two methods: one with no image processing on
test dataset, and the other with four different image processing methods. For the integration of the
predictions, we used majority voting and perfect agreement. The integration flow is shown in Figure 2.
For perfect agreement, the results were accepted only when all the results predicted by the five models
were in agreement, and rejected when they were not.
7. Submission Results
In this section we describe the results of our team’s submissions. The submissions included predictions
for each of the five models (Run ID: 891-896) and the integration of the predictions (Run ID: 301, 890).
The prediction for the model without added preprocessing (Run ID: 896) was 66.3%. The highest score
Figure 2: Integration of the five models’ predictions
Table 1
Submission Results
M1 M2
Run ID method name Score Acc Prec Recall F1 Acc Prec Recall F1
896 Non-Preprocessed 0.663 0.495 0.497 0.987 0.661 0.5 0.5 0.996 0.66
895 Binarization 0.638 0.484 0.49 0.838 0.619 0.503 0.501 0.951 0.656
894 Contrast Adjustment 0.660 0.491 0.495 0.973 0.656 0.499 0.499 0.993 0.664
892 Histogram Equalization 0.666 0.499 0.499 0.998 0.665 0.501 0.5 0.999 0.667
891 Laplacian Process 0.663 0.484 0.49 0.838 0.619 0.503 0.501 0.951 0.656
301 Majority Voting Non
890 Perfect Agreement 0.631 0.473 0.484 0.805 0.604 0.508 0.504 0.945 0.657
for the prediction using the model with histogram equalization (Run ID: 892) was 66.6%. No score was
returned for majority voting (Run ID: 301), one of the proposed methods. The reason for not returning a
score is believed to be that it produced the same prediction result for all test data. For perfect agreement
(Run ID: 890) the score was 63.1%.
8. Discussion
In this section, we describe the submitted results. The model with histogram equalization and laplacian
processing outperformed the baseline model with no preprocessing (Run ID: 896). Other models with
additional preprocessing underperformed the baseline. This suggests that histogram equalization is an
effective image processing method for detecting "fingerprints" within the synthetic biomedical image
data to determine which real images were used in training to produce the generated images. We were
not able to exceed the baseline for perfect agreement in predictive integration. One possible reason
for this is that histogram equalization was effective, but other image processing methods were not. It
is also possible that the acceptance method of rejecting all predictions if they did not match resulted
in the rejection of accurate predictions. No results were returned for majority voting for prediction
integration. A possible reason for this could be that the prediction was not accepted because it was
used for all images.
9. Conclusion
This paper describes an approach to the identify training data "fingerprints" task of ImageCLEFmedical
GANs 2024[2]. We applied image processing as a preprocessing step, and attempted training and
prediction. We also made predictions for each model, and attempted to integrate the predictions
using majority voting and perfect agreement methods. The results showed that only the models with
histogram equalization and laplacian processing were able to exceed the 66.3% of the models without
image processing that were set as the baseline. Both predictions integration failed to exceed the baseline.
This paper describes an approach to the task of identifying training data "fingerprints" of Image-
CLEFmedical GANs 2024 [2]. We applied image processing as a preprocessing step and attempted
training and prediction. We also made predictions for each model and attempted to integrate the
predictions using majority voting and perfect agreement methods. The results showed that only the
models with histogram equalization and laplacian processing were able to exceed the 66.3% of the
models without image processing, which was set as the baseline. Both prediction integrations failed to
outperform the baseline.
10. Acknowledgments
A part of this research was carried out with the support of the Grant for Toyohashi Heart Center
Smart Hospital Joint Research Course and the Grant-in-Aid for Scientific Research (C) (issue numbers
22K12149 and 22K12040).
References
[1] B. Ionescu, H. Müller, A. Drăgulinescu, J. Rückert, A. Ben Abacha, A. Garcıa Seco de Herrera,
L. Bloch, R. Brüngel, A. Idrissi-Yaghir, H. Schäfer, C. S. Schmidt, T. M. Pakull, H. Damm, B. Bracke,
C. M. Friedrich, A. Andrei, Y. Prokopchuk, D. Karpenka, A. Radzhabov, V. Kovalev, C. Macaire,
D. Schwab, B. Lecouteux, E. Esperança-Rodier, W. Yim, Y. Fu, Z. Sun, M. Yetisgen, F. Xia, S. A. Hicks,
M. A. Riegler, V. Thambawita, A. Storås, P. Halvorsen, M. Heinrich, J. Kiesel, M. Potthast, B. Stein,
Overview of ImageCLEF 2024: Multimedia retrieval in medical applications, in: Experimental
IR Meets Multilinguality, Multimodality, and Interaction, Proceedings of the 15th International
Conference of the CLEF Association (CLEF 2024), Springer Lecture Notes in Computer Science
LNCS, Grenoble, France, 2024.
[2] A. Andrei, A. Radzhabov, D. Karpenka, Y. Prokopchuk, V. Kovalev, B. Ionescu, H. Müller, Overview
of 2024 ImageCLEFmedical GANs Task – Investigating Generative Models’ Impact on Biomedical
Synthetic Images, in: CLEF2024 Working Notes, CEUR Workshop Proceedings, CEUR-WS.org,
Grenoble, France, 2024.
[3] G. Bradski, The OpenCV Library, Dr. Dobb’s Journal of Software Tools (2000).
[4] N. Otsu, A threshold selection method from gray-level histograms, IEEE Transactions on Systems,
Man, and Cybernetics 9 (1979) 62–66. doi:10.1109/TSMC.1979.4310076.
[5] K. He, X. Zhang, S. Ren, J. Sun, Deep Residual Learning for Image Recognition, in: Proceedings of
the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 770–778.