Teaching AI to See Flowers: How Preprocessing Influences Deep Learning for Flower Segmentation
- Beschreibung
Are you curious about how artificial intelligence can advance science beyond all the buzzwords? In this project, you’ll explore the practical side of AI by evaluating how the preparation of data influences the performance of deep learning models in biological research. Using images of the Siberian Iris, you’ll test how different preprocessing strategies affect segmentation quality, helping to improve computer-based methods for studying plant ecology, genetics, and conservation.
Background
Methodology (Segmentation and deep learning??):
Flower morphology and color pattern plays an important role in plant-pollinator interactions and therefore in plant reproductive success. In recent years, computer-based image analysis has become an important tool for studying such traits. One powerful approach is the use of deep learning segmentation models, algorithms that can automatically identify and separate different parts of an image, such as the petal areas of a flower. These methods are attractive because, once trained, they can analyze large numbers of images quickly and with relatively high accuracy. However, their performance depends strongly on how the input images are prepared. Even small choices in preprocessing, for example adjusting brightness and color scales or artificially expanding the dataset, can make a noticeable difference in how well a model learns and how accurate the results are. For each new study system, systematic testing of preprocessing strategies is therefore essential.Ecology (Why Iris?):
The Siberian iris (Iris sibirica) provides an excellent case study for such tests. This perennial species is well known for its large flowers with complex and colorful patterns. Its petals typically display contrasting regions of violet, orange and white, with distinct venation, making them both ecologically important for pollinator attraction and technically challenging for computer vision tasks. Although I. sibirica still occurs in many local populations from Central Europe to Central Asia, it is considered endangered. Gaining insights into its reproductive strategies, such as how flower patterns contribute to pollinator attraction, can therefore support ongoing conservation efforts. The combination of striking visual diversity, ecological relevance, and conservation importance makes I. sibirica an ideal model for exploring how image preprocessing influences deep learning based flower pattern segmentation.Tasks
- Pipeline familiarization: Learn to use an R-based image segmentation pipeline and understand the preprocessing functions provided.
- Data preparation: Extend the training dataset by manually annotating additional flower images.
- Systematic testing: Combine preprocessing functions into a testing framework, train segmentation models under different preprocessing conditions.
- Evaluation: Compare results using segmentation accuracy metrics and statistically interpret which preprocessing steps have the strongest effects.
No prior coding experience is required, just curiosity, motivation, and a willingness to learn!
Expected Outcome
- A statistical comparison of preprocessing strategies, showing their relative effect on segmentation accuracy.
- A reproducible testing pipeline that can be applied to other plant species or imaging datasets.
- Insight into how biological flower features interact with deep learning models.
- Projektzeitraum
- Wintersemester 2025/2026
- Bewerbungszeitraum
- 13. bis 27.10.2025
- Durchführung
- nach Absprache
- Details zu Projektzeitraum und Durchführung
You can work on this project remotely or in person, with a schedule tailored to your own pace and semester plan.
- Studienfach
- offen für alle Studienfächer
- Betreuende
- Ryck Leberecht
- Institut
- Institut für Biologie (190) (Plant Evolutionary Biology, 190b)
- Sprache
- deutsch/englisch
- Teilnehmendenanzahl
- min. 1, max. 2
- Arbeitsaufwand
-
ca. 180 Stunden pro Teilnehmende:r
| 6
ECTS-Punkte
Arbeitsaufwand (Stunden und ggf. ECTS) sind ungefähre Angaben. Die tatsächlich vergebenen ECTS-Punkte ergeben sich aus der tatsächlich geleisteten Arbeit.
- Für dieses Projekt ist kein Motivationsschreiben des Studierenden erforderlich
- Projektart
- experimentell
- Lernziele
-
Die Teilnehmende lernen in diesem Projekt:
- To apply key techniques in image processing, computer vision, and deep learning.
- To understand image augmentation and preprocessing strategies and how they influence model performance.
- To use statistical methods to evaluate, interpret and compare segmentation accuracy across preprocessing strategies.
- To connect computational skills to ecological research and learn how deep learning can support ecological studies.
- Anmerkungen für Studierende
The project is planned to be in English but you’re going to be supervised by a native German speaker so don´t worry. Take it as an opportunity to practice your English.
If you have questions regarding the project, you can always contact me via: ryck.leberecht@uni-hohenheim.de