Bachelor's Thesis

Text-Based Search in Image Data

Final Thesis 10.81 MB

Author of thesis: Bc. Denys Dumych

Acad. year: 2025/2026

Supervisor: Ing. Markéta Juránková, Ph.D.

Reviewer: Ing. Michal Hradiš, Ph.D.

Abstract:

This thesis focuses on designing and implementing a system for text-based search in image data, addressing the semantic gap between low-level image representations and natural language. The system combines object detection with multimodal embedding models to index both full images and detected object regions into a shared vector space, enabling users to search image collections using natural language queries or example images, with optional negative constraints to refine results. To support efficient similarity search on larger image collections, the system utilizes vector database indexing techniques for fast nearest-neighbor retrieval. The system was evaluated on a standard benchmark dataset, confirming retrieval accuracy.

Keywords:

text-based image retrieval, semantic search, object detection, multimodal models, CLIP, SigLIP, YOLO, DETR, vector embeddings, similarity search

Date of defence

15.06.2026

Result of the defence

Defended (thesis was successfully defended)

znamkaCznamka

Grading

C

Process of defence

Student nejprve prezentoval výsledky, kterých dosáhl v rámci své práce. Komise se poté seznámila s hodnocením vedoucího a posudkem oponenta práce. Student následně odpověděl na otázky přítomných. Komise se na základě posudku oponenta, hodnocení vedoucího, přednesené prezentace a odpovědí studenta na položené otázky rozhodla práci hodnotit stupněm C.

Topics for thesis defence

  1. How did you come up with the idea of negative queries?

Language of thesis

English

Faculty

Department

Study programme

Information Technology (BIT)

Composition of Committee

prof. Ing. Adam Herout, Ph.D. (předseda)
doc. Mgr. Adam Rogalewicz, Ph.D. (místopředseda)
Ing. Vladimír Bartík, Ph.D. (člen)
Ing. Michal Hradiš, Ph.D. (člen)
Ing. Josef Strnadel, Ph.D. (člen)

Supervisor’s report
Ing. Markéta Juránková, Ph.D.

Student v průběhu celého roku aktivně pracoval, samostatně navrhoval řešení a své návrhy pravidelně konzultoval. Splnil všechny body zadání a práci dokončil s dostatečným předstihem. Proto uděluji výsledné hodnocení vedoucího práce A.

Evaluation criteria Verbal classification
Information about assignment

Práce byla průměrně náročná, zaměřená na vytvoření softwaru s důrazem na jednoduché a intuitivní ovládání za použití stávajících technologií. Všechny body zadání byly splněny a dosažené výsledky splňují požadavky zadání.

Work with literature

Student aktivně přicházel s vlastními návrhy řešení a úprav systému

Activity during solution, consultations, communication

Student práci v průběhu roku pravidelně konzultoval a na schůzky přicházel připraven s aktuálním stavem řešení a vhodnými dotazy ke konzultaci.

Activity during completion

Práce byla dokončena včas a její finální podoba byla průběžně konzultována.

Publication activity, awards

Výstup práce je vhodný ke zveřejnění jako open-source software.

Points proposed by supervisor: 92

Grade proposed by supervisor: A

Reviewer’s report
Ing. Michal Hradiš, Ph.D.

The student created a simple application that includes two interesting elements: a combination of image and object retrieval, and negative queries. However, he did not review the state of the art and did not work systematically when designing the application and its UI.

Evaluation criteria Verbal classification Points
The difficulty of the assignment

Evaluation level: more difficult assignment

The topic requires the student to create a web application working with potentially large datasets and using state-of-the-art vision-text models.

Presentation level of the technical report

The text of the thesis is understandable and sufficiently explains the developed application. However, it does not include a broader overview of the state of the art in text-based image search itself, an analysis of potential users and use cases, and it also lacks a systematic design of the application and the user interface. Conversely, I find the section on the “history” of object detection unnecessary, as well as some of the details about training the CLIP and SigLIP models.

70
Formal preparation of a technical report

The thesis is rather pain. In terms of language, it contains only a few minor problems. There are more typographical issues, but still within an acceptable amount. I appreciate that the adapted figures 2.3, 3.1, and 3.2 are vector graphics. I would include the following among the typographical issues:

  • very low quality of Figures 2.1 and 2.2
  • random placement of tables, figures, and code listings in the middle of pages
  • very frequent absence of text between headings
  • “ugly” tables
  • missing punctuation after equations
72
Realisation output

The student created a simple working application which is, however, suitable only for limited local single-user use, and it would probably be impractical to extend it into a “standard” web application. I like that the student incorporated objects and "negative queries".

The student evaluated retrieval quality on the COCO dataset of the CLIP and SigLIP models, of retrieval including objects, and a simple "prompt" modification. He further evaluated processing speed and retrieval latency. These tests are rather basic and leave many other options unexplored.

No user testing was performed.

67
Usability of results

The application can be used for experiments with image retrieval.

The extent to which the requirements of the assignment have been met

Evaluation level: assignment fulfilled

Extent of the technical report

Evaluation level: is within the usual extent

Work with literature

The thesis references 30 sources. I consider many of the sources on object detection unnecessary. Overall, the sources are reasonable, but I am missing a broader overview of the state of the art text-image models and content-based image search.

The sources should be used more diligently in the text, especially in the Introduction, the beginning of Chapter 2, Section 2.2 CNN-Based Detectors, Section 2.2.3 YOLO, and Section 3.1. The sources are mostly used elsewhere, but these sections should include them as well.

Often, an arXiv version of a paper is cited even though the paper was published at a standard peer-reviewed venue or in a journal.

70
Points proposed by reviewer: 70

Grade proposed by reviewer: C

Responsibility: Mgr. et Mgr. Hana Odstrčilová