Publikationsserver
Segmentation-based Image Similarity Search
With the rapid and unprecedented growth of digital images, the need for effective image similarity search systems has become more important than ever. The application scenarios for image similarity search are numerous, ranging from e-commerce (where it enables customers to find products through imag...
Saved in:
| Main Author: | |
|---|---|
| Contributors: | |
| Format: | Doctoral Thesis |
| Language: | English |
| Published: |
Philipps-Universität Marburg
2024
|
| Subjects: | |
| Online Access: | PDF Full Text |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | With the rapid and unprecedented growth of digital images, the need for effective image similarity search systems has become more important than ever. The application scenarios for image similarity search are numerous, ranging from e-commerce (where it enables customers to find products through image queries), over healthcare (where it supports diagnosis by comparing medical images), to digital archiving (where it helps to organize and access large volumes of visual data). Furthermore, there has been tremendous progress in the field of image segmentation in recent years, suggesting that image similarity search could possibly benefit from image segmentation.
This thesis provides contributions to two primary research areas: (1) detection and segmentation, and (2) image similarity search. Two approaches are presented in the area of detection and segmentation: (a) a novel deep learning based workflow for automatic detection, alignment, and recognition of textual stamps on digitized index cards; (b) a novel cell segmentation approach for fluorescence microscopy images of morphologically complex eukaryotic cells.
In the area of image similarity search, we propose a novel approach to better understand the user’s search intent. Moreover, we present a novel two-stage approach based on multi-index hashing to integrate deep hashing into Elasticsearch with query times comparable to state-of-the-art similarity search methods.
An important contribution of the thesis is a novel approach that combines insights from both domains into segmentation-based image similarity search, proposing the use of segmented images to enable querying image databases for specific regions within images. A novel versatile region-based similarity search approach for images couples two foundation models and enables users to utilize point, box, and text prompts to search for similar regions.
Finally, the thesis explores the practical implementation of image similarity search in different application domains. Real-world systems for analyzing large-scale image and video data benefit substantially from image similarity search, and image similarity search accelerates data acquisition and labeling when iteratively training specialized deep learning models. |
|---|---|
| DOI: | 10.17192/z2024.0212 |