Zero-Shot Sketch-Based Image Retrieval
Тип публікації :
Магістерська робота
Дата випуску :
2021
Автор(и) :
Korobka Anastasiia Mykolayivna
Мова основного тексту :
Англійська
eKNUTSHIR URL :
Цитування :
[APA 7] Korobka, A. M. (2021). Zero-Shot Sketch-Based Image Retrieval [Master's thesis, Київський національний університет імені Тараса Шевченка]. eKNUTSHIR. https://ir.library.knu.ua/handle/123456789/2976
[ДСТУ] Korobka A. M. Zero-Shot Sketch-Based Image Retrieval : кваліфікаційна робота магістра : 12 Інформаційні технології. Kyiv, 2021. 57 p. URL: https://ir.library.knu.ua/handle/123456789/2976 (date of access: 25.07.2026).
We have analyzed existing methods of sketch-based image retrieval and classification on ZSL. We have noticed restrictions for input data format as detailed keywords description or sketch-image pairs, which limit the possibility of using arbitrary dataset. In our proposed methodology we suggest a preprocessing method for image-to-sketch transfer to simplify conditions for training models. Also, we have trained existing solutions as MobileNet and ADGPM on sketch-based data and speed up retrieval process with FAISS indexing.
For preprocessing phase we analyze and compare the performance of existing approaches from Computer Vision and Deep Learning: Canny Edge Detector, a style transfer net, Adaptive Threshold, and Salient Contours network. Next, we have implemented some experiments based on different effective nets from classical Computer Vision (MobileNet), classical Zero-Shot (ADGPM), and sketch-based Classification (Doodle2Search) tasks. Furthermore, we have proposed to use the FAISS library for indexing to improve image search based on output embeddings of nets in an effective in time and accurate way. We have compared our performance based on local and general benchmarks.
Finally, we have implemented a Demo application on sketch-based image retrieval to demonstrate the effectiveness of our methodology and the possibility of using such a solution for searching some photos in the Gallery app on devices. Also, we have optimized our proposed approaches with the quantization process, so we can use the Demo in real time, without tangible delay locally without a huge amount of memory.
For preprocessing phase we analyze and compare the performance of existing approaches from Computer Vision and Deep Learning: Canny Edge Detector, a style transfer net, Adaptive Threshold, and Salient Contours network. Next, we have implemented some experiments based on different effective nets from classical Computer Vision (MobileNet), classical Zero-Shot (ADGPM), and sketch-based Classification (Doodle2Search) tasks. Furthermore, we have proposed to use the FAISS library for indexing to improve image search based on output embeddings of nets in an effective in time and accurate way. We have compared our performance based on local and general benchmarks.
Finally, we have implemented a Demo application on sketch-based image retrieval to demonstrate the effectiveness of our methodology and the possibility of using such a solution for searching some photos in the Gallery app on devices. Also, we have optimized our proposed approaches with the quantization process, so we can use the Demo in real time, without tangible delay locally without a huge amount of memory.
Галузі знань та спеціальності :
12 Інформаційні технології
122 Комп’ютерні науки
Файл(и) :![Ескіз]()
Вантажиться...
Формат :
Adobe PDF
Розмір :
1.76 MB
Контрольна сума :
(MD5):bade2e708155405211898f828c1a4f7f
Якщо не вказано інше, ця робота розповсюджується на умовах ліцензії Creative Commons Attribution-NonCommercial 4.0 International