Details

Handwritten character recognition using deep learning on a mobile device
ID Koshtrevska, Marija (Author), ID Pejović, Veljko (Mentor) More about this mentor... This link opens in a new window

.pdfPDF - Presentation file, Download (681,70 KB)
MD5: A72D01A22C55B4034FF90AD01BDAEF2F

Abstract
Adapting handwritten character recognition models to the unique writing styles of individual users presents a significant challenge. To address this, we developed a comprehensive system that enables on-device personalization. The system integrates a custom Android application for structured data col- lection from 19 participants, image preprocessing techniques to reduce do- main differences between public datasets and our collected user data, and a per-user model training strategy. Personalization significantly improved per- formance, with the average accuracy of the baseline model for general hand- written characters increasing from 43.77% to 83.51% after adaptation for individual users. Generalization was evaluated using Leave-One-Out Cross- Validation. The training pipeline was designed for on-device execution, con- suming only 15% of CPU utilization and 220 MB of memory, confirming the practical feasibility of implementing personalized recognition systems on mobile devices.

Language:English
Keywords:handwritten character recognition, on-device training, model personalization, deep learning
Work type:Bachelor thesis/paper
Typology:2.11 - Undergraduate Thesis
Organization:FRI - Faculty of Computer and Information Science
Year:2025
PID:20.500.12556/RUL-172857 This link opens in a new window
COBISS.SI-ID:250169347 This link opens in a new window
Publication date in RUL:11.09.2025
Views:978
Downloads:251
Metadata:XML DC-XML DC-RDF
:
Copy citation
Share:Bookmark and Share

Secondary language

Language:Slovenian
Title:Prepoznavanje ročno napisanih znakov z uporabo globokega učenja na mobilni napravi
Abstract:
Razlike v slogu pisanja med posamezniki predstavljajo izziv pri prepoz- navanju rokopisa. V ta namen smo razvili celovit sistem, ki omogoča per- sonalizacijo neposredno na mobilni napravi. Sistem vključuje namensko Android aplikacijo za zbiranje podatkov, pri čemer je v raziskavi sodelo- valo 19 udeležencev. Poleg tega uporablja tehnike predobdelave slik za zmanjšanje domenske razlike med javno in lastno zbirko podatkov ter strate- gijo treniranja modelov za vsakega uporabnika posebej. Personalizacija mod- ela je znatno izboljšala uspešnost, saj se je povprečna natančnost, dosežena z osnovnim modelom za splošno rokopisno pisavo, povprečno povečala z 43,77% na 83.51% po prilagoditvi za posamezne uporabnike. Zmožnost pos- ploševanja modela smo preverili z prečno preverjanje z izpuščanjem enega (Leave-One-Out Cross-Validation). Celoten postopek učenja je bil zasnovan za izvajanje na mobilni napravi, pri čemer je učenje na napravi porabilo 15% procesorske moči in 220 MB pomnilnika. To potrjuje praktično izvedljivost implementacije personaliziranih sistemov prepoznavanja na mobilni napravi.

Keywords:prepoznavanje rokopisnih znakov, učenje na napravi, personalizacija modelov, globoko učenje

Similar documents

Similar works from RUL:
Similar works from other Slovenian collections:

Back