On-Device Deep Learning for Retrieving System and User Timestamps from Noisy Chat Images
Year: 2025
Type: conference-paper
Abstract: Extracting information from scanned or digital images are helpful in different real time applications. Optical Character Recognition (OCR) is popular in extraction of information from digital images taken in the ideal lighting condition. But OCR method is not efficient when the images are scanned and clicked with the noisy and blurry environment. Deep learning-based OCR method performs good in text detection and localization in the digital images. Existing methods do not focus on the timestamp detection in screenshots. In this paper, YoloV5 (You Only Look Once) model and Faster RCNN (Faster Region-Convolutional Neural Network) are implemented to locate the region of time stamps from the screenshot images from various messaging applications. Our models take mobile screenshot of conversation applications as input and detect the system timestamps and User timestamps. 5k screenshot images have been collected from 16 messaging applications with various backgrounds. The images are annotated with 3 classes. YoloV5 and faster RCNN models are trained with the annotated dataset. Accuracy of YoloV5 model increased to 97.6 with training it with more dataset than RCNN and also the size is limited to 3.85MB.
Source: Procedia Computer Science
Institution: B.M.S. College of Engineering
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
Funder: Samsung
OpenAlex ID: W4410258146