Eyu, Jer Min (2022) Application development for product recognition on-shelf with deep learning. Final Year Project, UTAR.
Abstract
Negligence of empty shelf and high human intervention have been the issues that leads to low customer retention in brick-and-mortar stores. Hence, state-ofthe-art deep learning models are trained and compared for product recognition on-shelf and an application to detect empty shelf with the best deep learning model is developed. The three compared models are YOLOv3, YOLOv4, and YOLOv5 to recognise Philips 9w bulb and Philips 11w bulb in lighting equipment store. YOLOv5 outperformed in three compared models. Then, a dataset of 200 images of empty shelf had been filtered and annotated for empty shelf detection training. The accuracy of implemented model is as high as 99.5%. The developed application has successfully detected empty space on the shelf and sent Telegram message to remind the retailers to restock. To identify if the system works in real-life scenario, usability testing is carried out by three employees from the stores and two lecturers from the university. The overall SUS score is 91%. However, the limitations of the developed application should be overcame for real-world implementation. These included that the smooth preview of surveillance tool is not produced, all black and long items are detected as empty shelf, adaptation of different stocking method is low, and crashing of the application under low connectivity.
Actions (login required)
![[img]](http://eprints.utar.edu.my/style/images/fileicons/application_pdf.png)
