CARTERA: QUICK SHOPPING SOLUTION

Abstract

One of the busiest places in retail stores is the billing counter because it takes a lot of time to recognize the products and then the bill is finally accumulated, and we encounter these problems in our daily routine. This is where a self-checkout system is realized because it is about time when the traditional system is replaced by automated product recognition systems. Because the adaptation of a self-checkout system will be favorable to both social and economic factors. As for the current days of pandemic COVID-19 least interaction is encouraged, another advantage is that the self-checkout systems will replicate busy and labor-intensive stores. This way the overall process will be more frequent than physical operation and timesaving. Recent developments in Computer Vision and Deep Learning have provided a powerful repute to the field because the object detection and recognition techniques have widely improved in both speed and accuracy. However, product detection and recognition using images remains a very demanding task in the sphere of computer vision. This thesis investigates the suitability of running product detection and recognition on mobile devices, where the resources are limited. We experimented to develop a self-checkout system using a custom dataset utilizing state-of-the-art object detection model YOLOv4-tiny by means of an android application. The model YOLOv4-tiny provides much faster object detection on phone with 99.52% Mean Average Precision as compared to SSD MobileNetv2 which gives 95.33%. It is very useful for devices with limited resources. The results emphasize that a self-checkout system is a practicality for the future to come, but with the low processing power a mobile has to offer, and with added issues such as large-scale product classification and data limitations the viability of a self-checkout system is yet to be improved.