Retail Store Objects Recognition and Counting System

Choo, Su Ming (2023) Retail Store Objects Recognition and Counting System. Final Year Project (Bachelor), Tunku Abdul Rahman University of Management and Technology.

Text Full Text - Choo Su Ming.pdf Restricted to Registered users only Download (4MB)

Abstract

The retail industry is witnessing significant advancements in AI technology. However, research has revealed that there is a need to improve the accuracy of object recognition and counting systems. To address this, a Retail Store Objects Recognition and Counting System, which uses a dataset of eight classes, namely Box, Bottle, Instant Noodle, Canned, Toilet Paper, Tupperware, Jar, and Oil is proposed. This dataset is loaded into a YOLO v5 model for training and object detection, and is further improved using YOLO v6, YOLO v7, and YOLO v8 algorithms. These algorithms consist of key processes such as convolution, max pooling, prediction, and Non-Maximum Suppression (NMS). The system's performance is evaluated using the confusion matrix, which determines accuracy, precision, recall, and mean average precision (mAP). The results demonstrate that YOLO v8x achieves the highest training map_0.5 score of 98.2%, showcasing remarkable accuracy compared to YOLO v5, v6, and v7. At the same times, highest testing accuracy of 82.55% was obtained. The work demonstrates that increasing model complexity with more layers, parameters, gradients, and FLOPs leads to improved accuracy in identifying and classifying objects. This study paves the way for further advancements in object recognition and counting systems for the retail industry.