Object Detection System for Self-Checkout Cashier System Based on Faster Region-Based Convolution Neural Network and YOLO9000

Abstract

This paper will discuss an alternative to replace barcodes in the form of object detection based on deep learning which reads the overall feature of an object, so defects and irregularly shaped products do not hinder the reading of the product in processing purchases. Two deep learning-based object detection models, Faster Regional Convolutional Neural Network (Faster R-CNN) and You Only Look Once 9000 (YOLO9000), were tested for their performance in the training phase and real-time implementation. The result of training phase testing shows that the Faster R-CNN model is more accurate and efficient with an mAP of 88.2%, a training time of 1175.6 seconds/epoch, and memory usage of 1.611 GB. The result of the real-time testing of the model shows that Faster RCNN has a high accuracy of 67.1%, but YOLO9000 has a very fast prediction speed of 0.023 seconds/frame. The result of simulation testing shows that YOLO9000 can read products at a speed of 67.40 seconds which is comparable to the speed of a barcode scanner-based cash register that can read products at a speed of 65.77 seconds.