Boosting Checkout Speed and Accuracy with AI

In this article, we explore how artificial intelligence (AI) can dramatically improve the speed and accuracy of retail checkout processes, helping businesses reduce training costs and improve customer satisfaction.

Why Use AI at Checkout?

Manually checking and totaling items is time-consuming. Cashiers must accurately remember the price of each item, yet mistakes can still happen, which negatively impacts checkout efficiency. What if we let AI take care of these repetitive tasks with speed and precision?

Cashiers have to check each item individually and calculate the total price, which is inefficient and prone to human error such as incorrectly entering the prices of the items or misremembering them, resulting in extended checkout times and even potential disputes with customers. In addition, companies must invest time and resources to train new employees to memorize item prices and the use of traditional cash registers, further impacting checkout efficiency.

Wouldn't it be attractive if AI could handle all the tasks and complete them accurately within just a few seconds?

How Does It Work?

With an AI model capable of recognizing customer-selected items, cashiers no longer need to memorize item names and prices. Selected items are detected and the total price is automatically calculated in a second.

So, how is such an AI model trained? Training a detection model requires annotated data of the target objects. Fortunately, the checkout scene is relatively simple compared to other use cases that involve more complex or cluttered environments. As a result, it's not necessary to prepare thousands of annotated images for each item, but having more data is always better.

General Scene vs. Checkout Scene

There are some key differences between general detection problems and detection in the checkout scene:

1. Large, clearly visible items – Small-object detection is rarely a concern.

2. Stable environment – Lighting, layout, and camera view are consistent.

3. Minimal clutter and occlusion – Items are placed with separation.

4. Simple backgrounds – Allows easy foreground extraction.

5. Limited, predefined item classes – Detection is confined to a closed set.

6. Fixed camera viewpoint – Input images are highly consistent.

7. Controllable data collection – Training images can be captured under ideal conditions.

However, you still need to ensure that the model is trained with data with enough diversity, so we should manage to "produce" more samples from the existing ones.

By applying data augmentation techniques, we can generate a diverse set of training samples from a limited number of raw images.

Those training samples are used to train a YOLO11 nano(YOLO11n) model, which is suitable for lightweight detection tasks. For edge deployment, the model is quantized to boost inference speed with minimal accuracy loss. Quantization is done on Qualcomm AI Hub and you need to upload the model and calibration set onto it. On the Qualcomm QCS6490 platform, quantized YOLO11n achieves the inference time of approximately 5 milliseconds per image thanks to its Hexagon Tensor Processor (HTP).

Below is the example of detecting 6 kinds of items using YOLO11n and totaling the price.

This animation demonstrates real-time item detection and automatic price totaling. Even when items are placed separately at varying positions, the YOLO11n model detects them reliably. For optimal detection performance, it is recommended that items be placed individually, ensuring they are not touching or overlapping. The distance between items can vary as long as they remain visually separated.

More Accurate and Time-Efficient Checkout Experience

With the help of AI detection models, businesses such as bakeries and delis, where cashiers typically identify items and manually input prices, can significantly reduce employee training costs, speed up the checkout process, and minimize human errors. By automating item recognition and price totaling, the system not only improves operational efficiency but also enhances customer satisfaction.