One of the most common problems slowing down cash desk operations is a necessity to correctly recognise products that are sold bulk, weighed, are seasonal or have no packaging. Commonly, both cashiers and customers make mistakes, but also purposely substitute low value items for expensive ones.
To eliminate the problem companies train cashiers and inform of new products and their varieties daily. Also, some sort of supervision is frequently present, often in a form of CCTV cameras overlooking cask line and self-checkout tills.
The aim of our software is to support cashiers in item recognition, removing uncertainty in their work and provide companies with a tool to eliminate cash line mistakes and frauds.
Golden AI comprises of two elements: graphical overlay for a cash desk software, providing cashier with a picture and code of a product placed on a desk and an AI, self-learning algorithm, located on store servers or in the cloud, which makes the actual recognition process. Implementation of Golden AI begins with creating an environment necessary to train AI in existing and, later, in new products. Appropriate equipment must be used to take high quality pictures of each product from many angles, with and without foil bag, until AI reaches target recognition level. The efficiency of this process largely depends on lightning conditions at the cash desk and transparency of the packaging foil.
The learning process takes place in one, central location by trained personnel. Ready models for AI are then sent to store servers. This setup makes store independent from external connection lines availability and quality. It also speeds up recognition process as massive video data from cash desk cameras is transmitted inside the LAN. As a result, the entire recognition process takes under 1 second per product. Recognition itself works better if the product is rotated under the camera rather than being static. The cashier is presented with 3 pictures of products of which he confirms one. It is also possible to manually override the AI proposal and enter the product code manually.
For the purpose of fraud detection, we suggest negative verification. The system is trained in recognising only items that are expensive, have frequent stock discrepancy or are known to be used to substitute for more expensive merchandise. If such item is recognised by Golden AI but a different item has been scanned by cashier or customer, system will issue a warning message, triggering an appropriate in-store procedure.
The entire process of recognition is logged and stored in system logs. It is used later to analyse reasons for failed recognitions, such as insufficient lighting at a particular cash desk or broken camera, broken package or fraud.