The initial deployment of this Ready Solution includes the following activities:

• Initial model training—Malong has a portfolio of pretrained AI models that can be deployed for retail POS inventory loss detection. Various models can be tested to determine if a model performs well when it is deployed to retail stores. Typically, the base model requires additional training to fit the customer’s retail inventory and the local environment, including SCO configuration and camera positioning. If the target retail stores have consistent SCOs and camera positions, the model can be deployed to all retail stores after local training. If retailers have more than one model of SCOs and cameras, RetailAI Protect can adjust and perform well in these scenarios.
• Integration with customer SCO—The Malong RetailAI Protect system can be integrated with a POS using APIs from the SCO system. SCO systems, which are either commercially available or custom engineered for large retail chains, typically have an API that sends and receives information through a message brokering service like ActiveMQ. When an item is scanned, the SCO system can return information about the transaction, including the UPC barcode, to the Malong RetailAI Protect solution.
• Deploying across multiple retail branches—This Ready Solution can be deployed easily across various retail stores in several geographic locations. Each retail store can have one or more PowerEdge servers as the edge node. The PowerEdge servers are configured and registered in Azure IoT Hub. Then, AI models such as Malong RetailAI Protect can be deployed to the IoT devices by using a few clicks. If SCOs across various retail branches have similar camera positioning, scanner positioning, lighting, and environmental factors, it is easy to train and deploy the initial model rapidly across all the retail stores, without needing further customization. The AI algorithm uses the visual image of the item being scanned to identify mis-scans or ticket switching. The image of the item being scanned depends on the position of the camera and the lighting of the store. If different retail stores have different camera positioning (for example, one store has an overhead fixed-dome camera while another has cameras next to the scanner that point up), the algorithm must be customized for the two scenarios.