Malong Technologies specializes in AI that provides computer vision technology for retail applications. Malong has developed a state-of-the-art computer vision system that uses deep neural network technology to identify retail products. This technology has been applied successfully to identify both accidental and intentional product scan errors. Malong RetailAI Protect provides a state-of-the art software solution for inventory loss at the POS, which is among the leading causes of inventory loss in brick-and-mortar retail environments. The technology works for both SCO terminals and staffed lanes.
Malong developed an algorithm called CurriculumNet, which is a breakthrough approach to learning directly from noisy and unbalanced visual data—the kind of data that is found in retail environments. With CurriculumNet, the model starts learning the more clearly defined and easier-to-recognize characteristics of a product. It gradually includes more complex learning tasks, such as learning rare and hard-to-distinguish items, into the learning process.
The learning model with CurriculumNet follows three main steps:
These three steps enable the Malong algorithm to learn during normal operations without the need for additional human supervision for annotation. Manual labeling incurs one of the highest costs involved in implementing a deep learning-based solution and is therefore not feasible in large-scale retail scenarios. The Malong algorithm was tested in 2017 at a worldwide image recognition competition called WebVision at CVPR, the premier conference in computer vision. More than 100 scientific research organizations participated in the competition, which was held by Google Research. The Malong algorithm won first place by a wide margin, outperforming the second-place submission by a reduction in relative error rate of nearly 50 percent, as shown in the .
Malong RetailAI Protect is the latest generation of a deep learning model designed for the retail store market. It builds on CurriculumNet to deliver a state-of-the-art algorithm performance. The AI model runs in an Intelligent Video Analytics (IVA) pipeline, using proprietary weakly supervised learning. In this model, noisy and imprecise data is used effectively to develop an AI algorithm that maps visual information about products to their UPC barcodes. As customers scan their items, the AI model compares visual images of the items to the scanned UPC barcodes. If there is a mismatch or if a mis-scan occurs, the AI model raises an alert by using a configured mechanism.