Decoupled Edge Guidance Network for Automatic Checkout.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Automatic checkout (ACO) systems face challenges due to the domain gap between single-product training data and multi-product real-world checkout images.
• Existing methods struggle to generalize from synthesized images to realistic checkout scenarios, limiting ACO performance.

Purpose of the Study:

• To propose a novel Decoupled Edge Guidance Network (DEGNet) to address the domain gap in automatic checkout.
• To improve the accuracy of generating complete shopping lists from checkout images.

Main Methods:

• Implemented a supervised domain adaptation approach integrating synthesized and checkout images.
• Developed an edge embedding module to introduce edge information and a decoupled feature extractor for joint image and edge data utilization.
• Introduced a proposal divide-and-conquer strategy (PDS) for augmenting high-quality training samples.

Main Results:

• DEGNet achieved state-of-the-art performance on the retail product checkout (RPC) dataset.
• Achieved checkout accuracy (cAcc) of 93.47% with Faster R-CNN and 95.25% with Cascade R-CNN.

Conclusions:

• The proposed DEGNet effectively bridges the domain gap in automatic checkout systems.
• The integration of edge information and the PDS strategy significantly enhance ACO performance on real-world checkout images.