Mobile Visual Product Recognition for Offline Low-Power Checkout

The system segments the product recognition task into two distinct classification arms: a first classification arm for coarse product category identification and a second classification arm for fine-grained product identification. This segmentation allows the system to process images through different computational pathways depending on the required precision, reducing overall energy consumption while maintaining high productivity in both quick category-based checkout and detailed product recognition scenarios.

The system merges multiple classification functions into a single convolutional neural network architecture by appending both the first classification arm (for coarse classification) and the second classification arm (for fine classification) to the same feature extraction backbone. This merging approach allows the shared convolutional layers to learn general product features once, which are then reused by both classification arms, thereby improving measurement precision without proportionally increasing device complexity.

The system performs preliminary action by pre-training the convolutional neural network with comprehensive product image data and classification rules before deployment in offline environments. The first and second classification arms are pre-configured with sufficient product information and decision logic during an initial setup phase, enabling the system to operate autonomously without active data connections. This preliminary preparation minimizes information loss by embedding necessary product recognition knowledge locally in the mobile device.