AI-Powered Image Recognition for Fresh Produce Retail: System Design and Implementation

Background: The fresh fruit and vegetable sector in retail relies heavily on weighing scales, traditionally using barcode scales with PLU codes. With hundreds of SKUs, PLU lookup is time‑consuming, leading to long queues and higher labor costs.

To address this, a machine‑learning based image‑recognition solution was proposed. Cameras attached to the scales capture items, and a model identifies the product, presenting a list of possible matches, thus reducing cashier interactions and training requirements.

Framework selection: TensorFlow was chosen as the core ML framework because its Lite version supports mobile deployment and aligns with existing internal expertise.

Product association: Instead of relying on barcodes, the system leverages the hierarchical product category (up to four levels) and specific fruit varieties. Example category path:

食品酒水 > 水产肉类/新鲜蔬果/熟食/现做食品 > 新鲜水果 > 苹果

Specific apple varieties are represented as:

金帅

国光

冰糖心

Feedback loop: User selections are cached offline and uploaded during idle periods to an ODS database. The data enriches the model for each store, improving accuracy over time.

Process optimization includes automated category inference from product titles and images, image preprocessing (cropping, sharpening, distortion correction), and handling of empty‑tray images to filter out irrelevant data.

Offline capability: An on‑device index stores recognized items, allowing immediate response even with poor network connectivity. The model is updated incrementally as more data is collected.

Conclusion: The deployed image‑recognition system significantly improves checkout efficiency and reduces labor costs for fresh produce retailers, while providing a reusable framework for other industry segments.