Smart retail SKUs checkout using improved residual network.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Unmanned stores require efficient operational systems.
• Automated Stock Keeping Units (SKUs) recognition is crucial for expediting customer checkout.
• Existing generic detection algorithms struggle with challenges in retail environments such as dense item arrangements, scale variations, and product similarities.

Purpose of the Study:

• To develop an improved retail product detection algorithm for unmanned stores.
• To enhance the accuracy and efficiency of Stock Keeping Units (SKUs) recognition.
• To address the specific challenges of detecting densely arranged, varied scale, and similar-looking retail products.

Main Methods:

• Proposed a novel boundary regression neural network architecture for enhanced bounding box detection in dense arrangements, reducing computational cost and parameter size.
• Introduced a novel loss function for hierarchical detection to address imbalances in positive and negative samples.
• Enhanced non-maximum suppression (NMS) with weighted non-maximum suppression (WNMS), linking NMS ranking scores to candidate box accuracy.

Main Results:

• The proposed algorithm demonstrated improved reliability and efficiency compared to existing methods on the SKU-110K and RPC datasets.
• The novel boundary regression network effectively handled dense item arrangements.
• The hierarchical loss function and WNMS contributed to more accurate product detection.

Conclusions:

• The developed retail product detection algorithm significantly enhances SKUs recognition in unmanned stores.
• The proposed methods effectively mitigate challenges associated with dense item arrangements, scale variations, and product similarities.
• This advancement contributes to greater operational efficiency and improved customer experience in retail settings.