Shelf Auditing Based on Image Classification Using Semi-Supervised Deep Learning to Increase On-Shelf Availability in...

Area of Science:

• Computer Vision
• Machine Learning
• Retail Analytics

Background:

• High on-shelf availability (OSA) is crucial for grocery store profitability.
• Existing computer vision datasets lack product annotations, necessitating manual labeling for OSA monitoring.
• Current methods require substantial manual effort for product labeling in images.

Purpose of the Study:

• To introduce a novel method combining semi-supervised learning and on-shelf availability (SOSA) to address the product annotation challenge.
• To apply the You Only Look Once (YOLOv4) deep learning architecture for OSA monitoring for the first time.
• To demonstrate explainable artificial intelligence (XAI) for OSA monitoring through a new software application, SOSA XAI.

Main Methods:

• Proposed a new semi-supervised learning and on-shelf availability (SOSA) method.
• Utilized the YOLOv4 deep learning architecture for OSA monitoring.
• Developed the SOSA XAI software application integrating explainable artificial intelligence.

Main Results:

• The SOSA method demonstrated effectiveness across image datasets with varying labeled sample ratios (20% to 80%).
• Experimental results confirmed the proposed SOSA approach surpasses existing methods like RetinaNet and YOLOv3 in accuracy.
• The SOSA XAI application provides explainability for OSA monitoring.

Conclusions:

• The proposed SOSA method offers a more accurate and efficient solution for monitoring on-shelf availability.
• The integration of YOLOv4 and XAI in SOSA XAI represents a significant advancement in retail analytics.
• This approach reduces the need for extensive manual labeling, saving time and resources in grocery inventory management.