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


Creative Commons License

Yilmazer R., Birant D.

SENSORS, cilt.21, sa.2, 2021 (SCI-Expanded) identifier identifier identifier

  • Yayın Türü: Makale / Tam Makale
  • Cilt numarası: 21 Sayı: 2
  • Basım Tarihi: 2021
  • Doi Numarası: 10.3390/s21020327
  • Dergi Adı: SENSORS
  • Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Academic Search Premier, Aerospace Database, Aqualine, Aquatic Science & Fisheries Abstracts (ASFA), Biotechnology Research Abstracts, CAB Abstracts, Communication Abstracts, Compendex, EMBASE, INSPEC, MEDLINE, Metadex, Veterinary Science Database, Directory of Open Access Journals, Civil Engineering Abstracts
  • Anahtar Kelimeler: on-shelf availability, semi-supervised learning, deep learning, image classification, machine learning, explainable artificial intelligence
  • Dokuz Eylül Üniversitesi Adresli: Evet

Özet

Providing high on-shelf availability (OSA) is a key factor to increase profits in grocery stores. Recently, there has been growing interest in computer vision approaches to monitor OSA. However, the largest and well-known computer vision datasets do not provide annotation for store products, and therefore, a huge effort is needed to manually label products on images. To tackle the annotation problem, this paper proposes a new method that combines two concepts "semi-supervised learning" and "on-shelf availability" (SOSA) for the first time. Moreover, it is the first time that "You Only Look Once" (YOLOv4) deep learning architecture is used to monitor OSA. Furthermore, this paper provides the first demonstration of explainable artificial intelligence (XAI) on OSA. It presents a new software application, called SOSA XAI, with its capabilities and advantages. In the experimental studies, the effectiveness of the proposed SOSA method was verified on image datasets, with different ratios of labeled samples varying from 20% to 80%. The experimental results show that the proposed approach outperforms the existing approaches (RetinaNet and YOLOv3) in terms of accuracy.