Detekcija proizvoda na policama trgovina primjenom dubokog učenja

master's thesis

• Title (english): Detection of products on store shelves using deep learning
• Author: Marko Tutić
• Mentor: Zoran Kalafatić (mentor)
• Committee member: Zoran Kalafatić (predsjednik povjerenstva)
• Committee member: Tomislav Hrkać (član povjerenstva)
• Committee member: Vladimir Čeperić (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2023-07-03, Croatia

Abstract

Detekcija objekata u trgovini je izazovan zadatak koji zahtijeva precizno otkrivanje velikog broja objekata unutar ograničenog vremena. Ovo istraživanje analiziralo je različite modele detekcije objekata pomoću skupa podataka SKU110K kako bi se odredile njihove mogućnosti detekcije objekata u trgovini. Analiza je uzela u obzir gustoću objekata i pojave preklapanja unutar skupa podataka.
Četiri modela, Faster R-CNN, SSD, RetinaNet i YOLOv8, vrednovana su na skupu podataka SKU110K. Procjena je uključivala različite razlučivosti slike i veličine objekata. YOLOv8, najnoviji model, nadmašio je ostale modele u kontekstu preciznosti.
Uvidi iz analize skupa podataka otkrili su da je nizak prag omjera presjeka i unije (IoU) za algoritam potiskivanja ne-maksimalnih odziva (NMS) doveo do smanjene preciznosti. Dok je RetinaNet imao dobre rezultate na slikama visoke rezolucije, SSD je pokazao smanjenu točnost na slikama visoke rezolucije.
Općenito, YOLOv8 je pokazao vrhunske performanse u otkrivanju objekata u trgovini, iako su svi modeli imali problema s otkrivanjem malih predmeta.

Abstract (english)

Store object detection is a challenging task that requires the accurate detection of numerous objects within a limited time. This research analyzed various object detection models using the SKU110K dataset to determine their capabilities for in-store object detection. The analysis considered object density and overlapping occurrences within the dataset.
Four models, Faster R-CNN, SSD, RetinaNet, and YOLOv8, were evaluated on the SKU110K dataset. The evaluation involved different image resolutions and object sizes. YOLOv8, the latest model, outperformed the other models in terms of accuracy.
Insights from the dataset analysis revealed that a low Intersection over Union (IoU) threshold for the Non-Maximum Suppression (NMS) algorithm led to decreased accuracy. While RetinaNet performed well with high-resolution images, SSD showed reduced accuracy with increased image resolutions.
Overall, YOLOv8 demonstrated superior performance in store object detection, although all models struggled with detecting small objects.

• Keywords: Detekcija proizvoda; YOLOv8; RetinaNet; SSD; Faster R-CNN; Duboko učenje; Računalni vid
• Keywords (english): Store objects detection; YOLOv8; RetinaNet; SSD; Faster R-CNN; Deep Learning; Computer Vision
• Language: english
• URN:NBN: urn:nbn:hr:168:116569
• Study programme: Title: Computing; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra inženjer/inženjerka računarstva (magistar/magistra inženjer/inženjerka računarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2023-10-11 10:50:11