Uporaba računalniškega vida za napovedovanje zapečenosti piškotov

SEDEJ, NINA (author), Perš, Janez (mentor), Koporec, Gregor (comentor)

Peka jedi v pečici je priljubljen način priprave hrane, pri katerem so rezultati odvisni od številnih dejavnikov, ki jih je težko natančno nadzorovati. V okviru magistrske naloge predstavimo sistem računalniškega vida za nedestruktivno določanje stanja pečenja piškotov v pečici --- ali je piškot surov, nizko, srednje ali visoko zapečen. Cilj sistema je pomoč ljubiteljskemu kuharju, ki uporablja gospodinjsko pečico, pri doseganju optimalnih rezultatov peke piškotov. Za simulacijo odločitev ljubiteljskega kuharja smo uporabili model, ki se odloča glede na čas, pretečen od začetka peke. Ta nam je služil kot osnova za vrednotenje sistema računalniškega vida, ki je sestavljen iz pečice z vgrajeno digitalno kamero, ki zajema slike piškotov med peko ter modela za določanje stanja pečenja na podlagi zajetih slik. Preizkusili smo modele, ki se odločajo na podlagi ene same slike in modele, ki se odločajo na podlagi zaporedja slik. Ugotovili smo, da modeliranje s pomočjo ene same slike ne izboljša rezultatov ljubiteljskega kuharja. Po drugi strani smo z modeli ConvLSTM, ki se odločajo na podlagi zaporedja slik, dosegli izboljšanje v primerjavi z ljubiteljskim kuharjem. Iz tega sklepamo, da je za uspešno določanje stanja pečenja potrebna informacija o dinamiki pečenja. Kljub temu, da bi bile za splošno uporabo potrebne še dodatne izboljšave zaključimo, da je sistem računalniškega vida z modeli, ki modelirajo dinamiko peke, obetavna rešitev za pomoč ljubiteljskemu kuharju pri izboljšanju rezultatov pečenja piškotov.

• Language: Slovenian
• Keywords: stanje pečenja, dinamika pečenja, računalniški vid, globoke nevronske mreže, ConvLSTM, CNN, CNN-LSTM
• Work type: Master's thesis/paper
• Organization: FE - Faculty of Electrical Engineering
• Year: 2023
• PID: 20.500.12556/RUL-147856
• COBISS.SI-ID: 158978563
• Publication date in RUL: 14.07.2023

Secondary language

• Language: English
• Title: Utilizing Computer Vision for Predicting Cookie Doneness

Abstract

Baking in an oven is a popular food preparation method, where the results depend on numerous variables that are difficult to control precisely. In this master's thesis, we present a computer vision system for non-destructive determination of the baking state of cookies in an oven - whether the cookies are raw or their doneness is low, meduim, or high. The goal of the system is to assist the user of a household oven, particularly a home cook, in achieving optimal baking results. To simulate the decision-making process of a home cook, we used a model that makes decisions based on the time elapsed since the start of baking. This model served as a baseline for evaluating the computer vision system, which consists of an oven equipped with a built-in digital camera capturing images of the cookies during baking, and a model for determining the baking state based on the captured images. We tested models that make decisions based on a single image and models that make decisions based on a sequence of images. We found that modeling based on a single image did not improve the results compared to a home cook. On the other hand, with ConvLSTM models that make decisions based on a sequence of images, we achieved improvement compared to the amateur cook. This indicates that information about the dynamics of baking is crucial for successful determination of the baking state. Although further improvements would be needed for general use, we conclude that a computer vision system with models that capture the dynamics of baking shows promise in assisting home cooks in improving cookie baking results.

• Keywords: baking state, baking dynamics, computer vision, deep neural networks, ConvLSTM, CNN, CNN-LSTM