Penerapan Adaptive Smoothing Regularization untuk Otomatisasi Penilaian Estetika pada Foto Makanan

Yonia, Dwika Lovitasari (2025) Penerapan Adaptive Smoothing Regularization untuk Otomatisasi Penilaian Estetika pada Foto Makanan. Masters thesis, Institut Teknologi Sepuluh Nopember.

[thumbnail of 6025231051-Master_Thesis.pdf] Text
6025231051-Master_Thesis.pdf - Accepted Version
Restricted to Repository staff only until 1 April 2027.

Download (6MB) | Request a copy

Abstract

Di era digital, estetika visual makanan melalui fotografi memainkan peran strategis dalam menarik konsumen, baik bagi individu maupun industri kuliner. Teknologi deep learning telah mencatat kemajuan dalam mengembangkan algoritma untuk meniru penilaian manusia terhadap estetika makanan. Namun, penelitian sebelumnya masih menghadapi beberapa keterbatasan, seperti overfitting pada data pelatihan, minimnya eksplorasi terhadap fitur visual seperti pencahayaan dan tekstur, serta metode regulasi yang kurang adaptif terhadap data baru. Selain itu, subjektivitas dalam mendefinisikan dan mengukur estetika makanan sering kali menjadi tantangan. Untuk mengatasi masalah ini, penelitian ini mengintegrasikan Adaptive Smoothing Regularization (ASR), sebuah inovasi dalam regulasi adaptif, dengan fitur visual utama seperti komposisi, warna dominan, dan harmoni warna. Penelitian ini menggunakan dataset publik yang terdiri dari 24.000 sampel, dengan 13.088 foto makanan yang dianggap estetik secara visual dan 10.912 foto yang tidak estetik. Fitur komposisi digunakan untuk menangkap pola tata letak seperti simetri dan keseimbangan, sedangkan warna dominan berkontribusi terhadap persepsi visual. Harmoni warna, meskipun menilai kesesuaian warna, tidak memberikan dampak signifikan terhadap kinerja model dan meningkatkan kompleksitas sistem. Hasil eksperimen menunjukkan bahwa kombinasi komposisi dan warna dominan memberikan kinerja terbaik, dengan model InceptionV3 yang dilengkapi ASR mencapai akurasi 75,9% dan F1-Score 77,3%, mencerminkan keseimbangan optimal antara presisi (75,6%) dan recall (78,9%). Sementara itu, model Random Forest menunjukkan presisi tertinggi sebesar 80,1%, menjadikannya model paling konsisten. Meskipun ASR terbukti mengurangi overfitting, dampaknya lebih terbatas pada dataset yang bersih seperti yang digunakan dalam penelitian ini. Penelitian ini tidak hanya menawarkan pendekatan teknis yang lebih baik dalam otomatisasi penilaian estetika makanan, tetapi juga memberikan wawasan baru tentang elemen visual signifikan dalam estetika foto makanan. Kontribusi ini memberikan nilai praktis bagi industri kuliner, fotografi, dan desain, serta memperluas pemahaman dalam pengembangan metode deep learning berbasis regulasi adaptif.
==============================================================================================================================
In the digital era, the visual aesthetics of food through photography play a strategic role in attracting consumers, both for individuals and the culinary industry. Deep learning technology has made significant strides in developing algorithms to mimic human judgment of food aesthetics. However, previous studies have faced several limitations, such as overfitting on training data, minimal exploration of visual features like lighting and texture, and regulatory methods that are less adaptive to new data. Additionally, the subjectivity in defining and measuring food aesthetics often poses challenges. To address these issues, this study integrates Adaptive Smoothing Regularization (ASR), an innovation in adaptive regulation, with key visual features such as composition, dominant color, and color harmony. The study uses a public dataset comprising 24,000 samples, including 13,088 visually aesthetic food photos and 10,912 non-aesthetic ones. Composition features are used to capture layout patterns like symmetry and balance, while dominant color contributes to visual perception. Although color harmony assesses color compatibility, it does not significantly impact model performance and adds system complexity. Experimental results show that the combination of composition and dominant color provides the best performance, with the InceptionV3 model equipped with ASR achieving 75.9% accuracy and an F1-Score of 77.3%, reflecting an optimal balance between precision (75.6%) and recall (78.9%). Meanwhile, the Random Forest model exhibited the highest precision at 80.1%, making it the most consistent model. Although ASR effectively reduces overfitting, its impact is more limited on clean datasets like the one used in this study. This study not only offers a technically improved approach to automating food aesthetic assessment but also provides new insights into significant visual elements in food photo aesthetics. These contributions offer practical value for the culinary, photography, and design industries, while also broadening the understanding of developing adaptive regulation-based deep learning methods.

Item Type: Thesis (Masters)
Uncontrolled Keywords: Adaptive Smoothing Regularization; Convolutional Neural Networks, ekstraksi fitur, estetika makanan, Industry, Innovation, and Infrastructure, feature extraction, food aesthetics, Industry, Innovation, and Infrastructure
Subjects: T Technology > T Technology (General)
Divisions: Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55101-(S2) Master Thesis
Depositing User: Dwika Lovitasari Yonia
Date Deposited: 02 Feb 2025 07:38
Last Modified: 02 Feb 2025 07:38
URI: http://repository.its.ac.id/id/eprint/117573

Actions (login required)

View Item View Item