Electronic Theses and Dissertation

Jumlah penderita penyakit obesitas terus meningkat setiap tahunnya. Sehingga diperlukan suatu sistem diagnosa dini dari penyakit obesitas sehingga masyarakat mampu menjaga pola makan dan pola hidup sehat sejak dini. Namun metode diagnosis obesitas yang sudah ada sulit dijangkau dan digunakan oleh masyarakat. Tujuan penelitian ini adalah menghasilkan model sistem diagnosa obesitas berbasiskan thermal imaging dan deep learning yang mampu melakukan klasifikasi antara citra obesitas dan citra normal melalui aplikasi smart phone. Pada penelitian ini dibangun dataset citra termal peserta obesitas dan peserta normal menggunakan kamera termal. Dataset akan menjadi input dalam proses training model pre-trained CNN dan model custom CNN yang telah dilakukan fine tuning. Dalam proses desain model pre-trained CNN, digunakan arsitektur dan weight dari model pre-trained CNN yang sudah tersedia yaitu VGG19, MobileNet, Resnet152V2 dan DenseNet201. Sedangkan pada desain model custom CNN dibangun arsitektur model CNN C1, C2, dan C3. Hasil analisa dataset yang dibangun menunjukkan perbedaan suhu ROI antara citra normal dan obesitas berada pada bagian Abdomen sebesar 4.68 %. Hasil training model pre-trained CNN menunjukkan model DenseNet201 memiliki akurasi tertinggi sebesar 87 % dan model MobileNet memiliki time elapsed tercepat yaitu 13 detik. Sedangkan hasil training model custom menunjukkan model C2 memiliki akurasi tertinggi sebesar 76 % dan time elapsed sebesar 19 detik. Model C2 telah diimplementasikan ke dalam aplikasi smartphone sederhana berbasiskan android. Aplikasi akan menampilkan propabilitas diagnosa antara citra normal dan obesitas sehingga dapat digunakan masyarakat dalam melakukan deteksi dini penyakit obesitas.

The number of people with obesity continues to increase every year. So we need an early diagnosis system of obesity to maintain a healthy diet and lifestyle from an early age. However, the existing obesity diagnosis methods are difficult to reach and use. The purpose of this study is to produce a model of an obesity diagnosis system based on thermal imaging and deep learning that can classify obesity images and normal images through a smartphone application. In this study, a thermal image dataset of obese participants and normal participants was built using a thermal camera. The dataset will be the input in the training process for the pre-trained CNN model and the fine-tuned CNN custom model. In the design process of the pre-trained CNN model, the architecture and weights of the available pre-trained CNN models were used, namely VGG19, MobileNet, Resnet152V2, and DenseNet201. While on the CNN custom model design, the CNN architecture model C1, C2, and C3 is built. The results of the analysis of the dataset that was built showed the difference in ROI temperature between normal and obese images was in the abdomen of 4.68%. The results of the CNN pre-trained model training show that the DenseNet201 model has the highest accuracy of 87%, and the MobileNet model has the fastest elapsed time of 13 seconds. While the custom model training results show that the C2 model has the highest accuracy of 76% and time elapsed of 19 seconds. The C2 model has been implemented into a simple android-based smartphone application. The application will display the diagnostic probability between normal and obese images to be used by the public in conducting early detection of obesity.