Deteksi Anomali pada Bahan Makanan menggunakan Citra Hiperspektral dengan Diffusion Model

Amanullah, Arkan Arsalan (2026) Deteksi Anomali pada Bahan Makanan menggunakan Citra Hiperspektral dengan Diffusion Model. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025221129-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (29MB) | Request a copy

Abstract

Kasus keracunan makanan di Indonesia menunjukkan pentingnya sistem yang mampu mendeteksi kontaminasi bahan pangan secara akurat. Hyperspectral Imaging (HSI) digunakan karena mampu merekam informasi spektral dan spasial secara bersamaan. Namun, kompleksitas data hiperspektral serta keterbatasan data berlabel menyebabkan pendekatan klasifikasi konvensional kurang efektif. Oleh karena itu, penelitian ini mengusulkan metode deteksi anomali berbasis diffusion model pada citra hiperspektral. Penelitian ini menggunakan dataset citra hiperspektral bahan pangan yang terdiri atas citra normal dan citra dengan berbagai benda asing sebagai anomali. Pra-pemrosesan meliputi pemotongan citra menggunakan sliding window, injeksi anomali, reduksi dimensi menggunakan Principal Component Analysis (PCA), serta normalisasi data. Diffusion model pada penelitian ini dibangun menggunakan Variational Autoencoder (VAE) untuk pemetaan ke ruang laten, U-Net yang diintegrasikan dengan Anomaly-Masked Network (AMN) untuk proses denoising guna merekonstruksi citra normal, serta ResNet50 untuk ekstraksi fitur dalam pembentukan peta skor anomali. Hasil eksperimen menunjukkan bahwa penggunaan bobot pretrained VAE dari Stable Diffusion 1.5 serta konfigurasi neighbor size 15 × 15 pada Neighbor Masked Self-Attention (NMSA) di AMN dan kombinasi layer ResNet50 [0, 1, 2] memberikan performa deteksi terbaik. Model yang diusulkan memperoleh nilai rata-rata AUROC piksel sebesar 0,9694 dan AUPR piksel sebesar 0,8153, serta mengungguli seluruh metode pembanding. Hasil tersebut menunjukkan bahwa diffusion model efektif diterapkan pada citra hiperspektral untuk mendeteksi anomali pada bahan pangan.
==================================================================================================================================
Food poisoning incidents in Indonesia highlight the need for an accurate system to detect contamination in food materials. Hyperspectral Imaging (HSI) is utilized because it simultaneously captures both spectral and spatial information. However, the high complexity of hyperspectral data and the limited availability of labeled samples reduce the effectiveness of conventional classification approaches. Therefore, this study proposes a diffusion model-based anomaly detection method for hyperspectral images. The study employs a hyperspectral food image dataset consisting of normal samples and samples containing various foreign objects as anomalies. The preprocessing stage includes image cropping using a sliding window, anomaly injection, dimensionality reduction using Principal Component Analysis (PCA), and data normalization. The proposed diffusion model is built using a Variational Autoencoder (VAE) to map images into the latent space, a U-Net integrated with an Anomaly-Masked Network (AMN) to perform denoising for normal image reconstruction, and ResNet50 for feature extraction in anomaly score map generation. Experimental results show that using the pretrained VAE weights from Stable Diffusion 1.5, together with a neighborhood size of 15 × 15 in the Neighbor Masked Self-Attention (NMSA) module of the AMN and the combination of ResNet50 layers [0, 1, 2], achieves the best detection performance. The proposed model attains an average pixel-level AUROC of 0.9694 and a pixel-level AUPR of 0.8153, outperforming all baseline methods. These results demonstrate that diffusion models can be effectively applied to hyperspectral images for anomaly detection in food materials.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Hyperspectral Imaging, Deteksi Anomali, Diffusion Model, Hyperspectral Imaging, Anomaly Detection, Diffusion Model
Subjects:	T Technology > TA Engineering (General). Civil engineering (General) > TA1637 Image processing--Digital techniques. Image analysis--Data processing.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Arkan Arsalan Amanullah
Date Deposited:	24 Jul 2026 02:25
Last Modified:	24 Jul 2026 02:25
URI:	http://repository.its.ac.id/id/eprint/136932

Actions (login required)

View Item