Pramudya, Rafli Raihan (2026) Deteksi Anomali dalam Inspeksi Industri Agri-Food Menggunakan Algoritma PatchCore pada Data Hiperspektral. Other thesis, Institut Teknologi Sepuluh Nopember.
|
Text
5025221266-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (11MB) | Request a copy |
Abstract
Industri Agrifood menghadapi tantangan besar berupa food loss and waste yang sebagian besar disebabkan oleh cacat dan kontaminasi yang tidak terdeteksi pada lini produksi. Metode berbasis citra RGB memiliki keterbatasan dalam menangkap anomali karena minimnya informasi spektral. Hyperspectral Imaging (HSI) mampu mengatasi keterbatasan tersebut dengan ratusan pita spektral kontinu untuk setiap piksel. Penelitian ini mengembangkan dan mengevaluasi sistem deteksi anomali berbasis algoritma PatchCore yang diadaptasi untuk data hiperspektral pada produk Agrifood menggunakan dataset HSI-AgriFoodAnomaly. Sistem yang dirancang mencakup pipeline utuh berupa normalisasi min-max, reduksi dimensi spektral dengan PCA, ekstraksi fitur menggunakan 3D Convolutional Autoencoder, pembentukan memory bank melalui coreset subsampling, serta mekanisme inferensi yang menggabungkan skor PatchCore dengan skor rekonstruksi Autoencoder. Pengujian dilakukan dengan rancangan faktorial dua faktor, yaitu jumlah komponen PCA (3 dan 30) dan ukuran patch (3x3 dan 9x9 piksel), menghasilkan delapan konfigurasi yang dievaluasi pada 40 citra uji. Hasil penelitian menunjukkan bahwa konfigurasi terbaik, PCA30·9x9 dengan penambahan Autoencoder, mencapai Pixel-AUROC sebesar 0,883 dan Pixel-AUCPR sebesar 0,677. Metode yang diusulkan mengungguli tujuh pembanding (U-net AutoAD, BockNet, OTAD, SGLNet, DMS2F, dan PA2E) pada dataset yang sama. Variasi arsitektur dengan fungsi aktivasi sigmoid tanpa normalisasi setelah PCA mencatat performa tertinggi dengan Pixel-AUROC 0,941 dan Pixel-AUCPR 0,774, yang mengindikasikan bahwa kualitas representasi fitur pada memory bank lebih bermakna daripada kecilnya loss rekonstruksi semata.
====================================================================================================================================
The Agri-food industry faces significant challenges from food loss and waste, largely caused by undetected quality defects and contamination on production lines. Conventional RGB-based inspection methods are limited in capturing subtle anomalies due to insufficient spectral information. Hyperspectral Imaging (HSI) overcomes this limitation by recording hundreds of continuous spectral bands per pixel. This study develops and evaluates an anomaly detection system based on the PatchCore algorithm adapted for hyperspectral data in Agri-food products using the HSI-AgriFoodAnomaly dataset. The designed system comprises a complete pipeline consisting of min-max normalization, spectral dimensionality reduction via PCA, feature extraction using a 3D Convolutional Autoencoder, memory bank construction through coreset subsampling, and an inference mechanism that combines PatchCore scores with Autoencoder reconstruction scores. Experiments were conducted using a two-factor factorial design involving the number of PCA components (3 and 30) and patch size (3x3 and 9x9 pixels), yielding eight configurations evaluated on 40 test images. The results show that the best configuration, PCA30x9 with Autoencoder integration, achieved a Pixel-AUROC of 0.883 and a Pixel-AUCPR of 0.677. The proposed method outperformed six comparison methods (AutoAD, BockNet, OTAD, SGLNet, DMS2F, and PA2E) on the same dataset. An architectural variation employing a sigmoid activation function without post-PCA normalization recorded the highest performance with a Pixel-AUROC of 0.941 and a Pixel-AUCPR of 0.774, indicating that the quality of feature representation in the memory bank is determined more feature distribution than by low reconstruction loss alone.
| Item Type: | Thesis (Other) |
|---|---|
| Uncontrolled Keywords: | Deteksi Anomali, Hyperspectral Imaging, PatchCore, Autoencoder Anomaly Detection, Hyperspectral Imaging, PatchCore, Autoencoder. |
| Subjects: | Q Science > QA Mathematics > QA336 Artificial Intelligence T Technology > T Technology (General) > T57.5 Data Processing |
| Divisions: | Faculty of Information and Communication Technology > Informatics > 55201-(S1) Undergraduate Thesis |
| Depositing User: | Rafli Raihan Pramudya |
| Date Deposited: | 24 Jul 2026 03:56 |
| Last Modified: | 24 Jul 2026 03:56 |
| URI: | http://repository.its.ac.id/id/eprint/136747 |
Actions (login required)
![]() |
View Item |
