Komparasi classifier head swin transformer dan multi-layer perceptron pada klasifikasi kesegaran ikan berbasis citra multi-perspektif

Rus'an, Jasmein Al-baar Putri (2026) Komparasi classifier head swin transformer dan multi-layer perceptron pada klasifikasi kesegaran ikan berbasis citra multi-perspektif. Sarjana thesis, UIN Sunan Gunung Djati Bandung.

Preview	Text (COVER) 1_cover.pdf Download (176kB) | Preview
Preview	Text (ABSTRAK) 2_abstrak .pdf Download (193kB) | Preview
Preview	Text (SK BEBAS PLAGIARISM) 3_skbebasplagiarism.pdf Download (277kB) | Preview
Preview	Text (DAFTAR ISI) 4_daftarisi.pdf Download (191kB) | Preview
Preview	Text (BAB I) 5_bab1.pdf Download (218kB) | Preview
	Text (BAB II) 6_bab2.pdf Restricted to Registered users only Download (525kB) | Request a copy
	Text (BAB III) 7_bab3.pdf Restricted to Registered users only Download (622kB) | Request a copy
	Text (BAB IV) 8_bab4.pdf Restricted to Registered users only Download (990kB) | Request a copy
	Text (BAB V) 9_bab5.pdf Restricted to Registered users only Download (172kB) | Request a copy
	Text (DAFTAR PUSTAKA) 10_daftarpustaka.pdf Restricted to Registered users only Download (181kB) | Request a copy
	Text (LAMPIRAN) 11_lampiran.pdf Restricted to Repository staff only Download (128kB) | Request a copy

Abstract

Indonesia: Perkembangan teknologi deep learning telah membuka peluang untuk mengotomatisasi penilaian kesegaran ikan yang selama ini dilakukan secara manual melalui inspeksi visual dan bersifat subjektif. Penelitian ini melakukan studi komparasi antara dua konfigurasi classifier head pada arsitektur Swin Transformer, yaitu Swin Transformer Baseline dengan linear classifier dan Swin Transformer dengan Multi-Layer Perceptron (MLP), pada domain klasifikasi kesegaran ikan berbasis citra multi-perspektif. Metodologi penelitian mengikuti kerangka CRISP-DM dengan strategi fine-tuning dua tahap pada backbone Swin-Tiny dan evaluasi menggunakan 5-Fold Cross Validation. Dataset yang digunakan merupakan gabungan tiga sumber publik yang mencakup 6.930 citra dari 12 spesies ikan dengan dua perspektif utama yaitu close-up mata dan tubuh utuh, berlabel highly fresh, fresh, dan not fresh. Hasil evaluasi menunjukkan bahwa kedua model menghasilkan performa yang setara dengan accuracy sebesar 85,10% dan ROC-AUC di atas 0,96. Model MLP menunjukkan stabilitas pelatihan yang lebih konsisten namun tanpa peningkatan kemampuan generalisasi yang signifikan, mengindikasikan bahwa representasi fitur backbone Swin Transformer yang telah di-fine-tune sudah memadai tanpa memerlukan kompleksitas tambahan pada classifier head. Visualisasi Grad-CAM mengkonfirmasi bahwa model berfokus pada area mata dan permukaan tubuh ikan sesuai indikator organoleptik dalam SNI 2729:2021. English: The advancement of deep learning technology has opened opportunities to automate fish freshness assessment, which has traditionally relied on subjective manual visual inspection. This study conducts a comparative analysis of two classifier head configurations on the Swin Transformer architecture, namely the Swin Transformer Baseline with a linear classifier and the Swin Transformer with a Multi-Layer Perceptron (MLP), applied to multi-perspective image-based fish freshness classification. The research methodology follows the CRISP-DM framework with a two-stage fine-tuning strategy on the Swin-Tiny backbone and 5-Fold Cross Validation for evaluation. The dataset combines three public sources comprising 6,930 images from 12 fish species with two main perspectives — close-up eye and whole body images — labeled as highly fresh, fresh, and not fresh. Evaluation results indicate that both models achieved comparable performance with an accuracy of 85.10% and ROC-AUC exceeding 0.96. The MLP model demonstrated more consistent training stability but without significant improvement in generalization capability, suggesting that the feature representations produced by the fine-tuned Swin Transformer backbone are sufficient without requiring additional complexity in the classifier head. Grad-CAM visualization confirms that the model focuses on the eye region and body surface consistent with the organoleptic indicators in SNI 2729:2021.

Actions (login required)

View Item