HELMUD, Ellya and Widodo, Catur Edi and Nurhayati, Oky Dwi (2026) KLASIFIKASI CITRA IKAN MENGGUNAKAN ENSEMBLE DEEP LEARNING BERBASIS FEATURE-LEVEL FUSION DENGAN OPTIMASI DUAL-BACKBONE ADAPTIF SECARA OTOMATIS. Doctoral thesis, UNIVERSITAS DIPONEGORO.

	Text Halaman Judul.pdf Download (142kB)
	Text Cover lengkap.pdf Restricted to Repository staff only Download (726kB)
	Text BAB I.pdf Download (1MB)
	Text BAB II.pdf Download (1MB)
	Text BAB III.pdf Restricted to Repository staff only Download (602kB)
	Text BAB IV.pdf Restricted to Repository staff only Download (1MB)
	Text BAB V.pdf Restricted to Repository staff only Download (964kB)
	Text Daftar Pustaka.pdf Download (270kB)
	Text Lampiran.pdf Restricted to Repository staff only Download (865kB)

Abstract

Klasifikasi citra ikan merupakan instrumen krusial dalam menjamin kontrol kualitas dan efisiensi rantai pasok industri perikanan. Penelitian ini mengusulkan integrasi segmentasi U2-Net dan arsitektur ensemble deep learning berbasis feature-level fusion untuk meningkatkan akurasi klasifikasi spesies dan tingkat kesegaran ikan. Dataset penelitian menggunakan 1.050 citra yang bersumber dari tempat pelelangan ikan dan laboratorium mutu perikanan Kota Pangkalpinang, mencakup enam kategori target serta satu kelas dummy (Tidak_Dikenali) sebagai penampung data citra selain dari 6 kategori tersebut. Prosedur eksperimen menerapkan pembagian data dengan proporsi 80% training, 10% validasi, dan 10% testing. Metodologi penelitian menerapkan optimasi dual-backbone adaptif dengan menggabungkan EfficientNetB0 dan InceptionV3 melalui mekanisme concatenation. Pelatihan model dilakukan dalam dua fase sistematis. Fase pertama berfokus pada pelatihan head klasifikasi baru dengan membekukan seluruh lapisan backbone (pre-trained). Fase kedua menerapkan Asymmetric Fine-Tuning dengan mencairkan 70 lapisan terakhir dari kedua model dasar dan menggunakan learning rate yang lebih kecil guna menjaga stabilitas bobot pengetahuan sebelumnya. Hasil pengujian menunjukkan performa model yang signifikan dengan perolehan akurasi validasi sebesar 0,9714 (loss 0,4045) dan akurasi test set sebesar 0,9810 (loss 0,3658). Metodologi ini terbukti menghasilkan keputusan klasifikasi yang kokoh terhadap variasi citra dan berkontribusi secara signifikan pada standardisasi manajemen informasi perikanan secara otomatis.
Kata Kunci : Ensemble Deep Learning, Model Fusi, Feature-Level Fusion, EfficientNetB0, InceptionV3, Klasifikasi citra, Segmentasi citra, optimasi DualBackbone, Fine-Tuning

Fish image classification serves as a crucial instrument in ensuring quality control and supply chain efficiency within the fisheries industry. This study proposes the integration of U2-Net segmentation and a deep learning ensemble architecture based on feature-level fusion to enhance the classification accuracy of fish species and freshness levels. The dataset employed consists of 1,050 images sourced from fish auction sites and fisheries quality laboratories in Pangkalpinang City, encompassing six target categories and one dummy class (Unrecognized) to accommodate out-of-distribution image data. The experimental procedure implemented a data partition of 80% for training, 10% for validation, and 10% for testing.The methodology applies adaptive dual-backbone optimization by merging EfficientNetB0 and InceptionV3 through a concatenation mechanism. Model training was executed in two systematic phases; the first phase focused on training the new classification head while freezing all pre-trained backbone layers, followed by a second phase applying Asymmetric Fine-Tuning. This second phase involved unfreezing the final 70 layers of both base models and utilizing a reduced learning rate to maintain the stability of prior knowledge weights. The evaluation results demonstrate significant model performance, achieving a validation accuracy of 0.9714 (loss: 0.4045) and a test set accuracy of 0.9810 (loss: 0.3658). This methodology proves to yield robust classification decisions against image variations and contributes significantly to the standardization of automated fisheries information management.
Keywords : Ensemble Deep Learning, Model Fusion, Feature-Level Fusion, EfficientNetB0, InceptionV3, Image Classification, Image Segmentation, DualBackbone Optimization, Fine-Tuning

Actions (login required)

View Item