SETYAWAN, Budi and Prahasto, Toni and Santoso, Rukun (2025) IMPLEMENTASI RESIDUAL BLOCK PADA CNN DAN BIFPN UNTUK INSPEKSI CACAT PENGECORAN LOGAM. Masters thesis, UNIVERSITAS DIPONEGORO.

	Text cover1.pdf Download (227kB)
	Text Cover - 30000321410010.pdf Restricted to Repository staff only Download (2MB)
	Text Bab I -30000321410010.pdf Download (1MB)
	Text Bab II - 30000321410010.pdf Download (8MB)
	Text Bab III - 30000321410010.pdf Restricted to Repository staff only Download (6MB)
	Text Bab IV - 30000321410010.pdf Restricted to Repository staff only Download (20MB)
	Text Bab V - 30000321410010.pdf Restricted to Repository staff only Download (441kB)
	Text Daftar Pustaka - 30000321410010.pdf Download (768kB)

Abstract

Inspeksi kualitas produk logam secara manual memiliki keterbatasan akurasi dan konsistensi. CNN (Convolutional Neural Network) sebagai metode utama dalam inspeksi berbantuan komputer rentan terhadap vanishing gradient saat pelatihan jaringan. Penelitian ini bertujuan memodifikasi arsitektur CNN dengan residual block untuk klasifikasi, serta mengembangkan deteksi cacat logam berbasis YOLOv5 yang dimodifikasi dengan residual block dan BiFPN (Bidirectional Feature Pyramid Network). Selain itu, penelitian ini membandingkan performa model modifikasi dengan model konvensional dan model pretrained berdasarkan akurasi, presisi, recall, F1-score, serta reduksi ukuran model. Model klasifikasi CNN dengan residual block menghasilkan akurasi 99,77%, precision 100%, recall 99,61%, F1-score 99,80, dan ukuran file hanya 0,225 MB. Sebaliknya, model pretrained seperti EfficientNetV2 dan MobileNet memiliki ukuran jauh lebih besar (12–25 MB) dengan akurasi lebih rendah. Untuk deteksi objek, model terbaik (YOLOv5-BiFPN-residual dengan augmentasi pada tingkat bounding box) mencatat precision 64%, recall 58%, F1-score 59%, mAP@50 sebesar 0,49, dan mAP@50:95 sebesar 0,17, dengan ukuran file hanya 9,25 MB dibandingkan 88,9 MB pada model standar. Hasil ini menunjukkan bahwa integrasi residual block telah meningkatkan efektivitas pembelajaran CNN, sekaligus menghasilkan model yang lebih ringan secara komputasi. Pendekatan ini dapat menjadi alternatif bagi sektor manufaktur logam yang membutuhkan model inspeksi yang bersifat ringan (lightweight) dengan input berupa citra grayscale. Kata kunci : CNN, deteksi objek, inspeksi kualitas logam, klasifikasi citra, YOLOv5.

Manual inspection of metal product quality is limited in terms of accuracy and consistency. Convolutional Neural Networks (CNNs), as a primary method in computer-aided inspection, are prone to vanishing gradients during network training. This study aims to modify the CNN architecture by incorporating residual blocks for classification tasks and to develop a defect detection system based on YOLOv5, enhanced with residual blocks and a Bidirectional Feature Pyramid Network (BiFPN). In addition, the study compares the performance of the modified models with conventional and pretrained models in terms of accuracy, precision, recall, F1-score, and model size reduction. The CNN classification model with residual blocks achieved 99.77% accuracy, 100% precision, 99.61% recall, 99.80 F1-score, and a file size of only 0.225 MB. In contrast, pretrained models such as EfficientNetV2 and MobileNet exhibited significantly larger sizes (12–25 MB) with lower accuracy. For object detection, the best-performing model (YOLOv5-BiFPN with residual blocks and bounding box– level augmentation) achieved 64% precision, 58% recall, 59% F1-score, a mAP@50 of 0.49, and mAP@50:95 of 0.17, with a file size of only 9.25 MB compared to 88.9 MB in the standard model. These results demonstrate that integrating residual blocks improves CNN training effectiveness while producing computationally lighter models. This approach offers a promising alternative for the metal manufacturing sector, which requires accurate yet lightweight inspection models based on grayscale image inputs. Keywords : casting quality inspection, CNN, Image classification, object detection, YOLOv5

Actions (login required)

View Item