Deteksi Cacat Penanda Skala Botol Susu Dengan Convolutional Neutral Network

Lestari, Farrah Choirinisa and Prof. Dr. Eng. Agus Naba, S.Si., MT and Drs. Hari Arief Dharmawan, M.Eng., Ph.D. (2024) Deteksi Cacat Penanda Skala Botol Susu Dengan Convolutional Neutral Network. Sarjana thesis, Universitas Brawijaya.

Abstract

Dalam industri botol cairan, kualitas kemasan merupakan indikator penting untuk mempertahankan daya saing produk. Proses kontrol kualitas secara manual sering kali tidak efektif, terutama pada skala produksi besar. Dengan kemajuan teknologi seperti machine learning dan computer vision, proses deteksi cacat dapat dilakukan secara otomatis menggunakan metode Automated Vision Inspection (AVI). Penelitian ini bertujuan untuk mendeteksi cacat penanda skala botol susu menggunakan metode Convolutional Neural Network (CNN). Tiga arsitektur CNN, yaitu VGG16, EfficientNetV2B0, dan Xception, diterapkan dengan data set seimbang. Model dirancang menggunakan layer konvolusi dengan 32 filter, max pooling, dropout, flatten, dan layer dense dengan fungsi aktivasi softmax, serta dioptimalkan menggunakan optimizer Adam dan loss function categorical crossentropy. Hasil pelatihan menunjukkan bahwa VGG16 memiliki performa terbaik dengan akurasi 100%, dibandingkan dengan EfficientNetV2B0 (54,84%) dan Xception (64,52%). Saat diuji dengan data baru, VGG16 mencapai akurasi 98,5%, Xception 97,2%, dan EfficientNetV2B0 96,5%. Hasil ini menegaskan bahwa VGG16 adalah model yang paling optimal dalam mendeteksi cacat penanda skala botol susu, baik pada data set seimbang maupun tidak seimbang.

English Abstract

In the liquid bottle industry, packaging quality is a crucial indicator for maintaining product competitiveness. Manual quality control processes are often ineffective, especially for large-scale production. With advances in technology such as machine learning and computer vision, defect detection can be automated using Automated Vision Inspection (AVI) methods. This study aims to detect scale marker defects on milk bottles using the Convolutional Neural Network (CNN) method. Three CNN architectures, namely VGG16, EfficientNetV2B0, and Xception, were implemented with balanced datasets. The models were designed using a convolutional layer with 32 filters, max pooling, dropout, flattening, and dense layers with softmax activation functions, optimized using the Adam optimizer and categorical crossentropy loss function. The training results showed that VGG16 achieved the best performance with 100% accuracy, compared to EfficientNetV2B0 (54.84%) and Xception (64.52%). When tested on new data, VGG16 achieved 98.5% accuracy, Xception 97.2%, and EfficientNetV2B0 96.5%. These findings confirm that VGG16 is the most optimal model for detecting scale marker defects on milk bottles, both for balanced and unbalanced datasets.

Text (DALAM MASA EMBARGO) FARRAH CHOIRINISA LESTARI.pdf Restricted to Registered users only Download (2MB)

Actions (login required)

View Item