0415068801 - Yusup Miftahuddin, S.Kom., MT Dosen Pembimbing 1 Surya Reza Putra / 152020037 Penulis

Industri konveksi menghadapi tantangan dalam menjaga kualitas produk, khususnya pada deteksi kecacatan pakaian yang berpotensi menimbulkan kerugian. Inspeksi manual yang rentan kesalahan mendorong penerapan metode deep learning. Penelitian ini membandingkan dua arsitektur Convolutional Neural Network (CNN), yaitu ResNet-101 sebagai model utama dan LeNet-5 sebagai model pembanding, dengan optimasi menggunakan algoritma Adaptive Moment Estimation (Adam) serta hyperparameter tuning. Dataset citra pakaian diperoleh dari sumber publik dan data primer, diproses menjadi ukuran seragam 160×160 piksel, lalu dibagi untuk pelatihan, validasi, dan pengujian. Hasil menunjukkan bahwa LeNet-5 mencapai akurasi uji 88,89% dengan loss 0,2901, sedikit lebih tinggi dibanding ResNet-101 optimal dengan akurasi 85,42% dan loss 0,3538. Optimasi pada ResNet-101 berhasil meningkatkan akurasi dari 73,61% menjadi 85,42%. Penelitian ini menegaskan bahwa model sederhana seperti LeNet-5 dapat memberikan performa kompetitif sekaligus efisiensi sumber daya, sehingga lebih sesuai untuk implementasi industri. The garment industry faces challenges in maintaining product quality, particularly in detecting fabric defects that may lead to financial losses. Manual inspection methods are prone to errors, thus encouraging the adoption of deep learning solutions. This study compares two Convolutional Neural Network (CNN) architectures: ResNet-101 as the main model and LeNet-5 as the baseline. Optimization was performed using the Adaptive Moment Estimation (Adam) algorithm combined with hyperparameter tuning. The dataset was obtained from public sources and primary data, preprocessed into a uniform size of 160×160 pixels, and divided into training, validation, and testing sets. Experimental results show that LeNet-5 achieved a testing accuracy of 88.89% with a loss of 0.2901, slightly outperforming the optimized ResNet-101, which reached 85.42% accuracy with a loss of 0.3538. Moreover, optimization improved ResNet-101 accuracy from 73.61% to 85.42%. These findings demonstrate that simpler models such as LeNet-5 can deliver competitive performance with greater computational efficiency, making them suitable for industrial applications.