Electronic Theses and Dissertation

Penelitian ini mengevaluasi performa metode deep learning untuk mengklasifikasi citra emping melinjo berdasarkan kualitasnya menggunakan perangkat lunak mvtec halcon. emping melinjo merupakan produk ekspor indonesia yang kualitasnya memengaruhi harga dan permintaan pasar. klasifikasi awalnya dilakukan dengan proses penglihatan secara visual yang dilakukan oleh pakar (manusia) akan tetapi permasalahan muncul dikarenakan proses pengambilan keputusan yang subjektif, lambat, dan rentan kesalahan, sehingga peneliti mengusulkan otomatisasi dengan deep learning. tujuan penelitian adalah menguji akurasi metode deep learning dalam mengenali fitur visual emping seperti tekstur, warna, dan ketebalan. penelitian ini diharapkan meningkatkan efisiensi industri makanan dengan mengurangi ketergantungan pada inspeksi manual. metode yang digunakan meliputi pengambilan 900 citra emping melinjo dari tiga kelas kualitas (super, lose, barang/rusak) menggunakan kamera charge couple device (ccd) dan alas belt dari conveyor. dataset dibagi menjadi 70% pelatihan, 15% validasi, dan 15% pengujian. model deep learning yang diuji meliputi resnet18, alexnet, serta model compact dan enhanced dari mvtec halcon. proses transfer learning diterapkan dengan memanfaatkan pre-trained model seperti alexnet yang telah dilatih pada dataset besar seperti imagenet. tahap preprocessing meliputi normalisasi ukuran citra (224x224 piksel) dan konversi nilai piksel ke rentang -127 hingga 128 untuk optimasi komputasi. hasil pelatihan menunjukkan performa dari beberapa model: resnet18 mencapai akurasi sebesar 97% dengan loss 0.283, diikuti oleh model enhanced (84.8%), compact (85%), dan alexnet (74.6%). arsitektur resnet18 menggunakan residual block yang dirancang untuk memudahkan pelatihan jaringan dalam (18 layer) tanpa vanishing gradient dan mampu mengekstraksi fitur lokal seperti tepi dan tekstur. model ini memerlukan dataset yang besar dan waktu pelatihan selama 25 menit 51 detik. di sisi lain, performa model alexnet dipengaruhi oleh ukuran dataset yang kecil dan ketiadaan gpu, dengan waktu pelatihan 8 menit 8 detik. adapun model compact dan enhanced dari mvtec halcon menunjukkan akurasi sekitar 85% dan penggunaan memori yang hemat, sehingga cocok untuk aplikasi real-time. waktu pelatihan untuk model compact adalah 3 menit 25 detik. secara umum, waktu pelatihan dipengaruhi oleh faktor seperti jumlah epoch dan masalah overfitting. hasil ini sejalan dengan hipotesis bahwa cnn meningkatkan akurasi klasifikasi melalui ekstraksi fitur hierarkis.

This study evaluates the performance of deep learning methods for classifying melinjo crackers (emping melinjo) images based on their quality using MVTEC Halcon software. Melinjo crackers are an Indonesian export product whose quality affects price and market demand. Classification was initially performed through visual inspection by human experts, but problems arose due to subjective, slow, and error-prone decision-making processes, leading researchers to propose automation with Deep Learning. The research aims to test the accuracy of the Deep Learning method in recognizing visual features of the crackers, such as texture, color, and thickness. This research is expected to increase the efficiency of the food industry by reducing reliance on manual inspection. The methodology involved capturing 900 images of melinjo crackers from three quality classes (super, lose, and goods/damaged) using a Charge-Coupled Device (CCD) camera and a conveyor belt. The dataset was divided into 70% for training, 15% for validation, and 15% for testing. The Deep Learning models tested include ResNet18, AlexNet, and the compact and enhanced models from MVTEC Halcon. A transfer learning approach was applied by utilizing pre-trained models like AlexNet, which has been trained on large datasets such as ImageNet. The preprocessing stage included normalizing the image size (224x224 pixels) and converting pixel values to a range of -127 to 128 for computational optimization. The training results showed the performance of several models: ResNet18 achieved an accuracy of 97% with a loss of 0.283, followed by the enhanced model (84.8%), the compact model (85%), and AlexNet (74.6%). The ResNet18 architecture uses residual blocks designed to facilitate the training of deep networks (18 layers) without the vanishing gradient problem and is capable of extracting local features like edges and textures. This model required a large dataset and a training time of 25 minutes and 51 seconds. On the other hand, the performance of the AlexNet model was affected by the small dataset size and the absence of a GPU, with a training time of 8 minutes and 8 seconds. Meanwhile, the compact and enhanced models from MVTEC Halcon showed accuracies of around 85% and were memory-efficient, making them suitable for real-time applications. The training time for the compact model was 3 minutes and 25 seconds. In general, training time was influenced by factors such as the number of epochs and overfitting issues. These results align with the hypothesis that CNNs improve classification accuracy through hierarchical feature extraction.