Perbandingan Kinerja Inception- Resnetv2, Xception, Inception-v3, dan Resnet50 pada Gambar Bentuk Wajah

Abstrak

Saat ini, klasifikasi bentuk wajah banyak diterapkan dalam berbagai bidang. Dalam bidang industri fashion dapat digunakan untuk pemilihan gaya rambut, pemilihan bingkai kacamata, tata rias, dan mode lainnya. Selain itu, dalam bidang medis bentuk wajah digunakan untuk bedah plastik. Identifikasi bentuk wajah adalah tugas yang menantang karena kompleksitas wajah, ukuran, pencahayaan, usia dan ekspresi. Banyak metode yang dikembangkan untuk memberikan hasil akurasi terbaik dalam klasifikasi bentuk wajah. Deep learning menjadi tren dibidang komputer vision karena memberikan hasil yang paling baik dari pada metode sebelumnya. Makalah ini mencoba menyajikan perbandingan kinerja klasifikasi wajah dengan empat arsitektur deep learning Xception, ResNet50, InceptionResNet-v2, Inception-v3. Dataset yang digunakan berjumlah 4500 gambar yang terbagi lima kelas heart, long, oblong, square, round. Berbagai pengoptimal deep learning diantaranya; transfer learning, optimizer deep learning, dropout dan fungsi aktivasi diterapkan untuk meningkatkan kinerja model. Perbandingan antara berbagai model CNN didasarkan kinerja metrik seperti accuracy, recall, precision dan F1-score. Dengan demikian dapat disimpulkan bahwa model Inception-ResNet-V2 menggunakan fungsi aktivasi Mish dan optimizer Nadam mencapai nilai tertinggi dengan accuracy dan f1-score masing-masing 92.00%, dan penggunaan waktu 65.0 menit.

Abstract

Currently, face shape classification is widely applied in various fields. In the fashion industry, it can be used for hairstyle selection, eyeglass frame selection, makeup, and other modes. In the medical field, the face shape is used for plastic surgery. Identification of face shape is a challenging task due to the complexity of the face, size, lighting, age and expression. Many methods have been developed to provide the best accuracy results in the classification of face shapes. Deep learning is becoming a trend in the field of computer vision because it gives the best results than the previous method. This paper attempts to present a comparison of the performance of face classification with four deep learning architectures Xception, ResNet50, InceptionResNet-v2, Inception-v3. The dataset used is 4500 images divided into five classes heart, long, oblong, square, round. Various deep learning optimizers include; transfer learning, deep learning optimizer, dropout and activation functions are implemented to improve model performance. Comparisons between various CNN models are based on performance metrics such as accuracy, recall, precision and F1-score. Thus, it can be concluded that the Inception-ResNet-V2 model using the Mish activation function and the Nadam optimizer achieves the highest value with an accuracy and f1-score of 92.00%, and a time usage of 65.0 minutes. Thus, it can be concluded that the Inception-ResNet-V2 model using the Mish activation function and the Nadam optimizer achieves the highest value with an accuracy and f1-score of 92.00%, and a time usage of 65.0 minutes.