Modifikasi Pelatihan Model MetaFormer Menggunakan Balanced Softmax Cross Entropy Loss Untuk Klasifikasi Spesies Flora dan Fauna

Arif, Muhammad Riv'an (2023) Modifikasi Pelatihan Model MetaFormer Menggunakan Balanced Softmax Cross Entropy Loss Untuk Klasifikasi Spesies Flora dan Fauna. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 06111940000085-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only until 1 October 2025. Download (2MB) | Request a copy

Abstract

Proses identifikasi spesies secara visual di alam liar sangatlah menantang, karena banyaknya jenis spesies berbeda yang memiliki kemiripan secara visual (fine-grained). Selain itu, distribusi spesies flora dan fauna juga bersifat long-tailed, di mana jumlah populasi antar jenis spesies sangat bervariasi. Oleh karena itu, diperlukan pengembangan alat otomatis yang akurat untuk membantu para pakar dalam memantau keanekaragaman spesies flora dan fauna secara global. Pada Tugas Akhir ini telah dilakukan modifikasi pelatihan model MetaFormer menggunakan Balanced Softmax Cross Entropy Loss untuk klasifikasi spesies flora dan fauna. Model MetaFormer yang digunakan terdiri dari Convolutional Neural Network (CNN) dan Transformer Encoder. CNN berfungsi untuk mengekstraksi fitur pada data visual, sementara Transformer Encoder digunakan untuk menggabungkan informasi metadata dengan hasil ekstraksi fitur visual. Penambahan informasi metadata pada MetaFormer dapat membantu mengatasi permasalahan fine-grained. Untuk meningkatkan performa MetaFormer, digunakan Balanced Softmax Cross Entropy Loss pada fase pelatihan model, yang dapat membantu menyeimbangkan perbedaan distribusi label antara data latih dan data uji (permasalahan long-tailed). Modifikasi pelatihan model MetaFormer menggunakan Balanced Softmax Cross Entropy Loss (MetaFormerBSL) menghasilkan performa top-1 accuracy dan top-3 accuracy masing-masing sebesar 88.1% dan 95.1% pada dataset iNaturalist 2018, yang menunjukkan peningkatan sebesar 3.6% dan 0.9% dibanding MetaFormer (84.5% dan 94.2%). MetaFormerBSL juga lebih baik dalam mengklasifikasikan spesies dengan data latih sedikit (few-shot).
=================================================================================================================================
The process of visually identifying species in the wild is very challenging due to the large number of different species that have fine-grained visual similarities. Additionally, the distribution of flora and fauna species is long-tailed, meaning that the population sizes among different species vary significantly. Therefore, the development of accurate automated tools is needed to assist experts in monitoring the biodiversity of flora and fauna globally. In this study, a modification have been made to the training phase of MetaFormer using Balanced Softmax Cross Entropy Loss for classifying flora and fauna species. MetaFormer consists of Convolutional Neural Network (CNN) and Transformer Encoder. CNN is responsible for extracting features from visual data, while Transformer Encoder combines metadata information with the extracted visual features. The addition of metadata information to MetaFormer helps address the fine-grained issues. To improve the performance of MetaFormer, Balanced Softmax Cross Entropy Loss is applied during the model training phase, which helps balance the label distribution differences between the training and test data, addressing the long-tailed problem. Modification to the training phase of MetaFormer using Balanced Softmax Cross Entropy Loss (MetaFormerBSL) resulted in top-1 accuracy and top-3 accuracy of 88.1% and 95.1% respectively, on the iNaturalist 2018 dataset, which improves performance by 3.6% and 0.9% compared to MetaFormer (84.5% and 94.2%). MetaFormerBSL is also better at classifying species with limited training data (few-shot).

Item Type:	Thesis (Other)
Uncontrolled Keywords:	flora and fauna, classification, MetaFormer, Balanced Softmax Cross Entropy Loss, flora dan fauna, klasifikasi
Subjects:	Q Science > QA Mathematics > QA336 Artificial Intelligence Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science)
Divisions:	Faculty of Science and Data Analytics (SCIENTICS) > Mathematics > 44201-(S1) Undergraduate Thesis
Depositing User:	Muhammad Riv'an Arif
Date Deposited:	29 Nov 2023 01:11
Last Modified:	29 Nov 2023 01:11
URI:	http://repository.its.ac.id/id/eprint/103420

Actions (login required)

View Item