Hyperparameter Tuning of Identity Block Uses an Imbalance Dataset with Hyperband Method
Abdul Rachman Manga, Muhammad Acqmal Fadhilla Latief, Andi Widya Mufila Gaffar, Huzain Azis, Ramdan Satra, Yulita Salim
Abstract
Visual pattern recognition, selection of appropriate image processing techniques, and network architecture are key factors in achieving optimal model performance. This article focuses on the application of Identity Blocks in the context of image processing, especially on unbalanced datasets. Three different datasets, namely Plant Diseases, Rock & Paper Scissors, and Animal Faces, are used in this study, each with unique characteristics. Identity Block, implemented in the ResNet network architecture, helps to overcome the gradient loss problem that often occurs in deep neural networks (DNN) with deep layers. This research specifically explores Identity Block optimization using the hyperband method to improve model performance. The average performance improvement of all optimized models is 4.45% in accuracy, 5.39% in precision, 6.4% in recall, and 6.48% in F1-score. These results show that model optimization is very good at improving identity block performance using the hyperband method.