Claim Missing Document
Check
Articles

Found 1 Documents
Search

Improving Multi-Class Classification Performance with Multi-Scale Feature Extraction CNNs on a Limited Indonesian Spices Dataset Ahmad Zein Al Wafi; Alena Ghinna Kirana
JISKA (Jurnal Informatika Sunan Kalijaga) Vol. 11 No. 3 (2026): September 2026
Publisher : UIN Sunan Kalijaga Yogyakarta

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.14421/jiska.6185

Abstract

This study evaluates whether multi-scale feature extraction provides an advantage over single-scale CNN feature extraction under limited-data conditions. Using the Indonesian Spices Dataset as a controlled benchmarking testbed, which contains 31 distinct categories with only 210 images per class, we propose a multi-scale Convolutional Neural Network (CNN) architecture that captures features across varying spatial scales through three parallel branches with 3x3, 5x5, and 7x7 kernel sizes. The proposed model was evaluated against single-scale baselines using training and validation accuracy and loss, macro-averaged precision, recall, F1-score, and a confusion matrix. Experimental results show that the multi-scale approach provides the strongest learning capacity among the evaluated architectures, achieving the highest training accuracy (91.44%) and the lowest training loss (0.2713), while reaching a peak validation accuracy of 64.98% (62.90% final-epoch accuracy, 64.22% macro precision, and 62.59% macro F1-score) with a competitive training time of 3839.7s. The validation-accuracy advantage over the strongest single-scale baseline is modest; the clearer advantage is observed in training capacity and the broader validation metric profile. The late-epoch divergence between training and validation loss indicates that the additional representational capacity remains constrained by the limited number of training samples, highlighting the need for stronger regularization to improve generalization in scarce multi-class data.