OPTIMALISASI CNN RINGAN MENGGUNAKAN SPECAUGMENT DAN SHARPNESS-AWARE MINIMIZATION UNTUK MENINGKATKAN GENERALISASI PADA KLASIFIKASI SUARA BURUNG

David Suharjanto, NIM.: 24206052003 (2026) OPTIMALISASI CNN RINGAN MENGGUNAKAN SPECAUGMENT DAN SHARPNESS-AWARE MINIMIZATION UNTUK MENINGKATKAN GENERALISASI PADA KLASIFIKASI SUARA BURUNG. Masters thesis, UIN SUNAN KALIJAGA YOGYAKARTA.

Preview	Text (OPTIMALISASI CNN RINGAN MENGGUNAKAN SPECAUGMENT DAN SHARPNESS-AWARE MINIMIZATION UNTUK MENINGKATKAN GENERALISASI PADA KLASIFIKASI SUARA BURUNG) 24206052003_BAB-I_IV-atau-V_DAFTAR-PUSTAKA.pdf - Published Version Download (2MB) | Preview
	Text (OPTIMALISASI CNN RINGAN MENGGUNAKAN SPECAUGMENT DAN SHARPNESS-AWARE MINIMIZATION UNTUK MENINGKATKAN GENERALISASI PADA KLASIFIKASI SUARA BURUNG) 24206052003_BAB-II_sampai_SEBELUM-BAB-TERAKHIR.pdf Restricted to Registered users only Download (5MB) | Request a copy

Abstract

Bird sound classification based on artificial intelligence is a promising approach for supporting automatic bird species identification. However, the generalization capability of deep learning models remains a challenge, particularly when confronted with noisy environments and out-of-distribution (OOD) data. This study aims to evaluate the effects of SpecAugment (Aug) and Sharpness-Aware Minimization (SAM) on the performance, robustness, generalization capability, and computational efficiency of lightweight convolutional neural network (CNN) models for bird sound classification. Experiments were conducted using the Birdsdata dataset as the primary dataset, UrbanSound8K for robustness evaluation under noisy conditions, and the Shorebird dataset as the OOD dataset. Three lightweight CNN architectures were evaluated: MobileNetV3-Small, EfficientNet-B0, and ConvNeXt-Atto. The evaluation included performance on clean data, robustness under noise at signal-to-noise ratio (SNR) levels ranging from -5 to 20 dB, generalization capability on OOD data, along with confusion matrix and Grad-CAM analyses, and inference efficiency on the Radxa ROCK 4 SE edge device. The results demonstrate that the combination of SpecAugment and SAM consistently improved the performance of all evaluated architectures compared with the baseline models, yielding F1-score improvements of 2.9-5.8 percentage points. ConvNeXt-Atto achieved the best performance with an accuracy of 96.7% and an F1-score of 96.1%, while also exhibiting superior robustness to noise and improved OOD detection capability based on the AUROC metric. These findings were supported by confusion matrix and Grad-CAM analyses. In edge deployment, ConvNeXt-Atto achieved the best trade-off between accuracy and efficiency, with 484.3 ms inference latency and 356.6 MB RAM usage. These findings demonstrate that SpecAugment and Sharpness-Aware Minimization improve the accuracy, robustness, and generalization of lightweight CNNs while maintaining computational efficiency.

Share this knowledge with your friends :

Actions (login required)

View Item