OPTIMASI DETEKSI KELAS MINORITAS PADA DEEP LEARNING: EVALUASI KINERJA AUGMENTASI VISUAL MELAWAN TEKNIK SMOTE

Poetra Ebeline Abel; Odi Nurdiawan; Arif Rinaldi Dikananda; Aris Pratama Putra

Authors

Poetra Ebeline Abel STMIK IKMI, Indonesia
Odi Nurdiawan STMIK IKMI Cirebon, Indonesia
Arif Rinaldi Dikananda STMIK IKMI Cirebon, Indonesia
Aris Pratama Putra STMIK IKMI Cirebon, Indonesia

Keywords:

Ketidakseimbangan Data, SMOTE, Augmentasi Citra, MobileNetV2, Klasifikasi Biner

Abstract

Ketidakseimbangan kelas merupakan salah satu tantangan paling signifikan dalam pengembangan model klasifikasi citra, terutama pada tugas klasifikasi biner keberadaan objek. Ketimpangan jumlah sampel antara kelas positif dan negatif menyebabkan model cenderung bias terhadap kelas mayoritas sehingga mengurangi kemampuan mendeteksi kelas minoritas. Penelitian ini bertujuan melakukan analisis komparatif terhadap dua pendekatan penyeimbangan data, yaitu augmentasi citra berbasis transformasi visual dan Synthetic Minority Oversampling Technique (SMOTE) berbasis ruang fitur. Dataset yang digunakan terdiri dari 1.001 citra hasil pembersihan dan pelabelan, yang kemudian dibagi secara stratifikasi menjadi 800 citra latih dan 201 citra uji. Dua pipeline pengolahan dibangun secara terpisah: (1) pelatihan end-to-end menggunakan arsitektur MobileNetV2 dengan augmentasi visual on-the-fly, dan (2) ekstraksi fitur menggunakan MobileNetV2 yang diikuti oversampling fitur melalui SMOTE serta klasifikasi menggunakan Multi-Layer Perceptron (MLP). Evaluasi performa dilakukan menggunakan metrik akurasi, presisi, recall, dan F1-score untuk memberikan gambaran menyeluruh mengenai kemampuan prediktif model. Hasil penelitian menunjukkan bahwa model MLP + SMOTE unggul dalam akurasi (0.9552), presisi (0.9697), dan F1-score (0.9343), menandakan kestabilan prediksi yang tinggi. Sebaliknya, model CNN + augmentasi memperoleh recall tertinggi (0.9718), menunjukkan kepekaan lebih baik terhadap kelas minoritas. Temuan ini memperlihatkan adanya trade-off antara sensitivitas dan konsistensi performa, sehingga pemilihan metode penyeimbangan data perlu disesuaikan dengan kebutuhan aplikasi. Penelitian ini memberikan kontribusi empiris penting sebagai dasar pengembangan strategi penanganan ketidakseimbangan data pada klasifikasi citra.

References

N. Kaur and T. K. Gandhi, “A systematic review of class imbalance solutions in deep learning for image analysis,” Image Vis. Comput., vol. 138, p. 104752, 2023, doi: 10.1016/j.imavis.2023.104752.

W. Chen, K. Yang, Z. Yu, Y. Shi, and C. L. P. Chen, “A survey on imbalanced learning: latest research, applications and future directions,” Artif. Intell. Rev., 2024, doi: 10.1007/s10462-024-10759-6.

M. Hasan, T. Chowdhury, and G. Prince, “Challenges of object detection under severe class imbalance: A systematic analysis,” Image Vis. Comput., vol. 139, p. 104912, 2024, doi: 10.1016/j.imavis.2024.104912.

H. Wang, C. Sui, F. Jiang, S. Li, H. Liu, and A. Wang, “Value-guided adaptive data augmentation for imbalanced small object detection,” Electronics, vol. 13, no. 10, p. 1849, 2024, doi: 10.3390/electronics13101849.

D. Elreedy and A. Atiya, “A theoretical distribution analysis of synthetic minority oversampling technique (SMOTE) for imbalanced learning,” Mach. Learn., 2024, doi: 10.1007/s10994-022-06296-4.

S. Gholampour, “Impact of nature of medical data on machine and deep learning for imbalanced datasets: Clinical validity of SMOTE is questionable,” Mach. Learn. Knowl. Extr., vol. 6, no. 2, pp. 827–841, 2024, doi: 10.3390/make6020039.

S. Borah, S. Das, and S. Roy, “A novel data-level balancing strategy for minority object recognition in imbalanced datasets,” Expert Syst. Appl., vol. 233, p. 120846, 2023, doi: 10.1016/j.eswa.2023.120846.

E. Ersözlü, H. Taheri, and M. Koch, “Improved synthetic oversampling for deep feature-based classification on imbalanced datasets,” Sci. Rep., vol. 14, pp. 1124–1139, 2024.

Z. Wu and H. Zeng, “Image data augmentation techniques based on deep learning: A survey,” Math. Biosci. Eng., vol. 21, no. 6, pp. 6190–6224, 2024, doi: 10.3934/mbe.2024272.

H. Kaur, H. S. Pannu, and A. K. Malhi, “A Systematic Review on Imbalanced Data Challenges in Machine Learning: Applications and Solutions,” ACM Comput. Surv., vol. 52, no. 4, pp. 1–36, 2022, doi: 10.1145/3343440.

R. Gupta and A. Singh, “Analyzing annotation quality and its impact on object detection and classification performance,” Inf. Sci. (Ny)., vol. 660, p. 120097, 2024, doi: 10.1016/j.ins.2024.120097.

Y. Li, Q. Wang, and T. Zhang, “Revisiting object detection datasets for binary image classification through bounding-box guided supervision,” Pattern Recognit., vol. 140, p. 109573, 2023, doi: 10.1016/j.patcog.2023.109573.

J. Zhou, Y. He, and L. Huang, “A systematic evaluation of bounding-box transformations for downstream binary classification tasks,” Expert Syst. Appl., vol. 246, p. 123302, 2025, doi: 10.1016/j.eswa.2024.123302.

P. Thölke, G. Valente, B. Kießling, and T. Rohlfing, “Class imbalance should not throw you off balance: Choosing the right classifiers and performance metrics for brain decoding with imbalanced data,” Neuroimage, 2023, doi: 10.1016/j.neuroimage.2023.120253.

Y. Sun, Z. Zhang, and Z. Zhou, “A comprehensive evaluation framework for imbalanced data learning,” Pattern Recognit., vol. 141, p. 109537, 2023, doi: 10.1016/j.patcog.2023.109537.

M. Ahmad, F. Khan, and M. Alsaqer, “Explainable artificial intelligence for healthcare: A review,” Expert Syst. Appl., 2022, doi: 10.1016/j.eswa.2022.116952.

Z. Chen, H. Liu, and Y. Xu, “A unified representation of bounding box annotation for object-centric learning,” Neurocomputing, vol. 509, pp. 150–162, 2022, doi: 10.1016/j.neucom.2022.07.089.

S. Bhattacharya, S. Gupta, and P. Mitra, “Addressing extreme class imbalance in image classification using adaptive hybrid sampling,” Pattern Recognit., vol. 154, p. 110723, 2024, doi: 10.1016/j.patcog.2024.110723.

S. S. A. Zaidi, M. S. Ansari, A. Aslam, N. Kanwal, M. Asghar, and B. Lee, “A survey of modern deep learning based object detection models,” Digit. Signal Process., vol. 126, p. 103514, 2022, doi: 10.1016/j.dsp.2022.103514.