ConvNeXt-Tiny for High-Accuracy Natural Scene Image Classification with Test-Time Augmentation | ||
| Journal of Information Technology Management | ||
| دوره 18، شماره 3، 2026، صفحه 228-242 اصل مقاله (1.01 M) | ||
| نوع مقاله: Research Paper | ||
| شناسه دیجیتال (DOI): 10.22059/jitm.2026.108385 | ||
| نویسندگان | ||
| Mohammed Hamid Alkubaisi1؛ Issa Mohammed Mishaal* 2؛ Bassam Talib Sabri3؛ Kian Raheem Qasim4 | ||
| 1Department of Studies and Planning, University of Information Technology and Communications, Baghdad, Iraq. | ||
| 2Department of Scientific Affairs, University of Information Technology and Communications, Baghdad, Iraq. | ||
| 3Department of Business Information Technology, University of Information Technology and Communications, Baghdad, Iraq. | ||
| 4Department of Studies and Planning, University of Information Technology and Communications, Baghdad, Iraq. | ||
| چکیده | ||
| The ConvNeXt-Tiny architecture combined with Test-Time Augmentation is used in this paper to present a strong deep learning system for multi-class natural scene image classification. Scene classification is a basic computer vision problem that has a wide range of applications, including autonomous navigation and environmental monitoring. We are using the Intel Image Classification dataset, which consists of 6 categories: buildings, forest, glacier, mountain, sea, and street. Thanks to this cutting-edge ConvNeXt-Tiny backbone, based on the transfer learning paradigm with a pre-trained ConvNeXt-Tiny backbone, a state-of-the-art CNN incorporating the design principles of Vision Transformers, we achieve very high feature extraction efficiency.TTA is also adopted so as to further enhance prediction robustness and reliability with visual changes during the inference stage. Experimental results have shown that the proposed method provides a Macro F1-score of 94.83% and an accuracy on test data of 94.70%. Based on the comparisons made herein, our method had an improved performance for all benchmarks from 2023–2025, including modified ResNet50 models and architectures for the Swin Transformers. The performance of our method is better than many available current benchmarks of 2023-2025 that include modified ResNet50 models and architectures of Swin Transformers, while maintaining the computational efficiency of the pure convolutional models. | ||
| کلیدواژهها | ||
| Convolutional Neural Networks؛ ConvNeXt Architecture؛ Computer Vision؛ Deep Learning؛ Image Classification | ||
| مراجع | ||
|
| ||
|
آمار تعداد مشاهده مقاله: 63 تعداد دریافت فایل اصل مقاله: 31 |
||