城市戰中的坦克影像辨識問題初探

Abstract

無人機（UAV）在現代戰爭中已成為重要的情報與攻擊工具。坦克作為攻擊方核心單位，快速且準確的辨識系統對於戰場上的指揮、決策和自動化武器運用至關重要。城市戰因背景複雜且有煙霧、樹葉與建築物遮擋以及迷彩等干擾因素，使得傳統為開闊戰場所設計的無人機表現不佳，此外，目前全世界仍未有相關城市戰資料集以供機器學習。因此本研究首先設計了一套系統來快速將真實影像進行合成，用以生成多樣化的城市戰坦克資料集，並模擬多種城市戰遮擋情況。接著，本研究使用YOLOv10-S對合成圖像進行模型訓練與測試。結果顯示，在城市戰各種遮擋情況下，YOLOv10-S平均可達到97.375%的mAP50，Precision達到86.675%，Recall達到96.375%。於灰階影像中，模型仍能保持穩定表現，mAP50達94.525%。對抗性攻擊作為一種極端迷彩，對辨識率的破壞最為嚴重，使大部分坦克幾乎無法被辨識。複雜迷彩會顯著降低辨識之Recall，而灰階化處理能有效削弱迷彩干擾，使Recall回升至75%。在常見物體遮擋（如建築物與樹葉）下，辨識效能亦出現下降。最後，在熱源紅外線的影像中，mAP50、Precision和Recall分別可達99.5%、96.4%和99.6%，顯示模型能穩定學習熱度與形狀特徵。實驗也證明YOLOv10-S具有優異的泛化能力，透過合成影像增強資料集，不僅能有效提升模型在惡劣環境下的辨識效能，也為未來結合真實戰場影像與持續學習技術提供了可行的方向。

Unmanned Aerial Vehicles (UAVs) have become essential tools for intelligence gathering and offensive operations in modern warfare. As tanks serve as the core units of offensive forces, the development of rapid and accurate recognition systems is critical for battlefield command, decision-making, and autonomous weapon deployment. However, UAV performance in urban warfare is often hindered by complex backgrounds and interference factors such as smoke, foliage, buildings, and camouflage, which differ significantly from the open battlefield environments for which most UAV systems were originally designed. Moreover, no publicly available urban warfare dataset currently exists to support machine learning applications. To address this gap, this study designs a system that rapidly synthesizes realistic images to generate a diverse urban tank dataset and simulate various urban occlusion conditions. The YOLOv10-S model is then trained and evaluated on the synthesized images. The results show that under different urban occlusion conditions, YOLOv10-S achieves an average mAP@50 of 97.38%, with 86.68% precision and 96.38% recall. In grayscale images, the model maintains stable performance, achieving an mAP@50 of 94.53%. Adversarial camouflage, representing an extreme form of concealment, causes the most severe degradation in recognition accuracy, rendering most tanks nearly undetectable. Complex camouflage significantly reduces recall, while grayscale processing effectively mitigates such interference, raising recall to 75%. Under common object occlusions (e.g., buildings and foliage), recognition performance also declines. Finally, in thermal infrared imagery, the model achieves 99.5% mAP@50, 96.4% precision, and 99.6% recall, demonstrating its ability to effectively learn both heat and shape features. These results confirm that YOLOv10-S possesses excellent generalization capability. By enhancing datasets through synthetic imagery, the proposed approach not only improves recognition performance under adverse conditions but also provides a feasible direction for integrating real battlefield imagery and continual learning in future research.