Camouflage Target Detection Using Deep Learning

Abstract

This paper presents a novel approach for detecting camouflaged military targets using a hybrid architecture that combines a fine-tuned YOLOv8 detector with a Zero-Shot Learning (ZSL) classifier. Traditional object detection systems rely on extensive labeled datasets and predefined object classes, making them ineffective against dynamically camouflaged or unseen targets. The proposed system, CTD (Camouflage Target Detection), overcomes these limitations by decoupling the detection and classification tasks. It integrates a high- performance Spotter (YOLOv8) to find camouflaged objects in a class-agnostic manner, and a Vision-Language Identifier (CLIP) to classify them using flexible text prompts. This approach enables the system to identify and localize concealed objects across diverse environments and, crucially, to classify new targets without retraining. By eliminating the dependency on large annotated datasets for new classes, this approach significantly improves the accuracy, adaptability, and reliability of camouflage detection in modern aerial reconnaissance.