Automated camouflage pattern design based on conditional generative adversarial network and image quilting

The rapid advancement of reconnaissance technologies has posed significant challenges to traditional camouflage design. To address the limitations of manual dependency, poor environmental adaptability, and insufficient coverage for large targets, this study proposes an automated camouflage pattern design method that integrates a Conditional Generative Adversarial Network (CGAN) with image quilting. Background images are used as conditions for the CGAN to generate initial textures. Dominant colors are extracted using a K-means-based clustering method and combined with superpixel segmentation for color filling to synthesize camouflage patterns. To meet the requirements of large-scale targets, an image quilting algorithm is further implemented for seamless size expansion. A multidimensional evaluation system is developed to assess the camouflage performance, encompassing objective 2D quantitative metrics, human visual recognition experiments, and 3D forest-scene evaluations based on object detection algorithms. Experimental results show that the proposed method enhances Visual Information Fidelity (VIF) by 24.5%, reduces YOLOv8 detection confidence by 3.9%, increases human recognition time by 20.1%, and achieves more than 2× seamless texture expansion. Compared with traditional approaches, the method demonstrates significant improvements in both camouflage effectiveness and adaptability, highlighting its potential application in automated camouflage design for large-scale targets.