The rapid advancement of direct and indirect reconnaissance technologies with wide-area and full-spectrum coverage has exposed limitations in traditional digital camouflage methods. Current camouflage techniques primarily rely on background color distribution, achieving initial concealment but lacking precise target edge feature perception. This deficiency results in visible target contours and inadequate environmental integration, compromising camouflage effectiveness. Consequently, the accurate perception of target edges and generation of seamlessly integrated digital camouflage images has emerged as a critical challenge. This study addresses these limitations by proposing an adaptive digital camouflage generation method based on deep learning. The approach combines the you only look once version 8 (YOLOv8) object detection algorithm with resolution-robust large mask inpainting with Fourier convolutions (LAMA), establishing a comprehensive closed-loop pipeline from target perception to camouflage generation. The effectiveness of the proposed method is evaluated against existing approaches using three objective metrics: structural similarity index (SSIM), peak signal-to-noise ratio (PSNR), and mean squared error (MSE). Experimental results indicate that the proposed method effectively processes target edge regions and generates digital camouflage images that achieve superior background integration.

This study presents an adaptive digital camouflage image generation algorithm leveraging deep learning capabilities. The algorithm automatically generates digital camouflage images that closely match the surrounding environment based on acquired background images, enabling target concealment in the visible spectrum. The methodology employs the YOLOv8 deep learning model for target object detection and mask generation, establishing the foundation for edge-aware camouflage processing. To overcome limited training data constraints, the YOLOv8 training process incorporates a few-shot learning strategy to enhance detection accuracy. The target mask and background image are subsequently processed through the digital camouflage generation network, which perceives target edge features while extracting background texture and semantic information through deep learning. Furthermore, a single-objective optimization algorithm refines the YOLOv8-generated target mask, facilitating the production of digital camouflage images with enhanced texture, color, and structural consistency with the background.

The performance evaluation of the proposed method involved comparison with six representative baseline methods, all applied to identical target objects (Fig. 7). Traditional approaches demonstrate inadequate perception of target color space information and environmental edge characteristics, resulting in poor spatial localization of color distribution. Consequently, the target region remains visually distinct, yielding low fusion quality in the generated camouflage image. While deep learning-based methods show improved fusion, they are hindered by excessive edge sharpness that compromises camouflage effectiveness. The proposed method successfully addresses target edge regions, producing camouflage images with markedly enhanced background integration. Quantitative assessment employed three established image quality metrics—SSIM, PSNR, and MSE. Higher SSIM and PSNR values, coupled with lower MSE, indicate superior visual similarity and background integration. Statistical analysis of the results, expressed as mean±standard deviation (x_Mean±x_Std), demonstrates that the proposed method achieves an average SSIM increase of 0.028, PSNR improvement of 0.14 dB, and MSE reduction of 1.14, confirming its superior performance. Additional experiments conducted across diverse background scenarios (Figs. 8?10) demonstrate consistent production of high-quality camouflage images, validating the method’s robustness and broad applicability.

This paper presents a deep learning-based optical adaptive digital camouflage generation method to address the limitations of traditional approaches, particularly their inadequate perception of target edge features and subsequent poor background integration. The proposed methodology encompasses two primary stages: target detection and camouflage image generation. The approach implements data augmentation techniques to enhance YOLOv8 detection network generalization, followed by the LAMA image inpainting algorithm for high-quality digital camouflage image generation through target mask optimization. Performance evaluation utilizes objective metrics including SSIM, PSNR, and MSE. Experimental results demonstrate superior performance across all objective evaluation metrics compared to conventional methods, with improvements of 0.028 in average SSIM, 0.14 dB in PSNR, and a reduction of 1.14 in MSE. Additional experiments across varied environmental backgrounds confirm the method’s robustness and general applicability.