DMONet: deep multi-region joint fusion optimization network for camouflaged object detection

Abstract

While deep learning has significantly improved the accuracy of camouflaged object detection (COD) in recent years, several critical challenges still remain. Owing to the lack of effective multi-region joint analysis and the inconsistent semantic representations across different regions, existing methods often suffer from feature confusion among local areas. This issue can further result in redundant feature learning, inaccurate object localization, and severe over-segmentation around ambiguous or blurred boundaries. To address these limitations, we propose a Deep Multi-Region Joint Fusion Optimization Network (DMONet) for camouflaged object detection. Specifically, our framework integrates a multi-region perception network with a multi-scale feature fusion strategy to enhance both global contextual understanding and local structural awareness. It employs Global Vision Mamba Blocks (GVMB) and Local Vision Mamba Blocks (LVMB) to capture semantic dependencies across different regions, while a sparse Mamba mechanism is introduced to dynamically select informative features and suppress redundant responses. In addition, low-rank tensor decomposition is utilized to strengthen local-global feature interaction and improve cross-region complementary modeling. Finally, a hierarchical fusion module is designed to progressively refine and aggregate multi-region and cross-layer representations for accurate prediction. Extensive experiments on four public COD benchmarks show that DMONet achieves competitive performance and robust segmentation quality under complex backgrounds and challenging camouflage scenarios.