Scholarlink

<p>Camouflaged Object Detection (COD) aims to identify objects highly integrated with their environments. However, inherent scale diversity, complex contour features, and high similarity to the background of camouflaged objects pose significant challenges. Existing methods typically improve detection performance via complex model architectures while neglecting computational efficiency. Moreover, insufficient modeling of multi-scale and edge features leads to issues such as blurred boundaries and inaccurate localization. To address these challenges, this paper proposes a Progressive Edge-Aware Multi-Scale Fusion Network (PEMF-Net) for COD. The network leverages a progressive feature fusion framework to integrate edge semantics and multi-scale contextual information, enabling precise localization and refined segmentation of camouflaged objects. Specifically, we first design an Edge-Aware Enhancement Module (EAEM) to extract boundary information, providing edge semantic cues for object contour localization. Next, we introduce an Edge-Context Fusion Module (ECFM), which deeply integrates edge semantic cues with backbone features, guiding the model to focus on the correlation between object contours and contextual dependencies. Furthermore, we propose a progressive decoder that achieves multi-scale feature interaction and context-aware enhancement through the cascaded Cross-scale Fusion Units (CFUs). This decoder adopts a top-down fusion strategy, leveraging high-level semantic features to refine low-level detail representations, thereby further optimizing camouflaged prediction results. Through the collaborative learning among these modules, PEMF-Net generates prediction results with complete structures and distinct boundaries. Extensive experiments on four challenging benchmark datasets show that PEMF-Net outperforms 23 state-of-the-art models.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Progressive edge-aware multi-scale fusion network for camouflaged object detection

  • 期刊论文

  • 发表日期：

    2026年2月3日

查看全文
  • Jinqi Wang,
  • Mengyin Wang

摘要

Camouflaged Object Detection (COD) aims to identify objects highly integrated with their environments. However, inherent scale diversity, complex contour features, and high similarity to the background of camouflaged objects pose significant challenges. Existing methods typically improve detection performance via complex model architectures while neglecting computational efficiency. Moreover, insufficient modeling of multi-scale and edge features leads to issues such as blurred boundaries and inaccurate localization. To address these challenges, this paper proposes a Progressive Edge-Aware Multi-Scale Fusion Network (PEMF-Net) for COD. The network leverages a progressive feature fusion framework to integrate edge semantics and multi-scale contextual information, enabling precise localization and refined segmentation of camouflaged objects. Specifically, we first design an Edge-Aware Enhancement Module (EAEM) to extract boundary information, providing edge semantic cues for object contour localization. Next, we introduce an Edge-Context Fusion Module (ECFM), which deeply integrates edge semantic cues with backbone features, guiding the model to focus on the correlation between object contours and contextual dependencies. Furthermore, we propose a progressive decoder that achieves multi-scale feature interaction and context-aware enhancement through the cascaded Cross-scale Fusion Units (CFUs). This decoder adopts a top-down fusion strategy, leveraging high-level semantic features to refine low-level detail representations, thereby further optimizing camouflaged prediction results. Through the collaborative learning among these modules, PEMF-Net generates prediction results with complete structures and distinct boundaries. Extensive experiments on four challenging benchmark datasets show that PEMF-Net outperforms 23 state-of-the-art models.