Abstract: Region-prior-based camouflaged (COD) models extract target location information in the early stages and integrate it with visual features to enhance detection accuracy. Current COD algorithms, however, struggle to extract precise target localization under challenging scenarios, resulting in inaccurate regional priors or target omission. The proposed FLMCA-Net addresses these limitations through three dedicated components. Firstly, the positioning module aggregates multi-scale backbone features to capture target positions from global and local perspectives. Secondly, the receptive field enhancement module refines multi-scale feature representations while preserving spatial details. Finally, the context-aware fusion module establishes multi-scale interaction between regional priors and image features, simultaneously extracting global contextual patterns. Experimental results on three publicly available datasets demonstrate that FLMCA-Net outperforms 20 other representative models, with an average performance improvement of 2.5%, 1.8%, and 3.6% over RISNet in terms of , , and , respectively.

Key words: camouflage object detection, contextual information, region prior, multi-scale feature

摘要： 基于区域先验的伪装目标检测模型通常在早期阶段提取目标在图像中的位置信息，通过将目标的位置信息与图像特征相结合，网络能够更加有效地聚焦于目标区域，从而提高检测精度。而在许多具有挑战性的场景中，当前伪装目标检测算法对目标位置信息的提取能力有限，从而导致区域先验不准确甚至目标缺失。为此，提出了一种融合定位图与上下文感知的伪装目标检测网络FLMCA-Net。首先，构建定位模块融合来自主干网络的多尺度特征并从全局、局部两个维度上提取目标的位置信息；其次，设计了感受野增强模块细化不同尺度的特征表示；最后，引入上下文感知融合模块将多尺度特征和目标区域先验充分交互融合并挖掘目标的全局上下文特征。在3个公开数据集上的实验表明，FLMCA-Net相较于其他20种代表性模型均取得了最优表现，比RISNet在、和方面的性能平均提升了2.5%、1.8%和3.6%。

关键词: 伪装目标检测, 上下文信息, 区域先验, 多尺度特征