基于边界特征融合和前景引导的伪装目标检测

Boundary Feature Fusion and Foreground Guidance for Camouflaged Object Detection

伪装目标检测旨在检测隐藏在复杂环境中的高度隐蔽物体,在医学、农业等多个领域有重要应用价值.现有方法结合边界先验过分强调边界区域,对伪装目标内部信息的表征不足,导致模型对伪装目标的内部区域检测不准确.同时,已有方法缺乏对伪装目标前景特征的有效挖掘,使背景区域被误检为伪装目标.为解决上述问题,本文提出一种基于边界特征融合和前景引导的伪装目标检测方法,该方法由特征提取、边界特征融合、主干特征增强和预测等若干个阶段构成.在边界特征融合阶段,首先,通过边界特征提取模块获得边界特征并预测边界掩码;然后,边界特征融合模块将边界特征和边界掩码与最低层次的主干特征有效融合;同时,加强伪装目标边界位置及内部区域特征.此外,设计前景引导模块,利用预测的伪装目标掩码增强主干特征,即将前一层特征预测的伪装目标掩码作为当前层特征的前景注意力,并对特征执行空间交互,提升网络对空间关系的识别能力,使网络关注精细而完整的伪装目标区域.本文在4个广泛使用的基准数据集上的实验结果表明,提出的方法优于对比的19个主流方法,对伪装目标检测任务具有更强鲁棒性和泛化能力.

Camouflage object detection aims to detect highly concealed objects hidden in complex environments,and has important application value in many fields such as medicine and agriculture.The existing methods that combine bound⁃ary priors excessively emphasize boundary area and lack the ability to represent the internal information of camouflaged ob⁃jects,resulting in inaccurate detection of the internal area of the camouflaged objects by the model.At the same time,exist⁃ing methods lack effective mining of foreground features of camouflaged objects,resulting in the background area being mistakenly detected as camouflaged object.To address the above issues,this paper proposes a camouflage object detection method based on boundary feature fusion and foreground guidance,which consists of several stages such as feature extrac⁃tion,boundary feature fusion,backbone feature enhancement and prediction.In the boundary feature fusion stage,the boundary features are first obtained through the boundary feature extraction module and the boundary mask is predicted.Then,the boundary feature fusion module effectively fuses the boundary features and boundary mask with the lowest level backbone features,thereby enhancing the camouflage object’s boundary position and internal region features.In addition,a foreground guidance module is designed to enhance the backbone features using the predicted camouflage object mask.The camouflage object mask predicted by the previous layer of features is used as the foreground attention of the current layer features,and performing spatial interaction on the features to enhance the network’s ability to recognize spatial relation⁃ships,thereby enabling the network to focus on fine and complete camouflage object areas.A large number of experimental results in this paper on four widely used benchmark datasets show that the proposed method outperforms the 19 mainstream methods compared,and has stronger robustness and generalization ability for camouflage object detection tasks.