基于注意力机制和多尺度特征的伪装目标检测  

Camouflaged Object Detection Based on Attention Mechanism and Multi-scale Features

作　　者：蔡俊敏 孙涵[1] CAI Jun-min;SUN Han(School of Computer Science and Technology/Artificial Intelligence,Nanjing University of Aeronautics and Astronautics,Nanjing 211106,China)

机构地区：[1]南京航空航天大学计算机科学与技术学院/人工智能学院,江苏南京211106

出　　处：《计算机技术与发展》2023年第8期131-136,共6页Computer Technology and Development

基　　金：中央高校基本科研业务费专项资金资助项目(NZ2019009)。

摘　　要：针对伪装目标结构多样、尺度不一和目标边界与其背景具有高度相似性的情况,提出了一种基于注意力机制和多尺度特征的伪装目标检测算法。该算法主要分为两个部分,分别是基于多尺度特征的混合尺度解码器和基于反向注意力机制的注意力引导模块。混合尺度解码器通过级联的特征融合单元,融合高层特征的语义信息与低层特征的空间细节信息,对特征编码器生成的特征金字塔进行解码,得到初步的检测结果;之后引入反向注意力机制,通过擦除图像中已经识别到的目标区域,来引导网络挖掘新的伪装线索,最终得到识别位置更准确、更完整的伪装目标。实验中采用COD10K数据集、四种评价指标,与现有的十三种算法进行了对比。实验结果表明,该伪装目标检测算法具有更好的性能表现。An algorithm for detecting camouflaged objects based on attention mechanism and multi-scale features is proposed for the situation that camouflaged objects have diverse structures,different scales and the object boundaries are highly similar to their backgrounds.The proposed algorithm is mainly divided into two parts,which are a mixed-scale decoder based on multi-scale features and an attention-guiding module based on the reverse attention mechanism.The mixed-scale decoder fuses the semantic information of high-level features with the spatial detail information of low-level features through a cascaded feature fusion unit to decode the feature pyramid generated by the feature encoder and obtain the preliminary detection results.After that,the reverse attention mechanism is introduced to guide the network to mine new camouflage cues by erasing the already recognized object regions in the image,and finally obtain a more accurate and complete camouflage object.The COD10K dataset and four evaluation metrics are used in the experiments,and the comparison is conducted with thirteen existing algorithms.The experimental results show that the proposed algorithm has better performance.

关 键 词：伪装目标检测 注意力机制 多尺度特征 深度学习 卷积神经网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]