基于混合尺度感知与边缘特征交互的伪装目标检测

doi:10.13229/j.cnki.jdxbgxb.20241319

吉林大学学报(工学版) ›› 2026, Vol. 56 ›› Issue (7): 1996-2005.doi: 10.13229/j.cnki.jdxbgxb.20241319

• 计算机科学与技术 • 上一篇

基于混合尺度感知与边缘特征交互的伪装目标检测

蒲东兵¹(),张金峰¹,张美花¹,王一可²,孙英娟³()

^1.东北师范大学信息科学与技术学院，长春 130117
^2.沈阳城市学院智能与工程学院，沈阳 110112
^3.长春师范大学计算机科学与技术学院，长春 130032

收稿日期:2024-12-11 出版日期:2026-07-01 发布日期:2026-08-12
通讯作者: 孙英娟 E-mail:pudb@nenu.edu.cn;happysunyj@foxmail.com
作者简介:蒲东兵（1970-），男，副教授，博士.研究方向：计算机视觉，模式识别.E-mail：pudb@nenu.edu.cn
基金资助:
国家自然科学基金项目(62277009);吉林省重点科技研发项目(20220201157GX);吉林省重点科技研发项目(20220201140GX);吉林省教育厅科学技术研究项目(JJKH20241014KJ)

Camouflaged object detection based on mixed scale perception and edge feature interaction

Dong-bing PU¹(),Jin-feng ZHANG¹,Mei-hua ZHANG¹,Yi-ke WANG²,Ying-juan SUN³()

^1.College of Information Science and Technology，Northeast Normal University，Changchun 130117，China
^2.School of Intelligence and Engineering，Shenyang City University，Shenyang 110112，China
^3.College of Computer Science and Technology，Changchun Normal University，Changchun 130032，China

Received:2024-12-11 Online:2026-07-01 Published:2026-08-12
Contact: Ying-juan SUN E-mail:pudb@nenu.edu.cn;happysunyj@foxmail.com

摘要/Abstract

摘要：

针对已有伪装目标检测（COD）方法存在的边缘特征信息匮乏、多层次特征融合不充分等问题，提出了一种新颖的基于混合尺度感知与边缘特征交互的伪装目标检测网络（MSPEFI-Net）。MSPEFI-Net由混合尺度感知模块（MSP）、边缘特征交互模块（EFI）和全局特征聚合模块（GFA）组成。其中，MSP使用不同大小的卷积核提取不同尺度的特征，并进行跨尺度特征融合实现混合尺度感知，使模型能够有效扩大感受野，充分挖掘目标局部细节特征，以确保特征信息的完整性；EFI可以将高层语义信息与部分低层细节信息进行交互集成，以挖掘和补充边缘语义信息；GFA通过对上下文局部信息与边缘信息的跨层级特征聚合，以便有效地集成伪装目标及其边缘的特征信息，生成完整的预测图。实验结果表明，MSPEFI-Net在公开数据集CAMO、CHAMELEON、COD10K和NC4K上的 $S α$ 、 $F β w$ 、 $F β$ 、 $E φ$ 、 $M A E$ 等指标整体优于其他12种对比模型，具有更好的检测性能，消融实验也验证了各模块对模型性能提升的有效性。

关键词: 伪装目标检测, 混合尺度感知, 边缘特征交互, 特征增强

Abstract:

In view of the problems of lack of edge feature information and insufficient multi-level feature fusion in the existing amouflage object detection（COD） methods， a novel camouflaged object detection network MSPEFI-Net based on mixed-scale perception and edge feature interaction was proposed. MSPEFI-Net consists of a mixed-scale perception module（MSP），an edge feature interaction module（EFI）and a global feature aggregation module（GFA）.MSP extracts different scale features with different convolutional kernels，and fuses cross-scale features to achieve mixed-scale perception，so that MSPEFI-Net can effectively expand the receptive field and fully mine the local detail features of the object to ensure the integrity of feature information.EFI can interact and integrate high-level semantic information with some low-level detailed information to mine and supplement edge semantic information.GFA aggregates cross-level features from context local information and edge information to effectively integrate camouflage object features and its edge features，and generates a full prediction plot. Experimental results prove that the $S α$ 、 $F β w$ 、 $F β$ 、 $E φ$ and $M A E$ of MSPEFI-Net on the public datasets CAMO，CHAMELEON，COD10K and NC4K are superior to 12 comparison methods and show better detection performance of MSPEFI-Net.Ablation experiments also verify the effectiveness of each module to improve the performance of MSPEFI-Net.

Key words: camouflage object detection, mixed-scale perception, edge feature interaction, feature enhancements

中图分类号:

TP391.4

蒲东兵,张金峰,张美花,王一可,孙英娟. 基于混合尺度感知与边缘特征交互的伪装目标检测[J]. 吉林大学学报(工学版), 2026, 56(7): 1996-2005.

Dong-bing PU,Jin-feng ZHANG,Mei-hua ZHANG,Yi-ke WANG,Ying-juan SUN. Camouflaged object detection based on mixed scale perception and edge feature interaction[J]. Journal of Jilin University(Engineering and Technology Edition), 2026, 56(7): 1996-2005.

图/表 11

图1

图2

图3

图4

表1

基于CAMO、COD10K、CHAMELEON、NC4K数据集的性能对比"

方法	CAMO					CHAMELEON
方法	$S α$ ↑	$F β w$ ↑	MAE↓	$F β$	$E φ$ ↑	$S α$ ↑	$F β w$ ↑	MAE↓	$F β$	$E φ$ ↑
SINet^［1］	0.745	0.644	0.092	0.702	0.829	0.872	0.806	0.034	0.827	0.946
MGL-R^［3］	0.775	0.673	0.088	0.726	0.842	0.893	0.812	0.031	0.833	0.941
C²FNet^［2］	0.796	0.719	0.080	0.762	0.864	0.888	0.828	0.032	0.844	0.946
SINetv2^［11］	0.820	0.743	0.070	0.782	0.895	0.888	0.816	0.030	0.835	0.961
BGNet^［9］	0.812	0.749	0.073	0.789	0.882	0.901	0.850	0.027	0.860	0.954
ZoomNet^［4］	0.820	0.752	0.066	0.794	0.892	0.902	0.845	0.023	0.864	0.958
DGNet^［6］	0.839	0.769	0.057	0.806	0.915	0.890	0.816	0.029	0.834	0.956
FEDER^［5］	0.802	0.738	0.071	0.781	0.873	0.887	0.834	0.030	0.851	0.954
FDNet^［7］	0.841	0.775	0.063	0.807	0.908	0.897	0.835	0.027	0.854	0.964
FSPNet^［8］	0.856	0.799	0.050	0.830	0.928	0.908	0.851	0.023	0.867	0.965
FINet^［16］	0.828	0.752	0.065	0.791	0.900	0.883	0.808	0.031	0.825	0.957
UEDGNet^［15］	0.867	0.816	0.049	0.841	0.931	0.909	0.862	0.023	0.874	0.969
MSPEFI-Net	0.867	0.825	0.046	0.854	0.929	0.910	0.862	0.020	0.874	0.964
方法	COD10K					NC4K
方法	$S α$ ↑	$F β w$ ↑	MAE↓	$F β$	$E φ$ ↑	$S α$ ↑	$F β w$ ↑	MAE↓	$F β$	$E φ$ ↑
SINet^［1］	0.776	0.631	0.043	0.679	0.874	0.808	0.723	0.058	0.769	0.883
MGL-R^［3］	0.814	0.666	0.035	0.710	0.890	0.833	0.739	0.053	0.782	0.893
C²FNet^［2］	0.813	0.686	0.036	0.723	0.900	0.838	0.762	0.049	0.795	0.904
SINetv2^［11］	0.815	0.680	0.037	0.718	0.906	0.847	0.770	0.048	0.805	0.914
BGNet^［9］	0.831	0.722	0.033	0.753	0.911	0.851	0.788	0.044	0.820	0.916
ZoomNet^［4］	0.838	0.729	0.029	0.766	0.911	0.853	0.784	0.043	0.818	0.912
DGNet^［6］	0.822	0.693	0.033	0.728	0.911	0.857	0.784	0.042	0.814	0.922
FEDER^［5］	0.822	0.716	0.032	0.751	0.905	0.847	0.789	0.044	0.824	0.915
FDNet^［7］	0.840	0.729	0.030	0.757	0.935	0.834	0.750	0.052	0.784	0.905
FSPNet^［8］	0.851	0.735	0.026	0.769	0.930	0.879	0.816	0.035	0.843	0.937
FINet^［16］	0.817	0.686	0.034	0.723	0.908	0.847	0.771	0.047	0.805	0.911
UEDGNet^［15］	0.856	0.760	0.026	0.785	0.932	0.882	0.828	0.035	0.852	0.938
MSPEFI-Net	0.867	0.784	0.022	0.813	0.935	0.879	0.830	0.034	0.857	0.936

表1

图5

图6

图7

表2

基于CAMO、COD10K的消融实验结果"

模型	方法				CAMO					COD10K
模型	Basic	MSP	EFI	GFA	$S α$ ↑	$F β w$ ↑	MAE↓	$F β$	$E φ$ ↑	$S α$ ↑	$F β w$ ↑	MAE↓	$F β$	$E φ$ ↑
a	√				0.855	0.813	0.049	0.847	0.917	0.858	0.774	0.024	0.809	0.933
b	√	√			0.857	0.813	0.051	0.846	0.916	0.866	0.780	0.023	0.810	0.932
c	√	√	√		0.865	0.822	0.049	0.852	0.927	0.866	0.784	0.023	0.814	0.937
d	√	√	√	√	0.867	0.825	0.046	0.854	0.929	0.867	0.784	0.022	0.813	0.935

表2

表3

基于NC4K的消融实验结果"

模型	方法				NC4K
模型	Basic	MSP	EFI	GFA	$S α$ ↑	$F β w$ ↑	MAE↓	$F β$	$E φ$ ↑
a	√				0.869	0.820	0.037	0.853	0.931
b	√	√			0.875	0.824	0.035	0.853	0.932
c	√	√	√		0.878	0.827	0.035	0.855	0.934
d	√	√	√	√	0.879	0.830	0.034	0.857	0.936

表3

表4

基于CAMO、COD10K、NC4K对GFA的消融实验结果"

模型	方法	CAMO					COD10K					NC4K
模型	方法	$S α$ ↑	$F β w$ ↑	MAE↓	$F β$	$E φ$ ↑	$S α$ ↑	$F β w$ ↑	MAE↓	$F β$	$E φ$ ↑	$S α$ ↑	$F β w$ ↑	MAE↓	$F β$	$E φ$ ↑
a	W/CA	0.865	0.823	0.047	0.853	0.927	0.867	0.785	0.022	0.811	0.933	0.879	0.829	0.034	0.855	0.932
b	W/SA	0.862	0.819	0.050	0.850	0.920	0.865	0.777	0.024	0.808	0.932	0.876	0.825	0.035	0.853	0.933
c	W/MSCA	0.867	0.825	0.046	0.854	0.929	0.867	0.784	0.022	0.813	0.935	0.879	0.830	0.034	0.857	0.936

表4

参考文献 16

[1]	Fan D P, Ji G P, Sun G, et al. Camouflaged object detection[C]∥IEEE/CVF Conference on Computer Vision and Pattern Recognition, Seattle, USA, 2020: 2777-2787.
[2]	Sun Y, Chen G, Zhou T,et al.Context-aware cross-level fusion network for camouflaged object detection[C]∥International Joint Conferences on Artificial Intelligence, Montreal, Canada, 2021: 1025-1031.
[3]	Zhai Q, Li X, Yang F,et al.Mutual graph learning for camouflaged object detection[C]∥IEEE/CVF Conference on Computer Vision and Pattern Recognition,Nashville, USA,2021: 12997-13007.
[4]	Pang Y, Zhao X, Xiang T Z,et al.Zoom in and out:a mixed-scale triplet network for camouflaged object detection[C]∥IEEE/CVF Conference on Computer Vision and Pattern Recognition, New Orleans, USA,2022: 2160-2170.
[5]	He C, Li K, Zhang Y,et al.Camouflaged object detection with feature decomposition and edge reconstruction[C]∥IEEE/CVF Conference on Computer Vision and Pattern Recognition,Vancouver,Canada, 2023:22046-22055.
[6]	Ji G P, Fan D P, Chou Y C,et al.Deep gradient learning for efficient camouflaged object detection[J].Machine Intelligence Research, 2023, 20(1): 92-108.
[7]	Song Y, Li X, Qi L.Camouflaged object detection with feature grafting and distractor aware[C]∥2023 IEEE International Conference on Multimedia and Expo(ICME), Lisbon, Portugal, 2023: 2459-2464.
[8]	Huang Z, Dai H, Xiang T Z,et al.Feature shrinkage pyramid for camouflaged object detection with transformers[C]∥EEE/CVF Conference on Computer Vision and Pattern Recognition, Vancouver, Canada,2023: 5557-5566.
[9]	Sun Y, Wang S, Chen C,et al.Boundary-guided camouflaged object detection[C]∥International Joint Conference on Artificial Intelligence, Vienna, Austria,2022:1335-1341.
[10]	Dai Y, Gieseke F, Oehmcke S,et al.Attentional feature fusion[C]∥IEEE/CVF Winter Conference on Applications of Computer Vision, Waikoloa, USA,2021: 3560-3569.
[11]	Fan D P, Ji G P, Qin X,et al.Cognitive vision inspired object segmentation metric and loss function[J].Scientia Sinica Informationis,2021, 51(9): 1475-1489.
[12]	王春华,李恩泽,肖敏.多特征融合和孪生注意力网络的高分辨率遥感图像目标检测[J]. 吉林大学学报: 工学版, 2024,54(1): 240-250.
	Wang Chun-hua, Li En-ze, Xiao Min.Object detection in high-resolution remote sensing images based on multi-feature fusion and simeametric attention network[J].Journal of Jilin University (Engineering and Technology Edition), 2024,54(1):240-250.
[13]	Mei H, Ji G P, Wei Z,et al.Camouflaged object segmentation with distraction mining[C]∥IEEE/CVF Conference on Computer Vision and Pattern Recognition, Nashville, USA, 2021:8772-8781.
[14]	Fan D P, Ji G P, Cheng M M, et al. Concealed object detection[J]. IEEE transactions on pattern analysis and machine intelligence, 2021, 44(10): 6024-6042.
[15]	Lyu Y, Zhang H, Li Y,et al.Uedg:uncertainty-edge dual guided camouflage object detection[J].IEEE Transactions on Multimedia,2023,26: 4050-4060.
[16]	Liang W, Wu J, Wu Y, et al.FINet: frequency injection network for lightweight camouflaged object detection[J]. IEEE Signal Processing Letters, 2024,31:526-530.

Metrics

Viewed

Full text

Abstract

Cited

Shared

Discussed

基于混合尺度感知与边缘特征交互的伪装目标检测

Camouflaged object detection based on mixed scale perception and edge feature interaction

RICH HTML

PDF (PC)

摘要/Abstract

引用本文

使用本文

图/表 11

参考文献 16

相关文章 2

Metrics

本文评价

推荐阅读 0

[1]	张帆,王景波,朱杨,刘海英,郑煜,王文华. 面向遥感图像舰船识别的轻量化目标检测算法[J]. 吉林大学学报(工学版), 2026, 56(2): 533-542.
[2]	霍震,金立生,华强,贺阳. 基于边缘特征引导的智能汽车语义分割方法[J]. 吉林大学学报(工学版), 2025, 55(9): 3032-3041.