摘要
伪装目标检测的任务是找到因颜色、纹理等相似特征而与背景混合的目标,而现有方法没有充分考虑边缘特征对检测性能的影响,存在漏检、错分等情况,检测精度仍需提升。为了克服以上不足,提出一种基于层内双分支相互增强注意力的伪装目标检测方法,该方法在现有多监督机制的基础上,引入对象边缘的预测模块,使模型的检测性能得到提升。为增强模型对物体的空间定位和识别能力,以Swin Transformer模型作为主干网络,设计了一种新型的层内双分支相互增强注意力模块,该模块包含双注意力增强模块和简单互增强模块。在CAMO、COD10K、NC4K等3个主流基准数据集上开展实验评估模型的性能,并将其与现有18种典型算法进行比较。结果表明:该模型具有优越的性能,在S_(α)、αE、ωF、MAE 4个性能指标上显著地优于现有18种先进的方法。
The task of camouflaged object detection is to identify objects that blend into the background due to similar features such as color and texture.However,current research on camouflaged target detection has not fully considered the impact of edge features on detection performance,and there are still issues such as missed detection and incorrect classification,thus the detection accuracy needs further improvement.To overcome the above shortcomings,this paper proposes a camouflaged object detection method based on intra-layer dual-branch mutual enhancement attention,which introduces an object edge prediction module into the existing multisupervision mechanism to further boost detection performance of the model.To enhance the model’s spatial localization and recognition capabilities for objects,with the Swin Transformer model as the backbone,a novel intra-layer dual-branch mutual enhancement attention module composed of a dual attention enhancement module and a simple mutual enhancement module is developed.Comprehensive experiments were conduct to evaluate the performance of the model on three mainstream benchmark datasets,including CAMO,COD10K,and NC4K,comparing which with 18 typical lgorithms existing.Experimental results show that the presented model has superior performance,significantly outperforming current eighteen state-of-the-art methods in terms of four evaluation indexes,including S_(α)、αE、ωF、MAE.
作者
苏嘉文
周之平
莫燕
SU Jia-wen;ZHOU Zhi-ping;MO Yan(School of Information Engineering,Nanchang Hangkong University,Nanchang 330063,China)
出处
《南昌航空大学学报（自然科学版）》
CAS
2024年第2期10-17,48,共9页
Journal of Nanchang Hangkong University(Natural Sciences)
基金
国家自然科学基金(62261038)。
关键词
伪装目标检测
双注意力增强
融合互增强
注意力机制
camouflage object detection
dual-attention enhancement
fusion mutual enhancement
attention mechanism
