【作者】 朱琳；

【导师】 朱春华； 周飞； 刘浩；

【作者基本信息】 河南工业大学 ， 计算机技术（专业学位）， 2025， 硕士

【摘要】 红外弱小目标检测在军事侦察、导弹防御、电光系统等领域具有广泛应用,其关键技术之一是有效提取微弱目标信号并抑制复杂背景干扰。由于目标尺寸小、信噪比低,并且受大气扰动等因素的影响,现有的检测方法在精度和鲁棒性两个方面仍存在不足。在基于领域知识的算法中,基于滤波器的算法将目标作为离群值,通过原始背景和估算背景之间的差异来分离目标;基于显著性的方法在边缘出现高对比度特征时,会增强目标和杂波边缘,在强边缘干扰下也会更易产生误检;基于数据结构的方法利用目标的稀疏性将其进行分离,在分离过程中很可能将稀疏的目标和边缘一同滤除从而造成目标丢失。而基于数据驱动的算法虽然能够自适应提取特征,但它们的性能往往受到数据集和先验信息的限制。因此,在红外弱小目标检测中如何准确把握目标细节、提升背景抑制效果以及实现高鲁棒性和实时性检测,成为亟需解决的问题。本研究将探索性地把领域知识作为先验信息引入基于深度学习的数据驱动模型,构建了一个基于结构张量自适应特征融合的红外弱小目标检测网络,结合结构张量特征提取与稀疏注意力特征融合,以降低背景杂波对模型整体性能的影响,从而提高红外弱小目标检测效果。主要创新性工作包括:(1)针对红外弱小目标检测中连续的背景边缘杂波难以抑制的问题,提出了一种基于结构张量自适应权衡的红外弱小目标检测方法。该方法利用结构张量有效捕捉图像的纹理信息,通过构建边缘感知因子,实现了对背景中稀疏性杂波的刻画,有效地降低了背景杂波的连续性;引入平衡权重值来结合角点感知指标,使算法能够自适应的对杂波边缘和角点分别进行刻画,从而将目标从图像中进行差异化区分;为解决传统方法中阈值设定不自适应、目标与背景同时被强化的问题,进一步采用结构相似度指标设计自适应加权策略,动态调整平衡因子,实现不同场景下的弱小目标最优检测。在不同杂波背景下的六个数据集上进行实验,信噪比增益最高达到了253.52,背景抑制因子最高达到1777.38,表明所提方法具有良好的目标增强和背景抑制能力,为后续检测精度的提升提供了坚实的理论基础。(2)为进一步提高红外弱小目标的检测精度,提出了一种基于结构张量自适应特征融合的红外弱小目标检测网络。在该网络中,引入了已有的边缘感知因子,设计了基于结构张量的特征提取模块,该模块借助结构张量的优势,将先验信息引入到该网络模型中,对多尺度局部对比度进行加权处理,实现特征提取中细节信息的刻画;然后在自底向上注意力调制模块的基础上,引入了基于稀疏注意力的特征融合模块,通过自适应Top-K策略筛选最具判别力的目标特征,并结合结构相似度动态调整K值,实现跨层信息的优化融合。在SIRST数据集上进行实验,所提网络在保持网络轻量化和计算效率的同时,显著提升了检测精度,且在受试者工作特征曲线上检测率和虚警率的表现优于局部对比度网络等四种算法,在交并比和归一化交并比指标上分别最高达到了约0.781和0.752,表明了所提网络在复杂背景和低信噪比场景下具有更强的鲁棒性。更多还原

【Abstract】 Infrared small target detection has a wide range of applications in military reconnaissance,missile defense,electro-optical systems,etc.One of the key technologies is to effectively extract small target signals and suppress complex background interference.Due to the small size of the target,low signal-to-noise ratio,and the influence of atmospheric disturbances and other factors,the existing detection methods are still deficient in both accuracy and robustness.In domain-knowledge-based algorithms,filter-based algorithms treat the target as an outlier and separate the target by the difference between the original background and the estimated background;local information structure-based methods enhance the target and cluttered edges when high contrast features appear on the edges,and they are also more likely to produce false detections in the presence of strong edge disturbances;filter-based methods utilize the sparseness of the target to separate it,and in the process of separating,they are likely to filter out the sparse target and the data-driven algorithms can adaptively extract features,but their performance is often limited by the dataset and a priori information.Therefore,how to accurately grasp the details of the target,improve the background suppression effect,and realize the high robustness and real-time detection in infrared weak target detection has become an urgent problem to be solved.In this study,we will explore the introduction of domain knowledge as a priori information into a deep learning-based data-driven model,and construct a Structure Tensor&Attention Network based on Structure Tensor&Attention Network,which combines Structure Tensor Feature Extraction with Sparse Attention Feature Fusion,in order to reduce the impact of background clutter on the overall performance of the model,and thus to improve the Infrared small target detection.The main innovative work includes:(1)Aiming at the problem that the continuous background edge clutter is difficult to suppress in Infrared Small Target Detection,an Infrared Small Target Detection Based on Adaptive Balance Structure Tensor Indicator is proposed.The method utilizes the structure tensor to effectively capture the texture information of the image,and realizes the carving of sparsity clutter in the background by constructing the edge perception factor,which effectively reduces the continuity of the background clutter;the introduction of the balanced weight value to combine with the corner perception index enables the algorithm to adaptively carve the clutter edges and the corners separately,so as to differentially differentiate the target from the image;in order to solve the problem of the traditional method in which the In order to solve the problem that the threshold setting is not adaptive and the target and background are strengthened at the same time in the traditional method,the structural similarity index is further used to design an adaptive weighting strategy to dynamically adjust the balancing factor to realize the optimal detection of weak targets in different scenes.Experiments are carried out on six datasets under different clutter backgrounds,and the Signal-to-Noise Ratio Gain reaches up to 253.52 and the Background Suppression Factor reaches up to 1777.38,which indicates that the proposed method has good target enhancement and background suppression ability,and provides a solid theoretical foundation for the subsequent improvement of detection accuracy.(2)In order to further improve the detection accuracy of Infrared small targets,a Structure Tensor&Attention Network with adaptive feature fusion is proposed.In this network,the existing edge perception factor is introduced,and a structural tensor-based feature extraction module is designed,which introduces the a priori information into this network model with the advantage of structural tensor,weights the multi-scale Local Contrasts,and realizes the portrayal of detail information in the feature extraction;and then,on the basis of the bottom-up Attentional Modulation Module,a sparse-attention-based Feature Fusion module,which screens the most discriminative target features through the adaptive Top-K strategy and dynamically adjusts the K-value in combination with the structural similarity to achieve the optimal fusion of cross-layer information.Experimented on the SIRST dataset,the proposed network significantly improves the detection accuracy while maintaining the network lightweight and computational efficiency,and outperforms four algorithms such as Receiver-Operating-Characteristic Networks and four other algorithms in terms of detection and false alarm rates on Attentional Local Contrast Networks on the Receiver-Operating-Characteristic Curve,and reaches the highest of about 0.781 and 0.752 in the Intersection-Operating-Characteristic Ratio and Normalized Intersection-Operating-Characteristic Ratio metrics,which demonstrates the proposed network has a higher detection accuracy in the complicated background and low signal-to-noise ratio scenarios with stronger robustness.更多还原