红外小目标检测技术在军事侦查、精确制导、智慧交通领域具有重要的应用价值。然而，由于大气相关特性和较远的成像距离，红外图像中目标尺寸非常小且缺乏纹理形状。同时，目标的灰度显著性较弱且与背景像素占比极不均衡，导致红外小目标检测难度极大。针对上述问题，本文基于深度学习进行红外小目标检测相关研究。

针对复杂多变的红外成像场景以及图像中存在与小目标特征相似的背景干扰，导致现有算法检测精度低的问题，本文提出了一种基于双注意力机制的红外小目标检测算法，旨在提升模型检测性能。首先采用双注意力机制调整特征图中目标和背景的权重，实现背景与目标的区分。其次引入空洞卷积模块，以增强算法对红外小目标的感知能力和上下文关联性。此外，优化特征融合模块以聚合多尺度的特征信息，减少目标信息损失。在NUAA-SIRST、NUDT-SIRST和IRSTD-1K数据集上实验证明，本文算法主观评价优于现有检测算法，并且客观指标IoU和nIoU也表现出色。

针对红外小目标易受噪声和杂波干扰，以及低对比度小目标常被背景元素覆盖，从而引发虚警和漏检问题，本文提出了一种基于全局特征交互注意力的红外小目标检测算法，旨在减少漏检和虚警情况。首先结合Transformer结构和卷积神经网络构造卷积融合转换器模块，用于提取红外图像全局特征信息，并计算特征图的双线性注意力矩阵，抑制噪声杂波干扰。其次引入U形空洞卷积获取低分辨率图像的语义特征，以增强小目标的定位和辨别能力。此外，采用交互式注意力模块保留更多上下文信息，以提升小目标检测性能。实验结果表明，对比其他检测算法，本文所提算法能够有效减少漏检和降低虚警率。

Infrared small target detection technology has important application value in the fields of military reconnaissance, precision guidance, and smart transportation. However, due to the atmospheric correlation characteristics and long imaging distance, the target size in infrared images is very small and lacks texture shape. At the same time, the grayscale significance is weak and the proportion of pixels to the background is extremely unbalanced, making it extremely difficult to detect infrared small targets. In response to the above problems, this article conducts research on infrared small target detection based on deep learning.

In view of the complex and changeable infrared imaging scenes and the existence of background interference similar to small target features in the image, which leads to the low detection accuracy of existing algorithms, this thesis proposes an infrared small target detection algorithm based on a dual attention mechanism, aiming to improve model detection performance. First, a dual attention mechanism is used to adjust the weight of the target and the background in the feature map to achieve the distinction between the background and the target. Secondly, a dilated convolution module is introduced to enhance the algorithm's perception and contextual relevance of small infrared targets. In addition, the feature fusion module is optimized to aggregate multiscale feature information and reduce target information loss. Experiments on NUAA-SIRST, NUDT-SIRST and IRSTD-1K datasets prove that the subjective evaluation of this algorithm is better than existing detection algorithms, and the objective indicators IoU and nIoU are optimal.

In view of the fact that small infrared targets are susceptible to noise and clutter interference, and small targets with low contrast are often covered by background elements, causing false alarms and missed detection problems. This thesis proposes an infrared small target detection algorithm based on global feature interactive attention. Designed to reduce missed detections and false alarms. First, a convolution fusion Transformer module is constructed by combining the Transformer structure and the convolutional neural network to extract the global feature information of the infrared image, and calculate the bilinear attention matrix of the feature map to suppress noise and clutter interference. Secondly, U-shaped dilated convolution is introduced to obtain the semantic features of low-resolution images to enhance the localization and discrimination capabilities of small targets. In addition, an interactive attention module is used to retain more contextual information to improve small target detection performance. Experimental results show that compared with other detection algorithms, the algorithm proposed in this paper can effectively reduce missed detections and reduce the false alarm rate.