Multi-Spectral Object Detection Based on Multi-Scale Spatial-Spectral Interaction Network

近年来，可见光图像与热红外图像结合的多光谱目标检测，因其互补特性得到了广泛应用。然后，现有的大多数多光谱目标检测模型主要关注图像的局部特性，忽视了图像全局特征的提取，同时在特征提取过程中往往会丢失关键信息，如纹理和边缘等细节，导致提取的图像特征信息不足。针对这些问题，本文提出了多光谱目标检测模型SSIDet。该模型通过构建多尺度编码网络，分别从热红外图像和可见光图像中提取不同尺度的局部–全局特征；接着设计了一种空间–光谱交互注意力网络，充分融合空间特征和光谱特征，同时通过减少特征之间的冗余来增强其互补性；最后引入多尺度重建网络，进一步实现空间特征与光谱特征的协同增强。通过在FLIR和LLVIP数据集上的大量实验验证，本文方法在性能上优于现有方法。In recent years, multispectral object detection combining visible and thermal infrared images has been widely used due to its complementary characteristics. Then, most of the existing multispectral object detection models mainly focus on the local characteristics of the image, neglecting the extraction of global features of the image, and at the same time, key information, such as details of texture and edges, are often lost in the process of feature extraction, which leads to insufficient information of the extracted image features. Aiming at these problems, this paper proposes a multispectral object detection model SSIDet. The model extracts local-global features at different scales from thermal infrared images and visible images respectively by constructing a multiscale coding network; then a spatial-spectral interactive attention network is designed to fully integrate spatial and spectral features, and at the same time, its complementarity is enhanced by reducing redundancy between features; finally, a multiscale reconstruction network is introduced to further enhance feature extraction, and a multiscale reconstruction network is introduced to further enhance feature extraction. A multi-scale reconstruction network is introduced to further realize the synergistic enhancement of spatial and spectral features. Through extensive experimental validation on FLIR and LLVIP datasets, the method of this paper outperforms the existing methods in terms of performance.