重构迁移学习的红外目标分类

红外图像目标分类在目标识别等领域有重要的应用价值,目前卷积神经网络在可见光图像分类方面达到了优异的性能。但对于红外图像来说,由于有标记样本数量少和图像成像差异大,直接使用现有的网络模型来处理红外图像无法取得理想效果。该文将可见光图像作为源域,将红外图像作为目标域,在深度网络中使用迁移学习方法来解决此问题。在迁移学习中,目标域网络提取的特征越能体现出本域数据的真实分布,那么在此基础上进行两个域的分布适配就更加有效,迁移后的目标域网络性能和泛化能力越好。该文首先利用大量无监督的红外样本训练了红外图像深度卷积自编码器,增强了红外图像域网络的特征表达能力。其次,通过减小源域和目标域的特征分布距离,使得两个图像域特征分布相似,从而将源域中深度网络的学习能力迁移到目标域。经过上述改进,相比于可见光图像预训练微调的方法,分类准确率提升了11.27%。

Infrared Target Classification with Reconstruction Transfer Learning

Infrared target classification has important values in target recognition. At present, convolutional neural network has achieved excellent performance in visible image classification. However, for infrared images, the available networks can't achieve satisfying results due to the small number of annotated samples and large imaging differences. In this paper, visible images are used as source domain, infrared images as target domain. Transfer learning is used to address the challenges in the deep learning framework. In the transfer learning, if the target domain network can represent the distribution of its domain well, the performance and generalization of the target domain network should be more effective. Therefore, the convolutional autoencoder is trained with a large number of unannotated infrared samples, which greatly enhances the feature representation in the infrared image domain. By reducing the feature distribution distance between the two domains, the feature distributions become similar. The classification performance in the source domain is transferred to the target domain. With the changes above, the accuracy rate is improved by 11.27% compared with the method based on the visible images fine-tuning.