基于迁移学习的红外图像超分辨率方法研究

针对红外图像空间分辨率低、成像质量不高的问题,提出了基于迁移学习的红外图像超分辨率方法。该方法以基于卷积神经网络的自然图像超分辨率方法为基础进行改进:增加网络的层数进行更深层次的学习训练,串联多层小的卷积核使其能够利用更多的图像信息,以“相差图”为目标进行训练,减小网络训练时间,提升网络收敛速度;利用迁移学习知识,再以少量高质量红外图像为目标样本,对自然图像超分辨率的网络进行二次训练,将网络权重经过微调后迁移应用到红外图像的超分辨率上。实验结果表明:基于卷积神经网络的超分辨率方法能够有效迁移应用到红外图像的超分辨率上,且改进后的网络具有更好的自然及红外图像的超分辨率性能,验证了本文所提方法的有效性及优越性。

Research on infrared image super-resolution based on transfer learning

Aiming at the problem of low spatial resolution and low image quality of the infrared images,an infrared image super-resolution method based on transfer learning was proposed.The proposed method was improved from three aspects based on natural image super-resolution method using convolutional neural network.Firstly,it increased the number of network layers so as to carry on deeper learning and training.Besides,it cascades a serial of small convolution kernels to make more use of contextual information in original images,and trained with the goal of residual image in order to reduce the training time and lift the convergence speed.Lastly,according to the knowledge of transfer learning,the network of natural image super-resolution was trained one more time based on the existed network with a small amount of infrared images in high quality as the target samples,and the weights of improved network for natural image super-resolution was fine-tuned and then applied to the infrared image super-resolution.The results of experiments show that the super-resolution method based on convolutional neural network can be effectively transferred to the application of infrared image super resolution,and the improved network plays a largely better role in natural and infrared image super-resolution,which proves the validity and superiority of the proposed method.