基于多尺度耦合的密集残差网络红外图像增强

为了提升非制冷红外热像仪的图像质量,满足低对比度弱小区域的观瞄与锁定的需求,提出了一种基于多尺度密集残差网络的红外图像超分辨重建模型,该模型的基本框架是通过级联多个残差特征进行学习,以粗到细的方式重建高分辨率图像。首先提出一种多尺度跨域融合模块,通过对不同感受野的分支结果进行融合,不仅可以融合不同感受野的互补信息,还可有助于提升梯度收敛和特征传输;然后叠加多个跨域融合模块,并采用残差特征学习进行优化,最终学习出高分辨率细节信息。仿真实验结果表明,所提出的超分辨模型能够较好的超分辨重建效果,在微弱结构保持和点目标保持上的性能也更加突出。所提的模型已经在海思嵌入式深度学习平台上实现了高质量的红外增强,具有较高的工程应用价值。

Infrared image enhancement using dense residual network with multi-scale coupling

In order to improve the image quality of uncooled infrared thermal imager, and meet the needs of viewing and locking in low contrast and dim-area, a super-resolution reconstruction model of infrared image based on multi-scale dense residual network is proposed in this paper. The basic framework of the model is to reconstruct high-resolution image by cascading multiple residual features. Firstly, a multi-scale cross-channel fusion module is proposed. By fusing the branch results of different receptive fields, it not only fuses the complementary information of different receptive fields, but also helps to improve the gradient convergence and feature transmission. Then, multiple cross-fusion modules are cascaded and optimized by residual feature learning to learn high-resolution detail information. The simulation results show that the super-resolution model proposed in this paper can achieve better super-resolution reconstruction effect, and has better performance in weak structure maintenance and point target maintenance. Our proposed model has achieved highquality super-resolution reconstruction on the embedded deep learning platform of Hisilicon, and has high engineering application value.