Super-resolution Enhancement of Infrared Images Using a Lightweight Dense Residual Network
-
摘要: 现有的红外制导武器严重依赖操作手对目标的捕获,其捕获的精度与目标的纹理细节正相关。为了提升弱小区域的显示质量,满足现有导引头小型化、模块化、低成本的设计要求,本文设计了一种基于轻量级金字塔密集残差网络的图像增强模型,该模型在密集残差网络基础上通过密集连接层和残差网络来学习不同尺度图像之间的非线性映射,充分利用多尺度特征进行高频残差预测。同时,采用深度监督模块指导网络训练,有利于实现较大上采样因子的超分辨增强,提高其泛化能力。大量仿真实验结果表明本文所提出的超分辨模型能够获得高倍率的超分辨增强效果,其重建质量也优于对比算法。Abstract: Existing infrared-guided weapons heavily rely on operators to acquire targets, and the accuracy of acquisition is positively correlated with a target's texture details. To improve the display quality of weak small regions and meet the design requirements of miniaturization, modularization, and low-cost seekers, an image super-resolution(SR) reconstruction algorithm based on a pyramid dense residual network is proposed. The dense residual network is the basic framework of the proposed model. Through the dense connection layer and the residual network, the model can learn the non-linear mapping between images of different scales, and the multi-scale feature can be used to predict the high-frequency residual. In addition, using the deep supervision module to guide network training is conducive to the realization of SR reconstruction with a larger upper-sampling factor and improvements to its generalization ability. A large number of simulation results show that our proposed model outperforms comparison algorithms and that it has a high engineering application value.
-
Key words:
- SR reconstruction /
- lightweight /
- infrared image /
- dense residual-network /
- loss function /
- deep supervision /
-
表 1 不同模块性能分析
Table 1. Performance analysis of different modules
RL DC DS PSNR █ 29.8 █ 29.9 █ 29.7 █ █ 31.7 █ █ 31.0 █ █ 31.8 █ █ █ 32.5 表 2 不同算法的重建指标对比
Table 2. Comparison of reconstruction indexes of different algorithms
Images SRCNN EDSR Meta-SR GANSR SRMD Proposed 1 PSNR 32.087 32.237 32.297 32.347 32.507 32.777 SSIM 0.954 0.955 0.955 0.958 0.961 0.962 2 PSNR 22.907 23.507 23.187 23.547 23.187 23.827 SSIM 0.774 0.796 0.788 0.802 0.794 0.826 3 PSNR 24.147 25.617 24.467 25.477 24.567 26.287 SSIM 0.89 0.928 0.912 0.931 0.901 0.948 4 PSNR 32.767 32.697 32.777 32.787 33.047 33.087 SSIM 0.87 0.867 0.868 0.868 0.879 0.878 5 PSNR 29.297 30.207 29.557 30.167 29.567 30.487 SSIM 0.898 0.911 0.906 0.913 0.905 0.921 6 PSNR 28.657 28.557 28.557 28.697 28.627 29.037 SSIM 0.952 0.953 0.957 0.957 0.96 0.964 Average PSNR 28.307 28.807 28.467 28.837 28.577 29.247 SSIM 0.901 0.912 0.902 0.914 0.915 0.929 -
[1] 廖小华, 陈念年, 蒋勇, 等. 改进的卷积神经网络红外图像超分辨率算法[J]. 红外技术, 2020, 42(1): 075-80. http://hwjs.nvir.cn/article/id/hwjs202001011LIAO Xiaohua, CHEN Niannian, JIANG Yong, et al. Infrared image super-resolution using improved convolutional neural network[J]. Infrared Technology, 2020, 42(1): 075-80. http://hwjs.nvir.cn/article/id/hwjs202001011 [2] 田广强. 一种新颖高效的红外动态场景多目标检测跟踪[J]. 红外技术, 2018, 40(3): 259-263. http://hwjs.nvir.cn/article/id/hwjs201803010TIAN Guangqiang. A novel algorithm for efficient multi-object detection and tracking for infrared dynamic frames[J]. Infrared Technology, 2018, 40(3): 259-263. http://hwjs.nvir.cn/article/id/hwjs201803010 [3] 曾金发, 吴恩斯, 李能勇. 基于双核协同学习模型的红外目标跟踪算法[J]. 红外技术, 2018, 40(5): 438-443. http://hwjs.nvir.cn/article/id/hwjs201805006ZENG Jinfa, WU Ensi, LI Nengyong. Infrared object-tracking algorithm based on dual-kernelized collaborative learning[J]. Infrared Technology, 2018, 40(5): 438-443. http://hwjs.nvir.cn/article/id/hwjs201805006 [4] 韩团军, 尹继武. 一种鲁棒的自适应更新策略的弹载计算机红外目标跟踪算法[J]. 红外技术, 2018, 40(7): 625-631. http://hwjs.nvir.cn/article/id/hwjs201807001HAN Tuanjun, YIN Jiwu. Robust adaptive updating strategy for missile-borne infrared object-tracking algorithm[J]. Infrared Technology, 2018, 40(7): 625-631 http://hwjs.nvir.cn/article/id/hwjs201807001 [5] 艾志伟, 嵇建波, 李静, 等. 快速反射镜状态模型构建方法及其控制系统设计[J]. 红外技术, 2020, 42(1): 40-45. http://hwjs.nvir.cn/article/id/hwjs202001006AI Zhiwei, JI Jianbo, LI Jing, et al. State model construction method for fast steering mirror and its control system design[J]. Infrared Technology, 2020, 42(1): 040-45. http://hwjs.nvir.cn/article/id/hwjs202001006 [6] DONG C, Loy C C, He K, et al. Learning a deep convolutional network for image super-resolution[C]//European Conference on Computer Vision, 2014: 184-199. [7] Riegler G, Rther M, Bischof H. ATGV-Net: accurate depth super resolution[C]//Proc. Eur. Conf. Comput. Vis., 2016: 268-284. [8] SHI Y, LI S, LI W, et al. Fast and Lightweight Image Super-resolution based on dense residuals two-channel network[C]//2019 IEEE International Conference on Image Processing (ICIP), 2019: 2826-2830. [9] WANG Wei, JIANG Yongbin, LUO Yanhong, et al. An advanced deep residual dense network(DRDN) approach for image super-resolution[J]. International Journal of Computational Intelligence Systems, 2019, 12(2): 1592-1601. doi: 10.2991/ijcis.d.191209.001 [10] WU Y, JI X, JI W, et al. CASR: a context-aware residual network for single-image super-resolution[J]. Neural Computing and Applications, 2019: 1-16. doi: 10.1007/s00521-019-04609-8%3Fshared-article-renderer [11] 杨明, 王璇, 高宏伟. 基于多级深度网络的高清晰度红外电子变倍算法[J]. 弹箭与制导学报, 2020, 40(4): 14-19. https://www.cnki.com.cn/Article/CJFDTOTAL-DJZD202004004.htmYANG Ming, WANG Xuan, GAO Hongwei. High-definition infrared electronic zooming algorithm based on multi-level deep network[J]. Journal of Projectiles, Rockets, Missiles and Guidance, 2020, 40(4): 14-19. https://www.cnki.com.cn/Article/CJFDTOTAL-DJZD202004004.htm [12] HUI T, Loy C C, TANG X. Depth map super-resolution by deep multi-scale guidance[C]//Proc. Eur. Conf. Comput. Vis., 2016: 353-369. [13] QIU Y, WANG R, TAO D, et al. Embedded block residual network: a recursive restoration model for single-image super - resolution[C]//Proceedings of the IEEE International Conference on Computer Vision, 2019: 4180-4189. [14] CAO Y, HE Z, YE Z, et al. Fast and accurate single image super - resolution via an energy-aware improved deep residual network[J]. Signal Processing, 2019, 162: 115-125. doi: 10.1016/j.sigpro.2019.03.018 [15] Anwar S, Barnes N. Densely residual Laplacian super-resolution[J/OL]. IEEE Transactions on Pattern Analysis and Machine Intelligence [2020-09-20](DOI: 10.1109/TPAMI.2020.3021088). [16] WANG Z, Chen J, Hoi S C H. Deep learning for image super-resolution: a survey[J/OL]. IEEE Trans Pattern Anal Mach Intell.[2020-09-20]. https://ieeexplore.ieee.org/document/9185010 (doi: 10.1109/TPAMI.2020.2982166). [17] LONG J, Shelhamer E, Darrell T. Fully convolutional networks for semantic segmentation[C]// Proc. Conf. Comput. Vis. Pattern Recognit, 2015: 3431-3440. [18] HE K, ZHANG X, REN S, et al. Delving deep into rectifiers: Surpassing human-level performance on image net classification[C]//Proc. Int. Conf. Comput. Vis., 2015: 1026-1034. [19] Lim B, Son S, Kim H, et al. Enhanced deep residual networks for single image super-resolution[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2017: 136-144. [20] HUI T W, Loy C C, TANG X. Depth map super-resolution by deep multi-scale guidance[C]//European Conference on Computer Vision, Springer, 2016: 353-369. [21] Guei A C, Akhloufi M. Deep learning enhancement of infrared face images using generative adversarial networks[J]. Applied Optics, 2018, 57(18): 98-107. doi: 10.1364/AO.57.000D98 [22] ZHANG X, LI C, MENG Q, et al. Infrared image super resolution by combining compressive sensing and deep learning[J]. Sensors, 2018, 18(8): 2587-2599. doi: 10.3390/s18082587