Global-Local Attention-Guided Reconstruction Network for Infrared Image
-
摘要:
针对真实世界的红外图像恢复算法中存在的图像模糊、纹理失真、参数过大等问题,提出了一种用于真实红外图像的全局-局部注意力引导的超分辨率重建算法。首先,设计了一种跨尺度的全局-局部特征融合模块,利用多尺度卷积和Transformer并行融合不同尺度的信息,并通过可学习因子引导全局和局部信息的有效融合。其次,提出了一种新颖的退化算法,即域随机化退化算法,以适应真实红外场景图像的退化域。最后,设计了一种新的混合损失函数,利用权重学习和正则化惩罚来增强网络的恢复能力,同时加快收敛速度。在经典退化图像和真实场景红外图像上的测试结果表明,与现有方法相比,该算法恢复的图像纹理更逼真,边界伪影更少,同时参数总数最多可减少20%。
Abstract:To solve the problems of image blur smoothing, texture distortion, and excessively large parameters in real-world infrared-image recovery algorithms, a global-local attention-guided super-resolution reconstruction algorithm for infrared images is proposed. First, a cross-scale global-local feature fusion module utilizes multi-scale convolution and a transformer to fuse information at different scales in parallel and to guide the effective fusion of global and local information by learnable factors. Second, a novel domain randomization degradation model accommodates the degradation domain of real-world infrared images. Finally, a new hybrid loss based on weight learning and regularization penalty enhances the recovery capability of the network while speeding up convergence. Test results on classical degraded images and real-world infrared images show that, compared with existing methods, the images recovered by the proposed algorithm have more realistic textures and fewer boundary artifacts. Moreover, the total number of parameters can be reduced by up to 20%.
-
-
表 1 GFF模块数量设计的消融研究表
Table 1 Ablation study on GFF block number design
GFF block number 2 3 4 PSNR 32.60 32.68 32.64 SSIM 0.8999 0.9010 0.9011 表 2 权重因子的消融研究表
Table 2 Ablation study on weight factor
Weight Factor w0: w1=1 w0: w1=2 w0: w1=0.5 PSNR 32.15 32.68 32.04 SSIM 0.8887 0.9012 0.8786 表 3 损失函数的消融研究表
Table 3 Ablation study of the proposed hybrid loss
Index Loss1 Loss2 Loss3 Loss4 Loss5 L1 ✔ × × × ✔ Lp × ✔ × × ✔ Lg × × ✔ × ✔ Ld ✔ × × √ ✔ PSNR 32.62 32.58 32.53 32.51 32.68 SSIM 0.9000 0.8994 0.8985 08987 0.9011 表 4 基准数据集上双三次退化图像的超分辨率性能(PSNR/SSIM)与最新方法的定量比较
Table 4 Quantitative comparison of super-resolution performance (average PSNR/SSIM) with the state-of-the-art methods for bicubic degradation images on benchmark datasets
Method Scale Training dataset Set5[10] Set14[8] BSD100[7] Urban100[7] PSNR SSIM PSNR SSIM PSNR SSIM PSNR SSIM SRCNN[6] ×2 DIV2K 36.66 0.9542 32.45 0.9067 31.36 0.8879 29.50 0.8946 EDSR[9] ×2 DIV2K 38.11 0.9602 33.92 0.9195 32.32 0.9013 32.93 0.9773 RDN[14] ×2 DIV2K 38.24 0.9614 34.01 0.9212 32.34 0.9017 33.39 0.9353 RCAN[22] ×2 DIV2K 38.27 0.9614 34.12 0.9216 32.41 0.9027 33.34 0.9384 SAN[28] ×2 DIV2K 38.31 0.9620 34.07 0.9213 32.42 0.9028 33.10 0.9370 HAN[23] ×2 DIV2K 38.27 0.9614 34.16 0.9217 32.42 0.9027 33.35 0.9385 NLSA[2] ×2 DIV2K 38.34 0.9618 34.08 0.9231 32.43 0.9027 33.42 0.9394 GLAGSR (Ours) ×2 DIV2K 38.37 0.9616 34.17 0.9221 32.48 0.9029 33.49 0.9395 SRCNN[6] ×3 DIV2K 36.66 0.9542 32.45 0.9067 31.36 0.8879 29.50 0.8946 EDSR[9] ×3 DIV2K 34.76 0.9290 30.66 0.8481 29.32 0.8104 29.02 0.8685 RDN[14] ×3 DIV2K 34.58 0.9280 30.53 0.8447 29.23 0.8079 28.46 0.8582 GLAGSR (Ours) ×3 DIV2K 34.90 0.9314 30.80 0.8498 29.40 0.8130 29.55 0.8751 SRCNN[6] ×4 DIV2K 30.84 0.8628 27.50 0.7513 26.90 0.7101 24.52 0.7221 EDSR[9] ×4 DIV2K 32.46 0.8968 28.80 0.7876 27.71 0.7420 26.64 0.8033 RDN[14] ×4 DIV2K 32.47 0.8990 28.81 0.7871 27.72 0.7419 26.61 0.8028 RCAN[22] ×4 DIV2K 32.63 0.9002 28.87 0.7889 27.77 0.7436 26.82 0.8087 SAN[28] ×4 DIV2K 32.64 0.9003 28.92 0.7888 27.78 0.7436 26.79 0.8068 HAN[23] ×4 DIV2K 32.64 0.9002 28.90 0.7890 27.80 0.7442 26.85 0.8094 NLSA[2] ×4 DIV2K 32.59 0.9000 28.87 0.7891 27.78 0.7444 26.96 0.8109 GLAGSR (Ours) ×4 DIV2K 32.80 0.9029 29.03 0.7928 27.89 0.7461 27.02 0.8135 -
[1] 胡德敏, 闵天悦. 改进型轻量级GAN的红外图像超分辨率方法[J]. 小型微型计算机系统, 2022, 43(8): 1711-1717. https://www.cnki.com.cn/Article/CJFDTOTAL-XXWX202208021.htm HU Demin, MIN Tianyue. Infrared image super-resolution via improved lightweight GAN[J]. Journal of Chinese Computer Systems, 2022, 43(8): 1711-1717. https://www.cnki.com.cn/Article/CJFDTOTAL-XXWX202208021.htm
[2] MEI Y, FAN Y, ZHOU Y. Image super-resolution with non-local sparse attention[C]//Proc. CVPR, 2021: 3517-3526. DOI: 10.1109/CVPR46437.2021.00352.
[3] 马敏慧, 王红茹, 王佳. 基于改进的MSRCR-CLAHE融合的水下图像增强算法[J]. 红外技术, 2023, 45(1): 23-32. http://hwjs.nvir.cn/cn/article/id/69e5b90e-9c0c-43c6-b4e2-dedede3eb414 MA Minhui, WANG Hongru, WANG Jia. An underwater image enhancement algorithm based on improved MSRCR-CLAHE fusion[J]. Infrared Technology, 2023, 45(1): 23-32. http://hwjs.nvir.cn/cn/article/id/69e5b90e-9c0c-43c6-b4e2-dedede3eb414
[4] ZHANG D, SHAO J, LI X, et al. Remote sensing image super-resolution via mixed high-order attention network[J]. IEEE Trans. Geosci. Remote Sens. , 2021, 59(6): 5183-5196. DOI: 10.1109/TGRS.2020.3009918
[5] 刘云峰, 赵洪山, 杨晋彪, 等. 基于GNR先验的电力设备热成像超分辨率方法[J]. 红外技术, 2023, 45(1): 40-48. http://hwjs.nvir.cn/cn/article/id/3f88d6d0-ab5c-4cd7-999e-b50ffce93699 LIU Yunfeng, ZHAO Hongshan, YANG Jinbiao, et al. Super resolution method for power equipment infrared imaging based on gradient norm-ratio prior[J]. Infrared Technology, 2023, 45(1): 40-48. http://hwjs.nvir.cn/cn/article/id/3f88d6d0-ab5c-4cd7-999e-b50ffce93699
[6] DONG C, Loy C, HE K, et al. Learning a deep convolutional network for image super-resolution[C]//Proc. Computer Vision, ECCV, 2014: 184-199. DOI: 10.1007/978-3-319-10593-2_13.
[7] 聂丰英, 侯利霞, 万里勇. 自适应双边滤波与方向梯度的红外图像增强[J]. 红外技术, 2022, 44(12): 1309-1315. http://hwjs.nvir.cn/cn/article/id/8fbb7847-4550-463c-a220-9e97cd402968 NIE Fengying, HOU Lixia, WAN Liyong. Infrared image enhancement based on adaptive bilateral filtering and directional gradient[J]. Infrared Technology, 2022, 44(12): 1309-1315. http://hwjs.nvir.cn/cn/article/id/8fbb7847-4550-463c-a220-9e97cd402968
[8] Timofte R, Rothe R, Van Gool L. Seven ways to improve example-based single image super resolution[C]//Proc. CVPR, 2016: 1865-1873. DOI: 10.1109/CVPR.2016.206
[9] Lim B, Son S, Kim H, et al. Enhanced deep residual networks for single image super-resolution[C]//Proc. CVPRW, 2017: 136-144. Doi: 10.1109/CVPRW.2017.151
[10] Bevilacqua M, Roumy A, Guillemot C, et al. Low-complexity single-image super-resolution based on nonnegative neighbor embedding[C]// Proc. BMVC, 2012: 135-141. DOI: 10.5244/C.26.135.
[11] TANG Y, GONG W, CHEN X, et al. Deep inception-residual Laplacian pyramid networks for accurate single-image super-resolution[J]. IEEE Transactions on Neural Networks and Learning Systems, 2019, 31(5): 1514-1528.
[12] CAI J, ZENG H, YONG H, et al. Toward real-world single image super-resolution: a new benchmark and a new model[C]//Proc. ICCV, 2019: 3086-3095. DOI: 10.1109/ICCV.2019.00318.
[13] SHI W, Caballero J, Huszár F, et al. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network[C]//Proc. CVPR, 2016: 1874-1883. DOI: 10.1109/CVPR.2016.207.
[14] ZHANG Y, TIAN Y, KONG Y, et al. Residual dense network for image super-resolution[C]//Proc. CVPR, 2018: 2472-2481. DOI: 10.1109/CVPR.2018.00262.
[15] Esser P, Rombach R, Ommer B. Taming transformers for high-resolution image synthesis[C]//Proc. CVPR, 2021: 12873-12883. DOI: 10.1109/CVPR46437.2021.01268.
[16] JIANG Y, CHANG S, WANG Z. Transgan: two transformers can make one strong GAN[J/OL]. Computer Vision and Pattern Recognition, 2021, 34: https://arxiv.org/abs/2102.07074.
[17] ZHANG K, LIANG J, Van Gool L, et al. Designing a practical degradation model for deep blind image super-resolution[C]//Proc. ICCV, 2021: 4771-4780. DOI: 10.1109/ICCV48922.2021.00475.
[18] WANG X, XIE L, DONG C, et al. Real-ESRGAN: Training real-world blind super-resolution with pure synthetic data[C]//Proc. ICCVW, 2021: 1905-1914. DOI: 10.1109/ICCVW54120.2021.00217
[19] Ledig C, Theis L, Huszar F, et al. Photo-realistic single image super-resolution using a generative adversarial network[C]//Proc. CVPR, 2017: 4681-4690. DOI: 10.1109/CVPR.2017.19.
[20] ZHOU Y, WU G, FU Y, et al. Cross-MPI: cross-scale stereo for image super-resolution using multiplane images[C]//Proc. CVPR, 2021: 14842-14851. DOI: 10.1109/CVPR46437.2021.01460.
[21] 烟台艾睿光电科技有限公司. 艾睿光电红外开源数据库[EB/OL]. [2023-02-26]. http://openai.raytrontek.com/apply/Super_resolution.html/. IRay Technology Co., Ltd. IRay Optoelectronic Infrared Open Source Database[EB/OL]. [2023-02-26]. http://openai.raytrontek.com/apply/Super_resolution.html/.
[22] ZHANG Y, LI K, LI K, et al. Image super-resolution using very deep residual channel attention networks[C]//Proc. ECCV, 2018: 286-301. DOI: 10.1007/978-3-030-01234-2_18.
[23] NIU B, WEN W, REN W, et al. Single image super-resolution via a holistic attention network[C]//Proc. ECCV, 2020, 12357: 191-207.
[24] WANG X, YU K, WU S, et al. ESRGAN: Enhanced super-resolution generative adversarial networks[C]// Proc. ECCV, 2019: 63-79. DOI: 10.1007/978-3-030-11021-5_5.
[25] LIU Z, LIN Y, CAO Y, et al. Swin transformer: Hierarchical vision transformer using shifted windows[C]//Proc. ICCV, 2021: 9992-10002. DOI: 10.1109/ICCV48922.2021.00986.
[26] WANG Y, WANG L, WANG H, et al. Resolution-aware network for image super-resolution[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2018, 29(5): 1259-1269.
[27] Andreas Lugmayr, Martin Danelljan, Radu Timofte. Unsupervised learning for real-world super-resolution[J]. ICCV Workshop, 2019, 2(3): 3408-3416.
[28] DAI T, CAI J, ZHANG Y, et al. Second-order attention network for single image super-resolution[C]//Proc. CVPR, 2019: 11057-11066. DOI: 10.1109/CVPR.2019.01132.
[29] Agustsson E, Timofte R. Ntire 2017 challenge on single image super-resolution: Dataset and study[C]// Proc. CVPRW, 2017: 126-135. DOI: 10.1109/CVPRW.2017.150.
[30] LIANG J, CAO J, SUN G, et al. SwinIR: Image restoration using swin transformer[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021: 1833-1844. DOI: 10.1109/ICCVW54120.2021.00210.
[31] Blau Y, Mechrez R, Timofte R, et al. The 2018 PIRM challenge on perceptual image super-resolution[C]//Proceedings of the European Conference on Computer Vision (ECCV), 2018, DOI: 10.1007/978-3-030-11021-5_21
[32] MA C, YANG C Y, YANG X, et al. Learning a no-reference quality metric for single-image super-resolution[J]. Computer Vision and Image Understanding, 2017, 158: 1-16. DOI: 10.1016/j.cviu.2016.12.009
-
期刊类型引用(3)
1. 吴丹,张艳. 多特征匹配的高光谱图像配准方法. 现代电子技术. 2021(21): 56-59 . 百度学术
2. 史国川,拓浩男. SIFT和SURF算法在偏振图像匹配中的应用研究. 舰船电子工程. 2020(10): 93-95+148 . 百度学术
3. 刘丁瑜,易加维,张徐洲,张畅,刘飞,方慧,何勇. 单子叶植物叶片双向反射分布的测量与分析. 光谱学与光谱分析. 2019(07): 2100-2106 . 百度学术
其他类型引用(5)