基于多尺度特征融合的电子支气管镜暗区无监督视觉增强算法

Unsupervised visual enhancement for dark regions in electronic bronchoscopic images based on multi-scale feature fusion

摘要

目的：探讨基于深度学习的视觉增强网络在电子支气管镜图像检测价值。方法：纳入600例术后检测患者基于支气管镜检测图像制作数据集，向多尺度视觉增强网络中分别引用基于Haar离散小波变换下采样、Vision Transformer、空间残差增强模块和高低频特征合融合模块来增强网络对图像中暗区域的恢复。与EnGAN和URetinex模型进行定性定量对比实验。结果：本文模型可以有效恢复图像暗区域。其在数据集中NIQE和BTMQI指标中效果最优，相比于原图NIQE降低13.01%，BTMQI提升11.62%。BRISQUE指标效果也极其优秀，相比于原图降低9.15%。通过定性对比实验，本文模型可视化效果强于其他模型。结论：本文模型可提升电子支气管镜图像暗区域的恢复效果。

Abstract

Objective To explore the potential of a deep learning-based visual enhancement network in the analysis of electronic bronchoscopic images. Methods A dataset was constructed using bronchoscopic images acquired from 600 postoperative patients. A multi-scale visual enhancement network was developed to improve the restoration of dark regions through Haar discrete wavelet transform-based downsampling, Vision Transformer, spatial residual enhancement module, and high- and low-frequency feature fusion module. Qualitative and quantitative comparisons with EnGAN and URetinex models were performed. Results The proposed model effectively restored dark regions in bronchoscopic images, and achieved optimal NIQE and BTMQI on the dataset, with a 13.01% reduction in NIQE and an 11.62% increase in BTMQI compared with the original images. Additionally, it obtained a favorable BRISQUE value which was decreased by 9.15% relative to the original images. Qualitative comparisons further demonstrated its superior visual performance over other models. Conclusion The proposed model can improve the restoration of dark regions in electronic bronchoscopic images.

关键词

深度学习 / 离散小波变换 / 电子支气管镜 / 视觉增强

Key words

deep learning / discrete wavelet transform / electronic bronchoscope / visual enhancement

参考文献

[1] Yan PF, Sun WL, Li X, et al. PKDN: prior knowledge distillation network for bronchoscopy diagnosis[J]. Comput Biol Med, 2023, 166: 107486.

[2] 苏德尔, 高伟达, 王新伟, 等 . 微型化生物医学光学成像系统[J]. 光电工程, 2024, 51(12): 1-27.

[3] Su DE, Gao WD, Wang XW, et al. Miniaturized biomedical optical imaging systems[J]. Opto—Electronic Engineering, 2024, 51(12): 1-27.

[4] 薛丽荣, 王中秋, 李姝, 等 . 仿真滤线栅及图像增强后处理算法用于改善儿童低剂量X线片图像质量[J]. 中国介入影像与治疗学, 2024, 21(2): 105-109.

[5] Xue LR, Wang ZQ, Li S, et al. Sim grid and S—Enhance post processing algorithm for improving image quality of low—dose X—ray films in children[J]. Chinese Journal of Interventional Imaging and Therapy, 2024, 21(2): 105-109.

[6] 侯园园, 周萍 . 基于小波变换的数字胸片增强[J]. 中国医学影像技术, 2010, 26(10): 1976-1979.

[7] Hou YY, Zhou P . Approach on digital chest radiographs enhancement based on wavelet transform[J]. Chinese Journal of Medical Imaging Technology, 2010, 26(10): 1976-1979.

[8] 张文瀚, 王永雄, 曾福斌, 等 . 基于前馈注意力ConvNeXt模型分割腹部CT中的胰腺[J]. 中国医学影像技术, 2025, 41(3): 466-472.

[9] Zhang WH, Wang YX, Zeng FB, et al. Segmentation model of pancreas from abdominal CT based on feedforward attention ConvNeXt[J]. Chinese Journal of Medical Imaging Technology, 2025, 41(3): 466-472.

[10] Zhang HX, Ran T, Xiao WD, et al. Multi—scale progressive fusion network for low—light image enhancement[J]. IEEE Trans Instrum Meas, 2025, 74: 1-11.

[11] Wang LW, Liu ZS, Siu WC, et al. Lightening network for low—light image enhancement[J]. IEEE Trans Image Process, 2020, 29: 7984-7996.

[12] He KM, Zhang XY, Ren SQ, et al. Delving deep into rectifiers: surpassing human—level performance on ImageNet classification[C]// 2015 IEEE International Conference on Computer Vision (ICCV). Piscataway, NJ, USA: IEEE, 2015: 1026-1034.

[13] 李印, 左志超, 金观桥, 等 . 基于小波变换的PET/CT图像融合算法研究进展[J]. 中国医学影像技术, 2018, 34(8): 1267-1270.

[14] Li Y, Zuo ZC, Jin GQ, et al. Research progresses of PET/CT image fusion algorithm based on wavelet transform[J]. Chinese Journal of Medical Imaging Technology, 2018, 34(8): 1267-1270.

[15] 林燕凤, 赵舷宏, 付朝丽, 等 . 以人工智能肠道图像识别模型评估结肠镜检查前肠道准备[J]. 中国医学影像技术, 2023, 39(7): 1034-1038.

[16] Lin YF, Zhao XH, Fu ZL, et al. Artificial intelligence colonic image recognition model for evaluating bowel preparation before colonoscopy[J]. Chinese Journal of Medical Imaging Technology, 2023, 39(7): 1034-1038.

[17] He KM, Zhang XY, Ren SQ, et al. Deep residual learning for image recognition[C]// 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway, NJ, USA: IEEE, 2016: 770-778.

[18] Zhao H, Gallo O, Frosio I, et al. Loss functions for image restoration with neural networks[J]. IEEE Trans Comput Imaging, 2017, 3(1): 47-57.

[19] Eigen D, Puhrsch C, Fergus R . Depth map prediction from a single image using a multi—scale deep network[C]// Proceedings of the 28th International Conference on Neural Information Processing Systems. Cambridge, MA, USA: MIT Press, 2014: 2366-2374.

[20] Mittal A, Soundararajan R, Bovik AC . Making a "completely blind" image quality analyzer[J]. IEEE Signal Process Lett, 2013, 20(3): 209-212.

[21] Chow LS, Rajagopal H . Modified—BRISQUE as no reference image quality assessment for structural MR images[J]. Magn Reson Imaging, 2017, 43: 74-87.

[22] Patle MK, Chourasia B, Kurmi Y . High dynamic range image analysis through various tone mapping techniques[J]. Int J Comput Appl, 2016, 153(11): 14-17.

[23] Jiang YF, Gong XY, Liu D, et al. EnlightenGAN: deep light enhancement without paired supervision[J]. IEEE Trans Image Process, 2021, 30: 2340-2349.

[24] Wu WH, Weng J, Zhang PP, et al. URetinex—Net: retinex—based deep unfolding network for low—light image enhancement[C]// 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway, NJ, USA: IEEE, 2022: 5891-5900.

[25] Wu JX, Ran T, Xiao WD, et al. Multi—scale cascaded network with high—low frequency for low—light image enhancement[J]. Comput Graph, 2025, 132: 104380.

[26] 张甲杰, 陈树越, 刘瑞剑, 等 . 基于多尺度Retinex的医学图像增强方法研究[J]. 中国医学影像技术, 2007, 23(11): 1724-1726.

[27] Zhang JJ, Chen SY, Liu RJ, et al. Enhancement of medical images based multi—scale Retinex[J]. Chinese Journal of Medical Imaging Technology, 2007, 23(11): 1724-1726.

[28] Hosseinpour S, Sharafati A, Abghari H . Downscaling of two selected GCM data using a hybrid deep learning method of wavelet—CNN—LSTM in Iran[J]. Theor Appl Climatol, 2025, 156(9): 459.

[29] Pei X, Huang YD, Su WJ, et al. FFTFormer: a spatial—frequency noise aware CNN—transformer for low light image enhancement[J]. Knowl Based Syst, 2025, 314: 113055.

[30] Liu Q, Huang W, Hu T, et al. Efficient network architecture for target detection in challenging low—light environments[J]. Eng Appl Artif Intell, 2025, 142: 109967.

[31] Rajapaksha U, Sohel F, Laga H, et al. Deep learning—based depth estimation methods from monocular image and videos: a comprehensive survey[J]. ACM Comput Surv, 2024, 56(12): 315.

[32] Hui YM, Wang J, Li B . WSA—YOLO: weak—supervised and adaptive object detection in the low—light environment for YOLOV7[J]. IEEE Trans Instrum Meas, 2024, 73: 1-12.

[33] Bai KP, Li ZP, Wang LJ, et al. The low—illumination 3D reconstruction method based on neural radiation field[J]. J Radiat Res Appl Sci, 2025, 18(2): 101488.

[34] Li JL, Li BL, Tu ZZ, et al. Light the night: a multi—condition diffusion framework for unpaired low—light enhancement in autonomous driving[C]// 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway, NJ, USA: IEEE, 2024: 15205-15215.

基金资助

新疆维吾尔自治区自然科学基金青年科学基金(2024D01C291)