The blurred concrete crack images are typically the result of unexpected camera motion in engineering. Consequently, deblurring and segmentation of them represents a challenging task due to the complexity of the issue. This paper presents a novel Deep Multi-stage Network (DMNet), integrated with a multi-stage feature fusion strategy, for the purpose of deblurring crack images. The proposed model is compared to other models and found to be superior. Furthermore, a novel Discrete Wavelet Transform Network (DWTNet) which integrated with a multi-stage fusion approach is proposed to improve the semantic segmentation results of restored crack images. Two crack image datasets are employed to validate the proposed models and represent the typical blurring scenarios in engineering applications. The SSIM, Dice and IoU metrics are employed to provide a quantitative assessment of the proposed DMNet and DWTNet models’ capabilities in terms of image deblurring and segmentation. The values of these metrics exceed those of existing models, thereby demonstrating that the DMNet combined with the DWTNet is more effective than other state-of-the-art models. It exhibits superior performance in deblurring and segmentation of crack images, respectively.

中文翻译：

---

小波集成深度神经网络，用于裂纹图像的去模糊和分割


模糊的混凝土裂缝图像通常是工程中摄像机意外运动的结果。因此，由于问题的复杂性，对它们进行去模糊和分割是一项具有挑战性的任务。该文提出了一种新颖的深度多阶段网络 （DMNet），它与多阶段特征融合策略集成，用于去模糊裂纹图像。将所提出的模型与其他模型进行比较，发现其优越性。此外，该文提出一种新型的离散小波变换网络（DWTNet），该网络与多阶段融合方法相结合，以改善修复裂纹图像的语义分割结果。采用两个裂纹图像数据集来验证所提出的模型，并表示工程应用中的典型模糊场景。SSIM、Dice 和 IoU 指标用于对所提出的 DMNet 和 DWTNet 模型在图像去模糊和分割方面的能力进行定量评估。这些指标的值超过了现有模型的值，从而表明 DMNet 与 DWTNet 的结合比其他最先进的模型更有效。它分别在裂纹图像的去模糊和分割方面表现出卓越的性能。