摘要：【中文】本发明提供一种基于深度学习与声呐成像的水下桥墩构件分割方法，利用水下侧声声呐设备获取准备水下桥墩扫描图片；运用图像增强方法，增加数据集的数量；对数据集进行标注，将桥墩、桩基和河床分别用不同颜色进行多边形标记，并记录下多边形顶点坐标；将数据集划分为训练与验证集：建立深度学习语义分割网络中的Mask RCNN模型，进行训练，得到训练模型；水面上控制侧声声呐设备沿着水下桥墩部分进行扫描，获取扫描图片，利用训练好的Mask RCNN模型进行水下桥墩进行构件自动化分割。采用本发明效率高，成本低，相对于传统的人工潜水法以及声呐人工筛选法更具有明显的自动化和高效准确的优势。 【EN】The invention provides a method for dividing an underwater pier component based on deep learning and sonar imaging, which comprises the steps of obtaining a scanning picture of the underwater pier by using an underwater side sonar device; increasing the number of data sets by using an image enhancement method; labeling the data set, respectively carrying out polygon marking on the bridge pier, the pile foundation and the river bed by using different colors, and recording the coordinates of the vertex of the polygon; dividing the data set into training and verification sets: establishing a Mask RCNN model in a deep learning semantic segmentation network, and training to obtain a training model; the underwater side sonar equipment is controlled on the water surface to scan along the underwater bridge pier part, scanning pictures are obtained, and the trained Mask RCNN model is used for carrying out automatic segmentation on the underwater bridge pier. The method has the advantages of high efficiency, low cost and obvious automation, high efficiency and accuracy compared with the traditional manual diving method and sonar manual screening method.

摘要：【中文】本发明提供了一种基于深度学习的水下桥墩病害图像能见度增强方法，包括如下步骤：准备好低能见度水下桥墩照片以及高能见度水下桥墩照片，包括有病害与无病害图片；运用图像数据增强方法，扩充数据集的数量；将数据集划分为训练集与验证集，建立深度学习中的RetinexNet网络模型，将高能见度的照片作为标签，进行训练，得到训练模型；输入低能见度的水下病害照片，经过训练好的RetinexNet模型自动处理，即可得到较高能见度的水下病害照片。本发明效率高，成本低，相对于传统的图像亮度通道增强算法更具有明显的自动化优势。 【EN】The invention provides a method for enhancing visibility of an underwater pier disease image based on deep learning, which comprises the following steps: preparing low-visibility underwater pier photos and high-visibility underwater pier photos, including diseased and non-diseased pictures; expanding the number of data sets by using an image data enhancement method; dividing a data set into a training set and a verification set, establishing a RetinexNet network model in deep learning, and training by taking a high-visibility picture as a label to obtain a training model; and inputting the underwater disease photo with low visibility, and automatically processing the underwater disease photo by the trained RetinexNet model to obtain the underwater disease photo with high visibility. The method has high efficiency and low cost, and has obvious automation advantage compared with the traditional image brightness channel enhancement algorithm.

摘要：【中文】本发明提供了一种基于深度学习与声呐成像的水下桥墩病害识别方法，利用水下侧声声呐设备获取准备水下桥墩扫描图片，包括病害与无病害图片；运用图像增强方法，增加数据集的数量；对数据集进行标注，病害区域用矩形框标记并且保存坐标信息；将数据集划分为训练测试集、验证集和测试集；建立深度学习目标检测网络中的yolov3模型，进行训练，得到训练模型；水面上控制侧声声呐设备沿着水下桥墩部分进行扫描，获取扫描图片，利用训练好的yolov3模型进行水下桥墩病害自动识别。本发明效率高，成本低，相对于传统的人工潜水法以及声呐人工筛选法更具有明显的自动化和实时性优势。 【EN】The invention provides an underwater pier disease identification method based on deep learning and sonar imaging, which comprises the steps of obtaining and preparing underwater pier scanning pictures including disease and disease-free pictures by using an underwater side sonar device; increasing the number of data sets by using an image enhancement method; marking the data set, marking the disease area by using a rectangular frame and storing coordinate information; dividing a data set into a training test set, a verification set and a test set; establishing a yolov3 model in a deep learning target detection network, and training to obtain a training model; the control side sonar equipment on the surface of water scans along pier part under water, obtains the scanning picture, utilizes the good yolov3 model of training to carry out pier disease automatic identification under water. The invention has high efficiency and low cost, and has obvious advantages of automation and real-time performance compared with the traditional manual diving method and sonar manual screening method.

摘要：【中文】本发明公开了一种基于深度学习与声发射技术的拉索损伤识别方法，首先，利用声发射设备测量正常状态与损伤状态拉索的声发射信号各N组，形成数据集，将数据集中的信号序列打上标签，将N组信号归一化，并划分为训练集和测试集；建立长短期记忆全卷积神经网络，对声发射信号进行训练，保存训练好的模型和参数；利用声发射设备在现场获取拉索的声发射信号，将信号输入上述训练好的模型中，自动判别拉索是否存在损伤。本发明直接处理声发射的原始信号，效率高，成本低，相对于传统的人工检测法以及磁通量法更具有明显的自动化优势。 【EN】The invention discloses a inhaul cable damage identification method based on deep learning and acoustic emission technology, which comprises the steps of firstly, measuring N groups of acoustic emission signals of inhaul cables in normal states and damaged states by using acoustic emission equipment to form a data set, labeling a signal sequence in the data set, normalizing the N groups of signals, and dividing the signals into a training set and a testing set; establishing a long-short term memory full-convolution neural network, training an acoustic emission signal, and storing a trained model and parameters; and acquiring an acoustic emission signal of the inhaul cable on site by using an acoustic emission device, inputting the signal into the trained model, and automatically judging whether the inhaul cable is damaged or not. The method directly processes the original signal of the acoustic emission, has high efficiency and low cost, and has obvious automation advantages compared with the traditional manual detection method and the magnetic flux method.

摘要：【中文】本发明提供了一种基于深度学习的水下桥墩病害图像分辨率增强方法，包括如下步骤：准备好低分辨率水下桥墩照片以及高分辨率水下桥墩照片，包括有病害与无病害图片；运用图像数据增强方法，扩充数据集的数量；将数据集划分为训练集与验证集，建立深度学习中的WDSR网络模型，将高分辨率的照片作为标签，进行训练，得到训练模型；输入低分辨率的水下病害照片，经过训练好的WDSR模型自动处理，即可得到较高分辨率的水下病害照片。本发明效率高，成本低，相对于传统的图像分辨率增强算法更具有明显的自动化优势。 【EN】The invention provides a method for enhancing the resolution of an underwater pier disease image based on deep learning, which comprises the following steps: preparing low-resolution underwater pier photos and high-resolution underwater pier photos, including diseased and non-diseased pictures; expanding the number of data sets by using an image data enhancement method; dividing a data set into a training set and a verification set, establishing a WDSR network model in deep learning, and training by taking a high-resolution picture as a label to obtain a training model; and inputting the underwater disease photo with low resolution, and automatically processing the underwater disease photo by the trained WDSR model to obtain the underwater disease photo with higher resolution. The method has high efficiency and low cost, and has obvious automation advantage compared with the traditional image resolution enhancement algorithm.