技术领域Technical Field
本发明涉及下水管道缺陷识别技术领域，尤其是指一种下水管道内部缺陷识别方法和系统。The present invention relates to the technical field of sewer pipe defect recognition, and in particular to a method and system for recognizing internal defects of a sewer pipe.
背景技术Background Art
地下管道系统是城市基础设施的关键组成部分。随着管道的老化，将逐渐出现各种缺陷。如管道淤积、破裂、泄漏、变形，甚至崩塌，从而导致安全问题。由于管道网络的隐蔽性，城市管理部门经常难以及时检测和解决管道缺陷。因此，定期检查地下管道已成为城市安全监管的重要任务。Underground pipeline systems are a key component of urban infrastructure. As pipelines age, various defects will gradually appear. Such as pipeline siltation, rupture, leakage, deformation, and even collapse, leading to safety issues. Due to the hidden nature of the pipeline network, urban management departments often find it difficult to detect and resolve pipeline defects in a timely manner. Therefore, regular inspections of underground pipelines have become an important task for urban safety supervision.
市政排水系统中使用的管道的内径范围通常为0.1到1米，并且可能含有有毒气体，检查人员无法直接进入这些管道。随着传感器和计算机视觉的不断发展，出现了较为先进的下水道内部缺陷检测技术，如超声波检测、激光扫描、雷达检测、磁通漏技术和闭路电视检测。其中，闭路电视检测主要利用微型机器人携带的摄像头实时获取地下管道的内部信息。工作人员可以通过摄像头的反馈控制机器人的前进方向，并分析污水管道缺陷的类型、位置和程度，进行定性判断。与其他技术相比，闭路电视检测更经济、方便和可视化，因此逐渐成为主流的管道检查方法。然而，它仍然需要技术人员基于视频进行相对主观的手动分析，自动化程度不高，效率较低。因此，迫切需要智能检测方法来识别污水管道损伤的像素大小、位置和类型，为进一步的维护操作提供定量指标。The inner diameter of pipes used in municipal drainage systems usually ranges from 0.1 to 1 meter, and may contain toxic gases, which cannot be directly accessed by inspectors. With the continuous development of sensors and computer vision, more advanced sewer internal defect detection technologies have emerged, such as ultrasonic detection, laser scanning, radar detection, flux leakage technology, and CCTV detection. Among them, CCTV detection mainly uses the camera carried by the micro robot to obtain the internal information of the underground pipeline in real time. The staff can control the robot's forward direction through the feedback of the camera, and analyze the type, location and degree of sewage pipe defects to make qualitative judgments. Compared with other technologies, CCTV detection is more economical, convenient and visual, so it has gradually become the mainstream pipeline inspection method. However, it still requires technicians to perform relatively subjective manual analysis based on videos, with low automation and low efficiency. Therefore, there is an urgent need for intelligent detection methods to identify the pixel size, location and type of sewage pipe damage to provide quantitative indicators for further maintenance operations.
为了解决这一挑战，基于机器学习的缺陷检测技术被应用于下水管道状况诊断场景。然而，这些方法存在固有的缺点：(1)基于滑动窗口的区域选择算法具有较高的时间复杂度和许多冗余窗口；(2)手动设计的特征提取器对目标对象的多样性鲁棒性差。基于卷积神经网络的目标检测框架近年来在计算机视觉研究中取得了显著进展。这些方法可分类为两个主要类别：两阶段网络(R-CNN系列)和一阶段网络(单镜头多帧检测器(SSD)和YouOnly Look Once(YOLO)系列)。一阶段网络，如YOLO，已经表现出更平衡的性能，可以直接在图像中检测对象边界框和类别，无需进行候选区域提取。To address this challenge, machine learning-based defect detection techniques have been applied to sewer condition diagnosis scenarios. However, these methods have inherent disadvantages: (1) sliding window-based region selection algorithms have high time complexity and many redundant windows; (2) manually designed feature extractors are not robust to the diversity of target objects. Object detection frameworks based on convolutional neural networks have made significant progress in computer vision research in recent years. These methods can be categorized into two main categories: two-stage networks (R-CNN series) and one-stage networks (Single Shot Multi-Frame Detector (SSD) and YouOnly Look Once (YOLO) series). One-stage networks, such as YOLO, have shown more balanced performance and can directly detect object bounding boxes and categories in images without the need for candidate region extraction.
但是基于机器学习的缺陷检测技术检测准确性并不高，需要提高缺陷检测的准确性。However, the detection accuracy of defect detection technology based on machine learning is not high, and the accuracy of defect detection needs to be improved.
发明内容Summary of the invention
为此，本发明所要解决的技术问题在于克服现有技术中缺陷检测技术检测准确性并不高的问题。Therefore, the technical problem to be solved by the present invention is to overcome the problem that the defect detection technology in the prior art has low detection accuracy.
为解决上述技术问题，本发明提供了一种下水管道内部缺陷识别方法，包括：In order to solve the above technical problems, the present invention provides a method for identifying internal defects of a sewer pipe, comprising:
步骤S1：采集管道内部图像，进行数据清洗并筛选出有效的管道缺陷图像，按照预先设置的管道缺陷类型对有效的管道缺陷图像进行标注，得到标注好的数据集；Step S1: Collect images of the interior of the pipeline, perform data cleaning and screen out valid pipeline defect images, and annotate the valid pipeline defect images according to the preset pipeline defect types to obtain an annotated data set;
步骤S2：对标注好的数据集进行预处理；Step S2: preprocess the labeled data set;
步骤S3：利用预处理后的数据集对改进的YOLOv5模型进行训练，得到训练好的YOLOv5模型；Step S3: Using the preprocessed data set to train the improved YOLOv5 model to obtain a trained YOLOv5 model;
步骤S4：通过所述训练好的YOLOv5模型对待检测的管道内部图像进行识别，得到识别结果。Step S4: The trained YOLOv5 model is used to identify the internal image of the pipeline to be detected to obtain a recognition result.
在本发明的一个实施例中，所述步骤S2中对标注好的数据集进行预处理的方法包括：In one embodiment of the present invention, the method of preprocessing the labeled data set in step S2 includes:
将管道缺陷图像随机裁剪一块区域，以适应不同角度下拍摄的图片，并随机调整图像的色调、饱和度和明度；Randomly crop an area of the pipeline defect image to adapt to the pictures taken at different angles, and randomly adjust the hue, saturation and brightness of the image;
对管道缺陷图像进行Mosaic数据增强。Mosaic data enhancement is performed on pipeline defect images.
在本发明的一个实施例中，所述步骤S3中改进的YOLOv5模型的构建方法包括：In one embodiment of the present invention, the method for constructing the improved YOLOv5 model in step S3 includes:
YOLOv5模型包括Backbone主干网络、Neck网络和Head网络，所述Backbone主干网络用于提取图像特征，所述Neck网络用于对Backbone主干网络提取的图像特征进行多尺度融合，所述Head网络用于对Neck网络得到的多尺度融合后的特征进行预测；The YOLOv5 model includes a Backbone network, a Neck network and a Head network. The Backbone network is used to extract image features, the Neck network is used to perform multi-scale fusion of the image features extracted by the Backbone network, and the Head network is used to predict the multi-scale fused features obtained by the Neck network.
改进的YOLOv5模型包括：The improved YOLOv5 model includes:
在YOLOv5模型的Backbone主干网络中引入SK注意力机制；Introduce the SK attention mechanism into the Backbone network of the YOLOv5 model;
在YOLOv5模型的Neck网络中使用渐进特征金字塔结构AFPN代替FPN，改进特征融合方式；In the Neck network of the YOLOv5 model, the progressive feature pyramid structure AFPN is used instead of FPN to improve the feature fusion method;
将YOLOv5模型在训练时的损失函数替换为Scylla-IoU损失函数。Replace the loss function of the YOLOv5 model during training with the Scylla-IoU loss function.
在本发明的一个实施例中，所述SK注意力机制是一个动态选择性注意力模块，用于根据输入信息的多个尺度自适应调整其感受野的大小；在该动态选择性注意力模块中，使用softmax算法对由不同内核大小的多个分支的信息进行融合，获得各分支的关注权重；对各分支的不同关注权重导致融合层神经元有效感受野的大小不同；为了使神经元自适应调整其感受野大小，构建一种自动选择操作，即在不同内核大小的多个核之间进行SK卷积，包括：通过三个操作符，拆分、融合和选择来实现SK卷积；In one embodiment of the present invention, the SK attention mechanism is a dynamic selective attention module, which is used to adaptively adjust the size of its receptive field according to multiple scales of input information; in the dynamic selective attention module, a softmax algorithm is used to fuse information from multiple branches with different kernel sizes to obtain the attention weight of each branch; different attention weights for each branch result in different sizes of effective receptive fields of neurons in the fusion layer; in order to make the neurons adaptively adjust the size of their receptive fields, an automatic selection operation is constructed, that is, SK convolution is performed between multiple kernels with different kernel sizes, including: implementing SK convolution through three operators, splitting, fusion and selection;
所述拆分具体为：对于任何给定特征映射X∈RH'×W‘×C'进行4次转换，分别为表示对X采用卷积核大小为1×1的卷积操作，表示对X采用卷积核大小为3×3的卷积操作，表示对X采用卷积核大小为3×3和膨胀因子为2的空洞卷积操作，表示对X采用卷积核大小为3×3和膨胀因子为3的空洞卷积操作，R表示所有特征映射的集合，H表示特征图的高度，W表示特征图的宽度，C表示通道数；关于的卷积核分别采用1×1，3×3，5×5，7×7；将5×5和7×7的卷积核分别换成3×3的内核和膨胀因子为2和3的空洞卷积；F由分组/深度卷积、批量归一化和ReLU组成；The split is specifically as follows: for any given feature map X∈RH'×W'×C' , four transformations are performed, namely Indicates that a convolution operation with a convolution kernel size of 1×1 is applied to X. Indicates that a convolution operation with a convolution kernel size of 3×3 is applied to X. Indicates that a dilated convolution operation with a kernel size of 3×3 and a dilation factor of 2 is applied to X. Indicates that a dilated convolution operation with a kernel size of 3×3 and a dilation factor of 3 is applied to X, R represents the set of all feature maps, H represents the height of the feature map, W represents the width of the feature map, and C represents the number of channels; The convolution kernels of are 1×1, 3×3, 5×5, and 7×7 respectively; the 5×5 and 7×7 convolution kernels are replaced with 3×3 kernels and dilation factors of 2 and 3 respectively; F consists of grouped/depth convolution, batch normalization, and ReLU;
所述融合具体为：为了使神经元根据输入内容自适应地调整感受野大小，首先通过元素求和融合多个分支结果：使用全局平均池化来嵌入全局信息以生成通道统计信息：Fgp()表示全局平均池化函数，Uc表示U的二维形式，Uc(i,j)表示U的第c个通道中第[i，j]个元素值；通过一个全连接层对通道统计信息sc进行处理得到特征描述因子z∈Rd×1，实现精确和自适应选择；其中，特征描述因子具体计算公式为：其中，Ffc()表示全连接层操作，s为sc的三维形式，δ为ReLU函数，为批量归一化操作，W∈Rd×C，d＝max(C/r,L)，L为d的最小值，r为一个可变系数；The fusion is specifically as follows: in order to make the neuron adaptively adjust the receptive field size according to the input content, firstly, multiple branch results are fused by element summation: Use global average pooling to embed global information to generate channel statistics: Fgp () represents the global average pooling function, Uc represents the two-dimensional form of U, and Uc (i, j) represents the value of the [i, j]th element in the cth channel of U. The channel statistical informationsc is processed through a fully connected layer to obtain the feature description factor z∈Rd×1 to achieve accurate and adaptive selection. The specific calculation formula of the feature description factor is: Where Ffc () represents the fully connected layer operation, s is the three-dimensional form ofsc , δ is the ReLU function, It is a batch normalization operation, W∈Rd×C , d＝max(C/r,L), L is the minimum value of d, and r is a variable coefficient;
所述选择具体为：在特征描述因子z的引导下，采用softmax算法和跨通道的软注意力自适应选择不同空间尺度的信息，公式为：The selection is specifically as follows: under the guidance of the feature description factor z, the softmax algorithm and cross-channel soft attention are used to adaptively select information of different spatial scales, and the formula is:
ac+bc+cc+dc＝1；Ac,Bc,Cc,Dc∈RC×d，ac，bc，cc和dc分别表示的软注意力向量；通过各特征上的关注权重，得到最终的特征图V：V＝[V1,V2,...,Vc]，Vc∈RH×W。 ac +bc +cc +dc =1; Ac ,Bc ,Cc ,Dc ∈RC×d , ac ,bc ,cc and dc represent The soft attention vector of ; through the attention weights on each feature, the final feature map V is obtained: V=[V1 , V2 ,...,Vc ], Vc ∈RH×W .
在本发明的一个实施例中，所述在YOLOv5模型的Neck网络中使用渐进特征金字塔结构AFPN代替FPN，具体为：In one embodiment of the present invention, the progressive feature pyramid structure AFPN is used instead of FPN in the Neck network of the YOLOv5 model, specifically:
所述渐进特征金字塔结构AFPN用于避免信息丢失，在AFPN中，将从Backbone主干网络的每个特征层中提取的不同尺度特征表示为{C3，C4，C5}，采用ASFF方法，首先将底层特征C3和C4融合到特征金字塔中，再加入C5，表达式定义为：其中，将C3、C4和C5的第i行j列特征元素值分别描述为和表示从s层到t层的特征图；融合过程中，为了减小非相邻层次特征之间的语义差距，采用自适应空间融合操作，使用softmax算法计算出α，β、γ的值：以α为例，其中，分别是对通过1×1的卷积核卷积得到；经过以上处理，最终得到一组关于α，β、γ的以不同权重融合了{C3，C4，C5}的多尺度特征{P3，P4，P5}。The progressive feature pyramid structure AFPN is used to avoid information loss. In AFPN, the different scale features extracted from each feature layer of the Backbone network are represented as {C3 , C4 , C5 }. Using the ASFF method, the bottom-level features C3 and C4 are first fused into the feature pyramid, and then C5 is added. The expression is defined as: Among them, the characteristic element values of the i-th row and j-th column of C3, C4 and C5 are described as and Represents the feature map from layer s to layer t; in the fusion process, in order to reduce the semantic gap between non-adjacent layer features, an adaptive spatial fusion operation is used, and the softmax algorithm is used to calculate the values of α, β, and γ: Take α as an example, in, Respectively It is obtained by convolution with a 1×1 convolution kernel; after the above processing, a set of multi-scale features {P3 , P4 , P5 } about α, β, and γ are finally obtained, which are fused with {C3 , C4 , C5 } with different weights.
在本发明的一个实施例中，所述Scylla-IoU损失函数包括角度损失、距离损失、形状损失和IoU损失，其中，In one embodiment of the present invention, the Scylla-IoU loss function includes angle loss, distance loss, shape loss and IoU loss, wherein,
所述角度损失考虑了真实框与预测框向量之间的角度，定义为：The angle loss takes into account the angle between the real box and the predicted box vector and is defined as:
式中，Λ为角度损失，cw和ch为预测框与真实框中心点的水平距离和垂直距离，α为两中心点连线与水平线之间的夹角，σ为两点的欧氏距离；Where Λ is the angle loss,cw andch are the horizontal and vertical distances between the center points of the predicted box and the real box, α is the angle between the line connecting the two center points and the horizontal line, and σ is the Euclidean distance between the two points.
所述距离损失考虑了真实框与预测框向量之间的距离，定义为：The distance loss takes into account the distance between the real box and the predicted box vector and is defined as:
γ＝2-Λγ＝2-Λ
式中，Δ为距离损失，cw和ch为真实框和预测框最小外接矩形的宽和高，γ为与角度损失相关的系数；Where Δ is the distance loss,cw andch are the width and height of the minimum bounding rectangle of the real box and the predicted box, and γ is the coefficient related to the angle loss;
所述形状损失考虑了地面真实值与预测框之间的形状差，其定义为：The shape loss takes into account the shape difference between the ground truth and the predicted box, and is defined as:
式中，Ω为形状损失，w，h，wgt和hgt分别为预测框和真实框的宽和高，θ控制对形状损失的关注程度；Where Ω is the shape loss, w, h, wgt and hgt are the width and height of the predicted box and the true box respectively, and θ controls the degree of attention to the shape loss;
Scylla-IoU损失函数定义如下：The Scylla-IoU loss function is defined as follows:
在本发明的一个实施例中，所述步骤S3中利用预处理后的数据集对改进的YOLOv5模型进行训练的方法为：将输入的图像大小设置为640×640，使用SGD优化器进行优化，批处理大小设置为32，训练迭代轮数为400，采用余弦退火学习速率策略，学习率动量为0.937，权值衰减为0.0005，初始学习速率为0.01。In one embodiment of the present invention, the method for training the improved YOLOv5 model using the preprocessed data set in step S3 is: setting the input image size to 640×640, using the SGD optimizer for optimization, setting the batch size to 32, the number of training iterations to 400, using the cosine annealing learning rate strategy, the learning rate momentum to 0.937, the weight decay to 0.0005, and the initial learning rate to 0.01.
为解决上述技术问题，本发明提供了一种下水管道内部缺陷识别系统，包括：In order to solve the above technical problems, the present invention provides a sewer pipe internal defect identification system, comprising:
采集模块：用于采集管道内部图像，进行数据清洗并筛选出有效的管道缺陷图像，按照预先设置的管道缺陷类型对有效的管道缺陷图像进行标注，得到标注好的数据集；Acquisition module: used to collect images inside the pipeline, perform data cleaning and filter out valid pipeline defect images, annotate valid pipeline defect images according to the preset pipeline defect types, and obtain annotated data sets;
预处理模块：用于对标注好的数据集进行预处理；Preprocessing module: used to preprocess the labeled data set;
训练模块：用于利用预处理后的数据集对改进的YOLOv5模型进行训练，得到训练好的YOLOv5模型；Training module: used to train the improved YOLOv5 model using the preprocessed data set to obtain a trained YOLOv5 model;
识别模块：用于通过所述训练好的YOLOv5模型对待检测的管道内部图像进行识别，得到识别结果。Recognition module: used to recognize the internal image of the pipeline to be detected through the trained YOLOv5 model to obtain a recognition result.
为解决上述技术问题，本发明提供了一种电子设备，包括存储器、处理器及存储在存储器上并可在处理器上运行的计算机程序，所述处理器执行所述计算机程序时实现如上述下水管道内部缺陷识别方法的步骤。To solve the above technical problems, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above-mentioned method for identifying internal defects of sewer pipes when executing the computer program.
为解决上述技术问题，本发明提供了一种计算机可读存储介质，其上存储有计算机程序，所述计算机程序被处理器执行时，实现如上述下水管道内部缺陷识别方法的步骤。In order to solve the above technical problems, the present invention provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the above-mentioned method for identifying internal defects of sewer pipes are implemented.
本发明的上述技术方案相比现有技术具有以下优点：The above technical solution of the present invention has the following advantages compared with the prior art:
本发明使用改进的YOLOv5算法，在背景复杂、内部情况多变的情况下能够对下水管道缺陷进行准确识别，具有良好的检测识别效果，提升对城市下水管道缺陷识别的精度，提高缺陷预警的自动化水平。The present invention uses an improved YOLOv5 algorithm, which can accurately identify sewer pipe defects under complex backgrounds and changeable internal conditions. It has good detection and recognition effects, improves the accuracy of urban sewer pipe defect recognition, and improves the automation level of defect warning.
附图说明BRIEF DESCRIPTION OF THE DRAWINGS
为了使本发明的内容更容易被清楚的理解，下面根据本发明的具体实施例并结合附图，对本发明作进一步详细的说明。In order to make the contents of the present invention more clearly understood, the present invention is further described in detail below based on specific embodiments of the present invention in conjunction with the accompanying drawings.
图1是本发明的方法流程图；Fig. 1 is a flow chart of the method of the present invention;
图2是本发明实施例中改进的YOLOv5网络结构图；FIG2 is a diagram of an improved YOLOv5 network structure in an embodiment of the present invention;
图3是本发明实施例中SK注意力机制网络结构图；FIG3 is a network structure diagram of the SK attention mechanism according to an embodiment of the present invention;
图4是本发明实施例中渐近特征金字塔网络AFPN结构图；FIG4 is a structural diagram of an asymptotic feature pyramid network AFPN according to an embodiment of the present invention;
图5是本发明实施例中自适应空间融合操作图；FIG5 is a diagram of an adaptive spatial fusion operation in an embodiment of the present invention;
图6是本发明实施例中基于YOLOv5的下水管道内部缺陷识别方法检测结果对比图。FIG6 is a comparison chart of detection results of the sewer pipe internal defect recognition method based on YOLOv5 in an embodiment of the present invention.
具体实施方式DETAILED DESCRIPTION
下面结合附图和具体实施例对本发明作进一步说明，以使本领域的技术人员可以更好地理解本发明并能予以实施，但所举实施例不作为对本发明的限定。The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it, but the embodiments are not intended to limit the present invention.
实施例一Embodiment 1
参照图1所示，本发明涉及一种下水管道内部缺陷识别方法，包括：Referring to FIG. 1 , the present invention relates to a method for identifying internal defects of a sewer pipe, comprising:
步骤S1，微型机器人携带摄像机实时采集管道内部图像数据，筛选并标注图像数据，具体包括以下步骤：Step S1, the micro robot carries a camera to collect real-time image data inside the pipeline, and screens and annotates the image data, which specifically includes the following steps:
步骤1.1：利用管道CCTV闭路电视检测技术，微型机器人携带摄像机下入地下管道，工作人员通过摄像机的反馈来控制机器人的前进方向，摄像机实时获取管道内部信息发送到上位机并保存拍摄视频，上位机将视频逐帧转化为图像数据集；Step 1.1: Using pipeline CCTV detection technology, the micro robot carries a camera and goes down into the underground pipeline. The staff controls the robot's forward direction through the feedback from the camera. The camera obtains the internal information of the pipeline in real time and sends it to the host computer and saves the video. The host computer converts the video frame by frame into an image data set.
步骤1.2：对采集的管道内部图像进行数据清洗，删除大量无意义重复图像和剔除像素模糊、管道特征残缺的图像，再进行人工筛查，将筛选处理好后的数据集进行统一编号处理。为保证缺陷识别的全面性，将下水管道缺陷划分为十类，分别为：错位、破裂、接口脱落、沉积、结垢、障碍、树根、渗漏、异物穿入和支管暗接。按照划分的缺陷类别对管道图像数据进行标注，采用Labelimg软件对图片进行手动标注，用矩形框标注出具体缺陷位置，生成XML格式文件。Step 1.2: Clean the collected internal images of the pipeline, delete a large number of meaningless duplicate images and remove images with blurred pixels and incomplete pipeline features, then perform manual screening and uniformly number the screened data sets. To ensure the comprehensiveness of defect identification, sewer pipe defects are divided into ten categories: dislocation, rupture, interface detachment, deposition, scaling, obstruction, tree roots, leakage, foreign body penetration and branch pipe blind connection. Label the pipeline image data according to the defect categories, use Labelimg software to manually label the images, use rectangular boxes to mark the specific defect locations, and generate XML format files.
步骤S2，对标注好的数据集进行预处理，并按照数量为7：3的比例划分为训练集和测试集，具体包括以下步骤：Step S2, preprocessing the labeled data set and dividing it into a training set and a test set in a ratio of 7:3, specifically including the following steps:
步骤2.1：将管道缺陷图像随机裁剪一块区域并将其作为新的输入，以适应不同角度下拍摄的图片，并随机调整图像数据的色调、饱和度和明度，提高模型在不同光线下检测缺陷的鲁棒性。采用Mosaic数据增强，针对当前一张图，从该epoch内已训练过的图中随机抽取3张图。对这四张图进行等比缩放，将当前图随机上下左右平移，然后根据当前图的位置，将其他3张图依次放在左边，下边、右下，随机缩放拼接提高小目标检测能力，帮助模型学习到不同目标之间的相对位置和尺度关系，并且在丰富模型数据集的同时提高训练效率。Step 2.1: Randomly crop an area of the pipeline defect image and use it as a new input to adapt to pictures taken at different angles, and randomly adjust the hue, saturation and brightness of the image data to improve the robustness of the model in detecting defects under different light. Mosaic data enhancement is used to randomly extract 3 pictures from the trained pictures in this epoch for the current picture. These four pictures are scaled proportionally, the current picture is randomly translated up, down, left and right, and then the other 3 pictures are placed on the left, bottom and bottom right in turn according to the position of the current picture. Random scaling and splicing improves the ability to detect small targets, helps the model learn the relative position and scale relationship between different targets, and improves training efficiency while enriching the model data set.
步骤2.2：将完成预处理的城市下水管道内部缺陷图像数据集中的.xml标签文件和.jpg图像文件按照7：3的比例划分成训练集和测试集；Step 2.2: Divide the .xml label files and .jpg image files in the preprocessed urban sewer internal defect image dataset into a training set and a test set in a ratio of 7:3;
步骤S3，建立改进的YOLOv5缺陷检测网络，并基于划分的训练集进行训练，本发明中改进后的YOLOv5网络模型结构如图2所示。整体可分为四部分：输入(input)、Backbone主干网络、Neck网络(颈部)和Head网络(检测头)，具体包括以下步骤：Step S3, establish an improved YOLOv5 defect detection network, and train it based on the divided training set. The improved YOLOv5 network model structure in the present invention is shown in Figure 2. The whole can be divided into four parts: input, Backbone backbone network, Neck network (neck) and Head network (detection head), which specifically includes the following steps:
步骤3.1：Backbone主干网络的结构为依次设置的2×CBS模块、C3_SK模块、CBS模块、C3_SK模块、CBS模块、C3_SK模块、CBS模块、C3_SK模块和SPPF模块,其中，C3_SK模块即在C3模块中引入SK卷积。C3模块包括并联的上支路和下支路，上支路为依次设置的CBS模块和RES模块(在该RES模块中引入SK卷积)，下支路为CBS模块，上支路和下支路的结果通过Concat模块连接之后再经过一个CBS模块。各模块的结构详见图2。Step 3.1: The structure of the Backbone network is a 2×CBS module, a C3_SK module, a CBS module, a C3_SK module, a CBS module, a C3_SK module, a CBS module, a C3_SK module, and an SPPF module, wherein the C3_SK module introduces SK convolution in the C3 module. The C3 module includes an upper branch and a lower branch connected in parallel. The upper branch is a CBS module and a RES module (SK convolution is introduced in the RES module) set in sequence, and the lower branch is a CBS module. The results of the upper branch and the lower branch are connected through a Concat module and then pass through a CBS module. The structure of each module is shown in Figure 2.
CBS模块包含卷积(Conv)、批归一化层(BN)和SiLU激活函数三部分。Res模块由两个CBS模块串联，其输出再与输入进行短接组成。SPPF模块由三个最大池化层(Maxpool)串联，通过Concat模块融合多尺度特征，将同一特征图不同尺度下的特征融合到一起，丰富特征图的语义特征。各模块的具体结构详见图2。The CBS module consists of three parts: convolution (Conv), batch normalization layer (BN) and SiLU activation function. The Res module consists of two CBS modules in series, and their output is short-circuited with the input. The SPPF module consists of three maximum pooling layers (Maxpool) in series. The Concat module fuses multi-scale features, and fuses features of different scales of the same feature map together to enrich the semantic features of the feature map. The specific structure of each module is shown in Figure 2.
Backbone主干网络中引入的SK注意力机制是一个动态选择性注意力模块，用于根据输入信息的多个尺度自适应调整其感受野的大小。模块结构图如图3所示。在该模块中，使用softmax算法对由不同内核大小的多个分支的信息引导进行融合，获得各分支的关注权重。对这些分支的不同关注权重导致融合层神经元有效感受野的大小不同。为了使神经元能够自适应调整其感受野大小，提出一种自动选择操作，即在不同内核大小的多个核之间进行“选择性核”(SK)卷积。具体来说，通过三个操作符，即拆分(Split)、融合(Fuse)和选择(Select)来实现SK卷积。其中，拆分具体为：对于任何给定特征映射进行4次转换，分别为表示对X采用卷积核大小为1×1的卷积操作，表示对X采用卷积核大小为3×3的卷积操作，表示对X采用卷积核大小为3×3和膨胀因子为2的空洞卷积操作，表示对X采用卷积核大小为3×3和膨胀因子为3的空洞卷积操作，R表示所有特征映射的集合，H表示特征图的高度，W表示特征图的宽度，C表示通道数。卷积核分别为1×1，3×3，5×5，7×7(即用了这四个卷积核能够得到)。为提高效率，把5×5和7×7的卷积核分别换成3×3的内核和膨胀因子为2和3的空洞卷积。F由高效的分组/深度卷积、批量归一化和ReLU组成。融合具体为：为了使神经元根据输入内容自适应地调整感受野大小，首先通过元素求和融合多个分支结果：使用全局平均池化来嵌入全局信息以生成通道统计信息：Fgp()表示全局平均池化函数，Uc表示U的二维形式，Uc(i,j)表示U的第c个通道中第[i，j]个元素值。设置一个特征描述因子z∈Rd×1，实现精确和自适应选择。特征描述因子具体计算公式如下：z＝其中，Ffc()表示全连接层操作，s为sc的三维形式，δ为ReLU函数，为批量归一化操作，W∈Rd×C，d＝max(C/r,L)。L为d的最小值，r为一个可变系数。在本申请中，L设置为32，r设置为16。选择具体为：在特征描述因子z的引导下，采用softmax算法和跨通道的软注意力自适应选择不同空间尺度的信息，公式为：The SK attention mechanism introduced in the Backbone network is a dynamic selective attention module, which is used to adaptively adjust the size of its receptive field according to multiple scales of the input information. The module structure diagram is shown in Figure 3. In this module, the softmax algorithm is used to fuse the information guided by multiple branches of different kernel sizes to obtain the attention weights of each branch. The different attention weights to these branches result in different sizes of effective receptive fields of neurons in the fusion layer. In order to enable neurons to adaptively adjust the size of their receptive fields, an automatic selection operation is proposed, that is, to perform "selective kernel" (SK) convolution between multiple kernels of different kernel sizes. Specifically, SK convolution is achieved through three operators, namely split (Split), fusion (Fuse) and selection (Select). Among them, splitting is specifically: for any given feature map, 4 transformations are performed, respectively. Indicates that a convolution operation with a convolution kernel size of 1×1 is applied to X. Indicates that a convolution operation with a convolution kernel size of 3×3 is applied to X. Indicates that a dilated convolution operation with a kernel size of 3×3 and a dilation factor of 2 is applied to X. = indicates that a dilated convolution operation with a kernel size of 3×3 and a dilation factor of 3 is applied to X. R represents the set of all feature maps, H represents the height of the feature map, W represents the width of the feature map, and C represents the number of channels. The convolution kernels are 1×1, 3×3, 5×5, and 7×7 (i.e., these four convolution kernels can be used to obtain ). To improve efficiency, the 5×5 and 7×7 convolution kernels are replaced with 3×3 kernels and dilation factors of 2 and 3 respectively. F consists of efficient grouped/depth convolution, batch normalization, and ReLU. The fusion is as follows: In order to make the neuron adaptively adjust the receptive field size according to the input content, firstly, multiple branch results are fused by element-wise summation: Use global average pooling to embed global information to generate channel statistics: Fgp () represents the global average pooling function, Uc represents the two-dimensional form of U, and Uc (i, j) represents the value of the [i, j]th element in the cth channel of U. Set a feature description factor z∈Rd×1 to achieve accurate and adaptive selection. The specific calculation formula of the feature description factor is as follows: z= Where Ffc () represents the fully connected layer operation, s is the three-dimensional form ofsc , δ is the ReLU function, is a batch normalization operation, W∈Rd×C , d＝max(C/r,L). L is the minimum value of d, and r is a variable coefficient. In this application, L is set to 32 and r is set to 16. The selection is as follows: under the guidance of the feature description factor z, the softmax algorithm and cross-channel soft attention are used to adaptively select information of different spatial scales, and the formula is:
ac+bc+cc+dc＝1。Ac,Bc,Cc,Dc∈RC×d，ac，bc，cc和dc分别表示的软注意力向量。通过各特征上的关注权重，得到最终的特征图V：V＝[V1,V2,...,Vc]，Vc∈RH×W。需要注意的是，字母带下标c的都是该字母的二维形式。 ac +bc +cc +dc = 1.A c ,Bc ,Cc ,Dc ∈RC×d , ac ,bc ,cc and dc represent The soft attention vector of . Through the attention weights on each feature, the final feature map V is obtained: V＝[V1 ,V2 ,...,Vc ], Vc ∈RH×W . It should be noted that letters with subscript c are the two-dimensional forms of the letters.
步骤3.2：改进Neck网络(颈部)特征融合方式，使用AFPN代替FPN。具体结构图如图4和图5所示。渐进特征金字塔结构AFPN避免了在FPN结构中由于非相邻层次之间出现的较大语义差距，导致的融合时部分信息丢失问题。在此结构中，将从Backbone主干网络的每个特征层中提取的不同尺度特征表示为{C3，C4，C5}，采用ASFF方法，首先将底层特征C3和C4融合到特征金字塔中，再加入C5。表达式定义为：本实施例中C3、C4和C5分别从第3、4、5个特征层中提取，相对次序为1、2、3，将它们的第i行j列特征元素值描述为和例如表示相应元素映射到l层。由于是渐进融合过程，的数值与这一融合轮次中l的最大值有关。其中，xl表示第l层特征，表示从s层到t层的特征图。融合过程中，为了减小非相邻层次特征之间的语义差距，采用自适应空间融合操作，使用softmax算法计算出α，β，γ的值：以α为例，其中，分别是对通过1×1的卷积核卷积得到。具体计算方式类比步骤3.1中的{ac，bc，cc，dc}。经过以上处理，最终得到一组关于α，β、γ的以不同权重融合了{C3，C4，C5}的多尺度特征{P3，P4，P5}。Step 3.2: Improve the feature fusion method of the Neck network (neck), and use AFPN instead of FPN. The specific structure diagram is shown in Figures 4 and 5. The progressive feature pyramid structure AFPN avoids the problem of partial information loss during fusion due to the large semantic gap between non-adjacent levels in the FPN structure. In this structure, the different scale features extracted from each feature layer of the Backbone network are represented as {C3 , C4 , C5 }. Using the ASFF method, the underlying features C3 and C4 are first fused into the feature pyramid, and then C5 is added. The expression is defined as: In this embodiment, C3, C4 and C5 are extracted from the 3rd, 4th and 5th feature layers respectively, with a relative order of 1, 2 and 3. Their feature element values in the i-th row and j-th column are described as and For example Indicates that the corresponding element is mapped to layer l. Since it is a gradual fusion process, The value of is related to the maximum value of l in this fusion round. Among them, xl represents the l-th layer feature, Represents the feature map from layer s to layer t. In the fusion process, in order to reduce the semantic gap between non-adjacent layer features, an adaptive spatial fusion operation is used, and the softmax algorithm is used to calculate the values of α, β, and γ: Take α as an example, in, Respectively It is obtained by convolution with a 1×1 convolution kernel. The specific calculation method is similar to {ac ,bc ,cc ,dc } in step 3.1. After the above processing, a set of multi-scale features {P3 ,P4 ,P5 } about α, β, and γ are finally obtained, which are fused with {C3 ,C4 ,C5 } with different weights.
如图2所示，本实施例中Neck网络的三条支路分别接收Backbone主干网络输出的三个特征层并进行处理，第一支路接收第一个特征层(即Backbone主干网络的第二个C3_SK模块)，进行依次设置的卷积、ASFF2、C3模块、ASFF3、C3模块；第二支路接收第二个特征层(即Backbone主干网络的第三个C3_SK模块)，进行依次设置的卷积、ASFF2、C3模块、ASFF3、C3模块；第三支路接收第三个特征层(即Backbone主干网络的SPPF模块)，进行依次设置的卷积、ASFF3、C3模块。其中，第一支路的卷积输出到第一支路、第二支路的ASFF2；第二支路的卷积输出到第一支路、第二支路的ASFF2；第一支路的第一个C3模块输出到第一支路、第二支路、第三支路的ASFF3；第二支路的第一个C3模块输出到第一支路、第二支路、第三支路的ASFF3；第三支路的卷积输出到第一支路、第二支路、第三支路的ASFF3。Head网络接收Neck网络三条支路的第二个C3模块输出，卷积后进行结果检测并输出。As shown in Figure 2, the three branches of the Neck network in this embodiment respectively receive and process the three feature layers output by the Backbone backbone network. The first branch receives the first feature layer (i.e., the second C3_SK module of the Backbone backbone network), and performs convolution, ASFF2, C3 module, ASFF3, and C3 module in sequence; the second branch receives the second feature layer (i.e., the third C3_SK module of the Backbone backbone network), and performs convolution, ASFF2, C3 module, ASFF3, and C3 module in sequence; the third branch receives the third feature layer (i.e., the SPPF module of the Backbone backbone network), and performs convolution, ASFF3, and C3 modules in sequence. Among them, the convolution of the first branch outputs to the ASFF2 of the first branch and the second branch; the convolution of the second branch outputs to the ASFF2 of the first branch and the second branch; the first C3 module of the first branch outputs to the ASFF3 of the first branch, the second branch, and the third branch; the first C3 module of the second branch outputs to the ASFF3 of the first branch, the second branch, and the third branch; the convolution of the third branch outputs to the ASFF3 of the first branch, the second branch, and the third branch. The Head network receives the output of the second C3 module of the three branches of the Neck network, performs result detection after convolution, and outputs it.
需要说明的是，本实施例中的ASFF2表示为x1→2α+x2β,其中，x1→2表示ASFF处理前当前层的特征(相当于上述的)，x2表示ASFF处理前下一层的特征(相当于上述的)。ASFF3表示为x1→3α+x2→3β+x3γ，其中，x1→3表示ASFF处理后当前层的特征，x2→3表示ASFF处理后下一层的特征，x3表示下下一层待融合的特征(相当于上述的)。It should be noted that ASFF2 in this embodiment is expressed as x1→2 α+x2 β, where x1→2 represents the feature of the current layer before ASFF processing (equivalent to the above ), x2 represents the features of the next layer before ASFF processing (equivalent to the above ). ASFF3 is expressed as x1→ 3α+x2→ 3β+x3γ , where x1→3 represents the features of the current layer after ASFF processing, x2→3 represents the features of the next layer after ASFF processing, andx3 represents the features of the next layer to be fused (equivalent to the above ).
步骤3.3：替换损失函数为Scylla-IoU：Step 3.3: Replace the loss function with Scylla-IoU:
步骤3.3.1：损失函数Scylla-IoU由角度成本、距离成本、形状成本和IoU成本四个部分组成，提高了目标检测模型的准确性和鲁棒性，避免过拟合问题。角度损失考虑了真实框与预测框向量之间的角度，定义为：Step 3.3.1: The loss function Scylla-IoU consists of four parts: angle cost, distance cost, shape cost, and IoU cost, which improves the accuracy and robustness of the object detection model and avoids overfitting problems. The angle loss takes into account the angle between the true box and the predicted box vector, and is defined as:
式中，Λ为角度损失，cw和ch为预测框与真实框中心点的水平距离和垂直距离，α为两中心点连线与水平线之间的夹角，σ为两点的欧氏距离。Where Λ is the angle loss,cw andch are the horizontal and vertical distances between the center points of the predicted box and the real box, α is the angle between the line connecting the two center points and the horizontal line, and σ is the Euclidean distance between the two points.
步骤3.3.2：距离损失考虑了真实框与预测框向量之间的距离，定义为：Step 3.3.2: The distance loss takes into account the distance between the true box and the predicted box vector and is defined as:
γ＝2-Λγ＝2-Λ
式中，Δ为距离损失，cw和ch为真实框和预测框最小外接矩形的宽和高，γ为与角度损失相关的系数。Where Δ is the distance loss,cw andch are the width and height of the minimum bounding rectangle of the real box and the predicted box, and γ is the coefficient related to the angle loss.
步骤3.3.3：形状损失考虑了地面真实值与预测框之间的形状差，其定义为：Step 3.3.3: The shape loss takes into account the shape difference between the ground truth and the predicted box, which is defined as:
式中，Ω为形状损失，w，h，wgt和hgt分别为预测框和真实框的宽和高，θ控制对形状损失的关注程度，为了避免过于关注形状损失而降低对预测框的移动，本方法中θ取4。Where Ω is the shape loss, w, h,wgt andhgt are the width and height of the predicted box and the true box respectively, and θ controls the degree of attention to the shape loss. In order to avoid paying too much attention to the shape loss and reducing the movement of the predicted box, θ is taken as 4 in this method.
步骤3.3.4：最终SIoU(Scylla-IoU)损失函数定义如下：Step 3.3.4: The final SIoU (Scylla-IoU) loss function is defined as follows:
步骤3.4：基于划分的训练集进行训练：Step 3.4: Training based on the divided training set:
步骤3.4.1：输入的图像大小设置为640×640，使用SGD优化器进行优化，批处理大小设置为32，训练迭代轮数为400，采用余弦退火学习速率策略，学习率动量为0.937，权值衰减为0.0005，初始学习速率为0.01。Step 3.4.1: The input image size is set to 640×640, and the SGD optimizer is used for optimization. The batch size is set to 32, the number of training iterations is 400, the cosine annealing learning rate strategy is adopted, the learning rate momentum is 0.937, the weight decay is 0.0005, and the initial learning rate is 0.01.
步骤S4，通过测试集测试改进的网络模型，并评估性能。Step S4, testing the improved network model through the test set and evaluating the performance.
为了更好的说明本发明的有效性，将改进YOLOv5和原YOLOv5的检测效果进行了对比，结果如表1所示。图6中的(a)展示了原YOLOv5检测效果，图6中的(b)展示了本实施例的方法检测效果，不难发现本实施例的方法检测效果更好。In order to better illustrate the effectiveness of the present invention, the detection effects of the improved YOLOv5 and the original YOLOv5 are compared, and the results are shown in Table 1. (a) in FIG6 shows the detection effect of the original YOLOv5, and (b) in FIG6 shows the detection effect of the method of this embodiment. It is not difficult to find that the detection effect of the method of this embodiment is better.
表1 YOLOv5和改进YOLOv5的检测效果对比表Table 1 Comparison of detection effects of YOLOv5 and improved YOLOv5
实施例二Embodiment 2
本实施例提供一种下水管道内部缺陷识别系统，包括：This embodiment provides a sewer pipe internal defect identification system, including:
采集模块：用于采集管道内部图像，进行数据清洗并筛选出有效的管道缺陷图像，按照预先设置的管道缺陷类型对有效的管道缺陷图像进行标注，得到标注好的数据集；Acquisition module: used to collect images inside the pipeline, perform data cleaning and filter out valid pipeline defect images, annotate valid pipeline defect images according to the preset pipeline defect types, and obtain annotated data sets;
预处理模块：用于对标注好的数据集进行预处理；Preprocessing module: used to preprocess the labeled data set;
训练模块：用于利用预处理后的数据集对改进的YOLOv5模型进行训练，得到训练好的YOLOv5模型；Training module: used to train the improved YOLOv5 model using the preprocessed data set to obtain a trained YOLOv5 model;
识别模块：用于通过所述训练好的YOLOv5模型对待检测的管道内部图像进行识别，得到识别结果。Recognition module: used to recognize the internal image of the pipeline to be detected through the trained YOLOv5 model to obtain a recognition result.
实施例三Embodiment 3
本实施例提供一种电子设备，包括存储器、处理器及存储在存储器上并可在处理器上运行的计算机程序，所述处理器执行所述计算机程序时实现实施例一所述下水管道内部缺陷识别方法的步骤。This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the method for identifying internal defects of sewer pipes described in Embodiment 1 are implemented.
实施例四Embodiment 4
本实施例提供一种计算机可读存储介质，其上存储有计算机程序，所述计算机程序被处理器执行时，实现实施例一所述下水管道内部缺陷识别方法的步骤。This embodiment provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the method for identifying internal defects of a sewer pipe described in the first embodiment are implemented.
本领域内的技术人员应明白，本申请的实施例可提供为方法、系统、或计算机程序产品。因此，本申请可采用完全硬件实施例、完全软件实施例、或结合软件和硬件方面的实施例的形式。而且，本申请可采用在一个或多个其中包含有计算机可用程序代码的计算机可用存储介质(包括但不限于磁盘存储器、CD-ROM、光学存储器等)上实施的计算机程序产品的形式。本申请实施例中的方案可以采用各种计算机语言实现，例如，面向对象的程序设计语言Java和直译式脚本语言JavaScript等。Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of complete hardware embodiment, complete software embodiment, or the embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code. The scheme in the embodiment of the present application can be implemented in various computer languages, for example, object-oriented programming language Java and literal translation scripting language JavaScript, etc.
本申请是参照根据本申请实施例的方法、设备(系统)、和计算机程序产品的流程图和/或方框图来描述的。应理解可由计算机程序指令实现流程图和/或方框图中的每一流程和/或方框、以及流程图和/或方框图中的流程和/或方框的结合。可提供这些计算机程序指令到通用计算机、专用计算机、嵌入式处理机或其他可编程数据处理设备的处理器以产生一个机器，使得通过计算机或其他可编程数据处理设备的处理器执行的指令产生用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的装置。The present application is described with reference to the flowchart and/or block diagram of the method, device (system) and computer program product according to the embodiment of the present application. It should be understood that each process and/or box in the flowchart and/or block diagram, and the combination of the process and/or box in the flowchart and/or block diagram can be realized by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a device for realizing the function specified in one process or multiple processes in the flowchart and/or one box or multiple boxes in the block diagram.
这些计算机程序指令也可存储在能引导计算机或其他可编程数据处理设备以特定方式工作的计算机可读存储器中，使得存储在该计算机可读存储器中的指令产生包括指令装置的制造品，该指令装置实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能。These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including an instruction device that implements the functions specified in one or more processes in the flowchart and/or one or more boxes in the block diagram.
这些计算机程序指令也可装载到计算机或其他可编程数据处理设备上，使得在计算机或其他可编程设备上执行一系列操作步骤以产生计算机实现的处理，从而在计算机或其他可编程设备上执行的指令提供用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的步骤。These computer program instructions may also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more processes in the flowchart and/or one or more boxes in the block diagram.
尽管已描述了本申请的优选实施例，但本领域内的技术人员一旦得知了基本创造性概念，则可对这些实施例作出另外的变更和修改。所以，所附权利要求意欲解释为包括优选实施例以及落入本申请范围的所有变更和修改。Although the preferred embodiments of the present application have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.
显然，上述实施例仅仅是为清楚地说明所作的举例，并非对实施方式的限定。对于所属领域的普通技术人员来说，在上述说明的基础上还可以做出其它不同形式变化或变动。这里无需也无法对所有的实施方式予以穷举。而由此所引申出的显而易见的变化或变动仍处于本发明创造的保护范围之中。Obviously, the above embodiments are merely examples for clear explanation and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived from these are still within the protection scope of the invention.
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410710803.XACN118736270A (en) | 2024-06-04 | 2024-06-04 | A method and system for identifying internal defects of sewer pipes |
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410710803.XACN118736270A (en) | 2024-06-04 | 2024-06-04 | A method and system for identifying internal defects of sewer pipes |
| Publication Number | Publication Date |
|---|---|
| CN118736270Atrue CN118736270A (en) | 2024-10-01 |
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202410710803.XAPendingCN118736270A (en) | 2024-06-04 | 2024-06-04 | A method and system for identifying internal defects of sewer pipes |
| Country | Link |
|---|---|
| CN (1) | CN118736270A (en) |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN119918164A (en)* | 2025-04-07 | 2025-05-02 | 吉林建筑大学 | A waterproof and thermal insulation construction method for building water supply and drainage and heating engineering |
| CN119992306A (en)* | 2025-04-14 | 2025-05-13 | 宁波上航测绘股份有限公司 | An automatic recognition system for pipeline shallow section images based on deep learning |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US10248664B1 (en)* | 2018-07-02 | 2019-04-02 | Inception Institute Of Artificial Intelligence | Zero-shot sketch-based image retrieval techniques using neural networks for sketch-image recognition and retrieval |
| CN117456257A (en)* | 2023-11-07 | 2024-01-26 | 云南大学 | Agricultural pest identification method based on improved YOLOv5 |
| CN117765373A (en)* | 2024-02-22 | 2024-03-26 | 山东大学 | Lightweight road crack detection method and system with self-adaptive crack size |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US10248664B1 (en)* | 2018-07-02 | 2019-04-02 | Inception Institute Of Artificial Intelligence | Zero-shot sketch-based image retrieval techniques using neural networks for sketch-image recognition and retrieval |
| CN117456257A (en)* | 2023-11-07 | 2024-01-26 | 云南大学 | Agricultural pest identification method based on improved YOLOv5 |
| CN117765373A (en)* | 2024-02-22 | 2024-03-26 | 山东大学 | Lightweight road crack detection method and system with self-adaptive crack size |
| Title |
|---|
| 张世强 等: "基于改进YOLOv5算法的钢铁表面缺陷检测", 《科学技术与工程》, vol. 23, no. 35, 31 December 2023 (2023-12-31), pages 15148 - 15157* |
| 徐振杰;陈庆奎;: "一种基于目标检测的偏斜图像校准方法", 《小型微型计算机系统》, vol. 41, no. 05, 15 May 2020 (2020-05-15), pages 1106 - 1111* |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN119918164A (en)* | 2025-04-07 | 2025-05-02 | 吉林建筑大学 | A waterproof and thermal insulation construction method for building water supply and drainage and heating engineering |
| CN119992306A (en)* | 2025-04-14 | 2025-05-13 | 宁波上航测绘股份有限公司 | An automatic recognition system for pipeline shallow section images based on deep learning |
| CN119992306B (en)* | 2025-04-14 | 2025-08-12 | 宁波上航测绘股份有限公司 | Pipeline shallow profile image automatic identification system based on deep learning |
| Publication | Publication Date | Title |
|---|---|---|
| CN113469177B (en) | Deep learning-based drainage pipeline defect detection method and system | |
| Kumar et al. | Deep learning–based automated detection of sewer defects in CCTV videos | |
| Xing et al. | A convolutional neural network-based method for workpiece surface defect detection | |
| Li et al. | Automatic pixel‐level multiple damage detection of concrete structure using fully convolutional network | |
| Xiao et al. | Surface defect detection using image pyramid | |
| Zhou et al. | Automatic sewer defect detection and severity quantification based on pixel-level semantic segmentation | |
| CN112258496A (en) | Underground drainage pipeline disease segmentation method based on full convolution neural network | |
| CN118736270A (en) | A method and system for identifying internal defects of sewer pipes | |
| Wang et al. | Development and improvement of deep learning based automated defect detection for sewer pipe inspection using faster R-CNN | |
| Ye et al. | Diagnosis of sewer pipe defects on image recognition of multi-features and support vector machine in a southern Chinese city | |
| Zhou et al. | Convolutional neural networks–based model for automated sewer defects detection and classification | |
| Cano-Ortiz et al. | An end-to-end computer vision system based on deep learning for pavement distress detection and quantification | |
| CN116092179A (en) | Improved Yolox fall detection system | |
| KR102721266B1 (en) | Deep learning-based sewerage defect detection method and apparatus | |
| CN117274355A (en) | Drainage pipeline flow intelligent measurement method based on acceleration guidance area convolutional neural network and parallel multi-scale unified network | |
| Feng et al. | Improved SOLOv2 detection method for shield tunnel lining water leakages | |
| CN114078106A (en) | Defect detection method based on improved Faster R-CNN | |
| CN116630242B (en) | Pipeline defect evaluation method and device based on instance segmentation | |
| CN113033371A (en) | CSP model-based multi-level feature fusion pedestrian detection method | |
| IZUMI | Low-cost training data creation for crack detection using an attention mechanism in deep learning models | |
| CN118537708A (en) | Hole detection image damage identification model based on improved convolutional neural network and application system | |
| CN118887544A (en) | A tunnel lining surface defect recognition method based on query matching attention | |
| CN118840336A (en) | Deep learning-based intelligent endoscopic defect detection method | |
| Cheng et al. | An efficient and scale-aware zero-shot industrial anomaly detection technique based on optimized CLIP | |
| JP7559247B2 (en) | Sewer pipe abnormality diagnosis support system, client machine, server machine for sewer pipe abnormality diagnosis support system, and related method |
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |