技术领域Technical field
本发明属于电力系统的输电技术领域，具体涉及一种基于深度卷积神经网络的输电线路巡检的图像检测方法。The invention belongs to the field of power transmission technology of power systems, and specifically relates to an image detection method for power transmission line inspection based on a deep convolutional neural network.
背景技术Background technique
随着输电线路的不断增多以及对电力系统可靠性的要求日益提高，无人机巡检作为一种高效、安全的检测手段逐渐被广泛采用。无人机巡检可通过航拍图像对输电线路和杆塔进行监视，并获取大量的图片数据。为了提高巡检效率，许多先进的目标检测算法如YOLO系列和Faster-RCNN系列已经被引入，它们能够有效地检测输电线路巡检图片中的缺陷。然而，由于输电线路和杆塔上的零部件大小和形状存在很大的差异，传统的单一尺度目标检测算法可能会因缺乏多尺度信息的处理而导致漏检或误检的问题。因此，在这个背景下，引入多尺度目标检测与注意力机制成为一种有潜力的技术选择，它可以显著提升算法性能，使其在不同尺度下都能准确地检测出输电设备的缺陷。With the increasing number of transmission lines and the increasing requirements for the reliability of power systems, drone inspections are gradually being widely adopted as an efficient and safe detection method. Drone inspection can monitor transmission lines and towers through aerial images and obtain a large amount of picture data. In order to improve the efficiency of inspection, many advanced target detection algorithms such as YOLO series and Faster-RCNN series have been introduced, which can effectively detect defects in transmission line inspection pictures. However, due to the large differences in size and shape of components on transmission lines and towers, traditional single-scale target detection algorithms may cause missed or false detections due to lack of multi-scale information processing. Therefore, in this context, the introduction of multi-scale target detection and attention mechanisms has become a potential technical choice, which can significantly improve the performance of the algorithm and enable it to accurately detect defects in power transmission equipment at different scales.
结合多尺度目标检测与注意力机制，可以将多尺度目标检测和注意力机制相互融合，充分发挥它们在输电线路巡检图像处理中的优势。多尺度目标检测允许模型同时处理不同尺度的目标，通过在输入图像上应用不同大小的滑动窗口或不同尺度的特征图，从而检测出不同大小的目标。而注意力机制则赋予模型对图像中不同区域的不同关注程度，使得模型能够有选择地集中注意力于与目标相关的区域，忽略与目标无关的背景信息。将这两者相结合，可以使模型在不同层级上关注与目标相关的部分，从而在整个图像中更全面地检测出不同大小的缺陷。无论是远距离还是近距离拍摄的航拍图像，算法都能够适应性地检测出输电设备的缺陷。此外，结合多尺度目标检测与注意力机制还能提高算法的鲁棒性，使其能够应对一些特殊场景，如遮挡严重或部分缺陷难以直接观察到的情况，从而进一步提高缺陷检测的准确性和可靠性。Combining multi-scale target detection and attention mechanisms, multi-scale target detection and attention mechanisms can be integrated with each other to give full play to their advantages in transmission line inspection image processing. Multi-scale object detection allows the model to process objects of different scales simultaneously by applying sliding windows of different sizes or feature maps of different scales on the input image to detect objects of different sizes. The attention mechanism gives the model different degrees of attention to different areas in the image, allowing the model to selectively focus on areas related to the target and ignore background information irrelevant to the target. Combining the two allows the model to focus on target-relevant parts at different levels, thereby more comprehensively detecting defects of different sizes throughout the image. Whether it is aerial images taken from a distance or close up, the algorithm is able to adaptively detect defects in power transmission equipment. In addition, combining multi-scale target detection and attention mechanism can also improve the robustness of the algorithm, allowing it to cope with some special scenarios, such as severe occlusion or situations where some defects are difficult to directly observe, thereby further improving the accuracy and accuracy of defect detection. reliability.
目前，输电线路巡检图像的处理和缺陷检测依赖主要依赖人工进行，耗费大量的时间和人力资源。而通过结合多尺度目标检测与注意力机制的方法，可以显著提升算法的性能和效率。这项技术的创新之处在于将多尺度目标检测与注意力机制相结合，使得算法能够同时适应不同尺度的输电设备，准确地检测出不同大小的缺陷，并更关注重要区域，从而大幅降低漏检和误检的风险。该技术的应用将有助于提高输电线路巡检的自动化程度，节约大量人力和时间成本，同时有效地改善电力系统的可靠性和安全性。因此，结合多尺度目标检测与注意力机制的技术在输电线路巡检领域具有广阔的应用前景，为电力行业带来显著的经济和社会效益。At present, the processing and defect detection of transmission line inspection images mainly rely on manual work, which consumes a lot of time and human resources. By combining multi-scale target detection and attention mechanisms, the performance and efficiency of the algorithm can be significantly improved. The innovation of this technology is that it combines multi-scale target detection with an attention mechanism, allowing the algorithm to adapt to power transmission equipment of different scales at the same time, accurately detect defects of different sizes, and pay more attention to important areas, thereby significantly reducing leakage. risk of detection and false detection. The application of this technology will help improve the automation of transmission line inspections, save a lot of manpower and time costs, and effectively improve the reliability and safety of the power system. Therefore, technology that combines multi-scale target detection and attention mechanisms has broad application prospects in the field of transmission line inspection, bringing significant economic and social benefits to the power industry.
发明内容Contents of the invention
本发明的目的在于提供一种基于深度卷积神经网络的输电线路巡检图像检测方法，结合多尺度目标检测与注意力机制，旨在解决传统目标检测算法在处理不同尺度输电设备缺陷时可能导致漏检或误检的问题，以提高巡检效率、准确性和鲁棒性，实现自动化的输电线路缺陷检测。The purpose of this invention is to provide a transmission line inspection image detection method based on a deep convolutional neural network, combined with multi-scale target detection and attention mechanism, aiming to solve the problem that traditional target detection algorithms may cause when dealing with defects in power transmission equipment of different scales. The problem of missed or incorrect detection can be improved to improve the efficiency, accuracy and robustness of inspection and realize automated transmission line defect detection.
本发明采用的技术方案为：一种基于深度卷积神经网络的输电线路巡检图像检测方法，包括如下步骤：The technical solution adopted by the present invention is: a transmission line inspection image detection method based on a deep convolutional neural network, which includes the following steps:
步骤S1：数据预处理；收集含标注信息的输电线路巡检图像数据，进行图像增强和大小归一化；Step S1: Data preprocessing; collect transmission line inspection image data containing annotated information, perform image enhancement and size normalization;
步骤S2：多尺度目标检测模型设计；构建基于深度学习的多尺度目标检测模型，引入特征金字塔实现不同尺度目标检测；Step S2: Multi-scale target detection model design; build a multi-scale target detection model based on deep learning, and introduce feature pyramid to achieve target detection at different scales;
步骤S3：注意力机制引入；将注意力机制嵌入多尺度目标检测模型中，加强对重要区域的关注；Step S3: Introduce the attention mechanism; embed the attention mechanism into the multi-scale target detection model to strengthen attention to important areas;
步骤S4：模型训练与优化；使用预处理后的数据集训练多尺度目标检测模型，采用优化算法和学习率调整策略防止过拟合；Step S4: Model training and optimization; use the preprocessed data set to train the multi-scale target detection model, and use optimization algorithms and learning rate adjustment strategies to prevent overfitting;
步骤S5：多尺度融合与后处理；在预测阶段，融合多尺度目标检测结果，采用NMS算法抑制重叠检测结果，并通过注意力机制调整检测框；Step S5: Multi-scale fusion and post-processing; in the prediction stage, the multi-scale target detection results are fused, the NMS algorithm is used to suppress overlapping detection results, and the detection frame is adjusted through the attention mechanism;
步骤S6：模型评估与优化；使用标准评估指标对模型进行评估，针对结果优化模型的超参数和网络结构；Step S6: Model evaluation and optimization; use standard evaluation indicators to evaluate the model, and optimize the model's hyperparameters and network structure based on the results;
具体的，所述步骤S1：数据预处理详细步骤如下：Specifically, the detailed steps of step S1: data preprocessing are as follows:
S1.1收集含标注信息的输电线路巡检图像数据：从实际输电线路巡检任务中获取大量带有缺陷标注的航拍图像数据。这些图像数据将成为训练和评估多尺度目标检测模型的基础。S1.1 Collect transmission line inspection image data containing annotated information: Obtain a large amount of aerial image data with defect annotations from actual transmission line inspection tasks. These image data will form the basis for training and evaluating multi-scale object detection models.
S1.2图像增强：采用图像处理技术对图像进行增强，以增加图像的对比度、亮度，提高图像质量和清晰度，进一步凸显目标区域的特征，例如可以采用以下公式来实现对比度增强，对于图像I有：S1.2 Image enhancement: Use image processing technology to enhance the image to increase the contrast and brightness of the image, improve the image quality and clarity, and further highlight the characteristics of the target area. For example, the following formula can be used to achieve contrast enhancement. For image I have:
I'＝α*I+βI'＝α*I+β
其中，I’是增强后的图像，I是原始图像，α和β是调节参数，可以通过试验或自适应算法确定。Among them, I' is the enhanced image, I is the original image, α and β are adjustment parameters, which can be determined through experiments or adaptive algorithms.
S1.3大小归一化：由于不同的输电线路巡检图像可能具有不同的分辨率和尺寸，为了保证多尺度目标检测模型的稳定性和效率，需要将图像统一缩放到固定的大小。S1.3 Size normalization: Since different transmission line inspection images may have different resolutions and sizes, in order to ensure the stability and efficiency of the multi-scale target detection model, the images need to be uniformly scaled to a fixed size.
据预处理步骤将按照上述要素，对采集的输电线路巡检图像数据进行预处理。首先，收集大量含有缺陷标注的图像数据。然后，针对每个图像，应用图像增强技术提高图像质量。接着，将图像大小归一化到指定的高度和宽度，以便后续的多尺度目标检测模型训练。通过这样的数据预处理步骤，输入到后续多尺度目标检测模型的图像数据将具有更好的质量和一致的大小，有助于提高模型训练和检测的效果。同时，图像增强和大小归一化的过程可以通过现有的图像处理库和算法实现。The data preprocessing step will preprocess the collected transmission line inspection image data according to the above elements. First, a large amount of image data containing defect annotations is collected. Then, for each image, image enhancement techniques are applied to improve image quality. Next, the image size is normalized to the specified height and width for subsequent multi-scale target detection model training. Through such data preprocessing steps, the image data input to the subsequent multi-scale target detection model will have better quality and consistent size, helping to improve the effect of model training and detection. At the same time, the process of image enhancement and size normalization can be implemented through existing image processing libraries and algorithms.
具体的，所述步骤S2：多尺度目标检测模型设计中，使用了特征金字塔、FPN结构、多尺度特征图、路径聚合模块和PANet网络结构。Specifically, in the step S2: multi-scale target detection model design, feature pyramid, FPN structure, multi-scale feature map, path aggregation module and PANet network structure are used.
其中，特征金字塔是一种用于处理不同尺度目标的技术，其主要目的是通过在网络中引入多尺度的特征图层，使模型能够检测不同大小的目标。在本方案中，采用了PANet(Path Aggregation Network)结构来实现特征金字塔。Among them, feature pyramid is a technology used to deal with targets of different scales. Its main purpose is to enable the model to detect targets of different sizes by introducing multi-scale feature layers into the network. In this solution, the PANet (Path Aggregation Network) structure is used to implement the feature pyramid.
多尺度特征图：在PANet中，通过多个分支来生成不同尺度的特征图。具体指，在每个分支中应用不同大小的卷积核或池化操作，使得网络能够同时获取来自输入图像的不同尺度信息。Multi-scale feature maps: In PANet, feature maps of different scales are generated through multiple branches. Specifically, it refers to applying convolution kernels or pooling operations of different sizes in each branch so that the network can obtain different scale information from the input image at the same time.
路径聚合模块：路径聚合模块是PANet的核心组成部分。它通过聚合不同分支的特征图来获得全局感知能力，将来自不同分支的特征图进行逐元素求和或拼接，从而实现不同尺度特征的融合。Path aggregation module: The path aggregation module is the core component of PANet. It obtains global perception capabilities by aggregating feature maps from different branches, and sums or splices feature maps from different branches element by element to achieve the fusion of features at different scales.
PANet网络结构：整个PANet网络结构由多个特征提取分支和路径聚合模块组成。PANet network structure: The entire PANet network structure consists of multiple feature extraction branches and path aggregation modules.
具体的，步骤S2.1通过骨干网络(如ResNet或VGG)提取得到初始特征图；Specifically, step S2.1 extracts the initial feature map through the backbone network (such as ResNet or VGG);
步骤S2.2对初始特征图进行多尺度卷积和池化操作得到多尺度特征图；Step S2.2 Perform multi-scale convolution and pooling operations on the initial feature map to obtain a multi-scale feature map;
步骤S2.3通过路径聚合模块将多尺度特征图进行融合，形成最终的多尺度特征图，以供后续目标检测任务使用。PANet的路径聚合模块数学表达式如下所示：Step S2.3 uses the path aggregation module to fuse the multi-scale feature maps to form a final multi-scale feature map for use in subsequent target detection tasks. The mathematical expression of PANet’s path aggregation module is as follows:
其中，Fij表示第i个特征图分支中的第j个像素特征，K表示特征图的通道数，sij是由一个1×1卷积层计算得到的路径聚合权重。路径聚合模块可以自适应地学习每个分支在不同位置的权重，从而实现路径的动态聚合，使得PANet能够更好地处理不同尺度目标。Among them, Fij represents the j-th pixel feature in the i-th feature map branch, K represents the number of channels of the feature map, and sij is the path aggregation weight calculated by a 1×1 convolution layer. The path aggregation module can adaptively learn the weight of each branch at different positions to achieve dynamic aggregation of paths, allowing PANet to better handle targets of different scales.
具体的，所述步骤S3：引入注意力机制，该步骤详细内容如下：Specifically, step S3: introduce the attention mechanism. The details of this step are as follows:
其中，CBAM(Convolutional Block Attention Module)注意力模块是一种用于增强深度卷积神经网络对重要特征的关注的模块。它由两个部分组成：通道注意力和空间注意力。通道注意力用于增强模型对不同通道间特征的关注，而空间注意力用于增强模型对不同空间位置特征的关注。这样可以帮助模型更加精确地捕捉重要的特征，提高目标检测的准确性。Among them, the CBAM (Convolutional Block Attention Module) attention module is a module used to enhance the attention of deep convolutional neural networks to important features. It consists of two parts: channel attention and spatial attention. Channel attention is used to enhance the model's attention to features between different channels, while spatial attention is used to enhance the model's attention to features at different spatial locations. This can help the model capture important features more accurately and improve the accuracy of target detection.
将CBAM注意力模块融合到多尺度目标检测模型的不同层级中，以实现对不同尺度的特征图都应用注意力机制。模型在检测不同大小的目标时都能有针对性地增强对重要区域的关注，从而提高缺陷检测的准确性。The CBAM attention module is integrated into different levels of the multi-scale target detection model to apply the attention mechanism to feature maps of different scales. The model can enhance focus on important areas in a targeted manner when detecting targets of different sizes, thereby improving the accuracy of defect detection.
具体的，在多尺度目标检测模型中的不同层级引入CBAM注意力模块。首先在特征金字塔的不同尺度上分别应用通道注意力和空间注意力，得到增强后的特征图。然后将增强后的特征图与原始特征图进行相乘或相加的操作，得到最终融合后的特征图。这样一来，模型在不同层级上都能对不同尺度的特征图进行注意力加权，从而使得模型能够有选择地关注重要区域。在检测缺陷时，由于CBAM注意力模块的引入，模型能够更加精确地捕捉重要的特征信息，提高缺陷检测的准确性和鲁棒性。Specifically, the CBAM attention module is introduced at different levels in the multi-scale target detection model. First, channel attention and spatial attention are applied respectively on different scales of the feature pyramid to obtain enhanced feature maps. Then the enhanced feature map is multiplied or added to the original feature map to obtain the final fused feature map. In this way, the model can weight the attention of feature maps of different scales at different levels, allowing the model to selectively focus on important areas. When detecting defects, due to the introduction of the CBAM attention module, the model can more accurately capture important feature information and improve the accuracy and robustness of defect detection.
步骤S3.1，引入通道注意力模块，通道注意力公式化表示如下：Step S3.1, introduce the channel attention module, and the channel attention is formulated as follows:
Fc＝σ(MLPc(AVGPool(F)))Fc =σ(MLPc (AVGPool(F)))
其中，Fc是经过通道注意力增强后的特征图，F是原始特征图，AVGPool是全局平均池化操作，MLPc是一个多层感知机用于学习通道注意力的权重，σ是激活函数。Among them, Fc is the feature map enhanced by channel attention, F is the original feature map, AVGPool is the global average pooling operation, MLPc is a multi-layer perceptron used to learn the weight of channel attention, σ is the activation function .
步骤S3.2，引入空间注意力模块：Step S3.2, introduce the spatial attention module:
Fs＝σ(MLPs(MAXPool(F)))Fs =σ(MLPs (MAXPool(F)))
其中，Fs是经过空间注意力增强后的特征图，MAXPool是全局最大池化操作，MLPs是一个多层感知机用于学习空间注意力的权重，σ是激活函数。Among them, Fs is the feature map enhanced by spatial attention, MAXPool is the global maximum pooling operation, MLPs is a multi-layer perceptron used to learn the weight of spatial attention, and σ is the activation function.
步骤S3.3，注意力融合：Step S3.3, attention fusion:
Ffinal＝F+FattentionFfinal ＝F+Fattention
其中，Fattention是通道注意力和空间注意力相乘得到的注意力图，表示元素级别的乘法操作，Ffinal是融合后的最终特征图，F是原始特征图。Among them, Fattention is the attention map obtained by multiplying channel attention and spatial attention. Represents element-level multiplication operations, Ffinal is the final feature map after fusion, and F is the original feature map.
通过以上步骤，可将CBAM注意力模块融合到多尺度目标检测模型中，实现了对不同尺度特征图的注意力加权，进一步提高了缺陷检测的性能。Through the above steps, the CBAM attention module can be integrated into the multi-scale target detection model, achieving attention weighting on feature maps of different scales, and further improving the performance of defect detection.
具体的，所述步骤S4：模型训练与优化详细如下：Specifically, the details of step S4: model training and optimization are as follows:
其中，步骤S4.1，迭代训练并计算损失函数。批量梯度下降用于在训练过程中更新模型的参数以最小化损失函数。其公式如下：Among them, step S4.1 is to iteratively train and calculate the loss function. Batch gradient descent is used to update the parameters of the model during training to minimize the loss function. The formula is as follows:
其中，θ是模型的参数，α是学习率，m是批量大小，是损失函数关于第i个样本的梯度。在训练过程中，逐渐降低学习率可以帮助模型更稳定地收敛。本发明采用定期衰减策略：在训练的每个固定轮次或步骤后，将学习率乘以一个衰减因子。Among them, θ is the parameter of the model, α is the learning rate, m is the batch size, is the gradient of the loss function with respect to the i-th sample. During the training process, gradually reducing the learning rate can help the model converge more stably. The present invention adopts a periodic decay strategy: after each fixed round or step of training, the learning rate is multiplied by a decay factor.
步骤S4.2，正则化防止过拟合，本发明采用L2正则化防止过拟合，在损失函数中添加参数平方的惩罚项，公式如下：Step S4.2, regularization to prevent over-fitting. The present invention uses L2 regularization to prevent over-fitting, and adds a penalty term of the square of the parameter to the loss function. The formula is as follows:
其中，λ是正则化强度参数。Among them, λ is the regularization strength parameter.
在多尺度目标检测中，由于不同尺度的目标可能具有不同的重要性，为了平衡不同尺度缺陷的检测，可以对损失函数进行权重调整。例如，可以引入一个尺度相关的权重系数，使得不同尺度上的损失函数得到平衡。调整后的损失函数如下：In multi-scale target detection, since targets of different scales may have different importance, in order to balance the detection of defects at different scales, the weight of the loss function can be adjusted. For example, a scale-dependent weight coefficient can be introduced to balance the loss functions at different scales. The adjusted loss function is as follows:
Jweighted(θ)＝w1·J1(θ)+w2•J2(θ)+…+wn•Jn(θ)Jweighted (θ)＝w1 ·J1 (θ)+w2 •J2 (θ)+…+wn •Jn (θ)
其中，wn是第n个尺度上的权重系数，Jn(θ)是第n个尺度上的损失函数。Among them, wn is the weight coefficient on the nth scale, and Jn (θ) is the loss function on the nth scale.
模型训练与优化中，通过采用批量梯度下降等优化算法，结合学习率调整策略和正则化方法，可以有效地优化模型参数，防止过拟合。同时，通过损失函数的权重调整，实现对不同尺度缺陷的平衡处理，以确保多尺度目标检测模型具备全局感知能力，并能在不同尺度下准确地检测输电设备的缺陷。In model training and optimization, by using optimization algorithms such as batch gradient descent, combined with learning rate adjustment strategies and regularization methods, model parameters can be effectively optimized to prevent overfitting. At the same time, through weight adjustment of the loss function, balanced processing of defects at different scales is achieved to ensure that the multi-scale target detection model has global perception capabilities and can accurately detect defects in power transmission equipment at different scales.
具体的，所述步骤S5：多尺度融合与后处理中用到的内容和要素具体指：Specifically, the content and elements used in step S5: multi-scale fusion and post-processing specifically refer to:
多尺度目标检测结果融合：在预测阶段，多尺度目标检测模型会得到不同尺度下的目标检测结果。为了综合这些结果并得到最终的缺陷检测框，需要将不同尺度的检测结果进行融合。常见的融合方法有加权融合和置信度融合等。具体可通过以下公式进行加权融合：Fusion of multi-scale target detection results: In the prediction stage, the multi-scale target detection model will obtain target detection results at different scales. In order to synthesize these results and obtain the final defect detection frame, the detection results of different scales need to be fused. Common fusion methods include weighted fusion and confidence fusion. Specifically, weighted fusion can be performed through the following formula:
ScoreFused＝∑iwi×ScoreiScoreFused ＝∑iwi×Scorei
其中，ScoreFused是融合后的综合得分，Scorei是第i个尺度的目标得分，wi是相应尺度的权重。Among them, ScoreFused is the comprehensive score after fusion, Scorei is the target score of the i-th scale, and wi is the weight of the corresponding scale.
其中，多尺度融合后可能会出现重叠较多的检测结果，为了去除冗余的框并得到最终的缺陷检测框，采用非极大值抑制算法。该算法通过计算候选框的重叠面积，并保留得分最高的框，去除重叠区域内其他低得分的框。其具体步骤如下：Among them, after multi-scale fusion, there may be more overlapping detection results. In order to remove redundant frames and obtain the final defect detection frame, a non-maximum suppression algorithm is used. This algorithm calculates the overlapping area of candidate boxes, retains the box with the highest score, and removes other low-scoring boxes in the overlapping area. The specific steps are as follows:
步骤S5.1，对融合后的所有检测框按得分从高到低进行排序。Step S5.1: Sort all the fused detection frames according to their scores from high to low.
步骤S5.2，选择得分最高的框，将其加入最终结果集合，并从列表中删除该框。Step S5.2, select the box with the highest score, add it to the final result set, and delete the box from the list.
步骤S5.3，对剩余的框进行IoU(交并比)计算，若IoU大于设定阈值，则将该框从列表中删除。Step S5.3: Calculate the IoU (intersection over union) ratio of the remaining boxes. If the IoU is greater than the set threshold, the box is deleted from the list.
步骤S5.4，重复步骤步骤S5.2，和步骤步骤S5.3c，直到所有框都被处理完。Step S5.4, repeat step S5.2, and step S5.3c until all boxes have been processed.
在多尺度融合和非极大值抑制后，得到的缺陷检测框可能仍存在一定程度的不准确性或对重要区域的关注不足。为了进一步提高对重要区域的关注程度，可以结合注意力机制的结果对最终的检测框进行调整。具体来说，可以采用以下公式来融合注意力机制的结果：After multi-scale fusion and non-maximum suppression, the resulting defect detection frame may still have a certain degree of inaccuracy or insufficient attention to important areas. In order to further improve the degree of attention to important areas, the final detection frame can be adjusted based on the results of the attention mechanism. Specifically, the following formula can be used to fuse the results of the attention mechanism:
ScoreFinal＝ScoreFused×AttentionMapScoreFinal = ScoreFused × AttentionMap
其中，ScoreFinal是最终的检测得分，ScoreFused是融合后的综合得分，AttentionMap是注意力机制得到的关注程度图，通过这个关注程度图来增强对重要区域的关注程度。Among them, ScoreFinal is the final detection score, ScoreFused is the comprehensive score after fusion, and AttentionMap is the degree of attention map obtained by the attention mechanism. This attention degree map is used to enhance the degree of attention to important areas.
通过以上步骤，可以对多尺度目标检测的结果进行融合和后处理，得到最终的缺陷检测框，并通过注意力机制调整提高对重要区域的关注程度，从而进一步增强算法的检测准确性和鲁棒性。Through the above steps, the results of multi-scale target detection can be fused and post-processed to obtain the final defect detection frame, and the degree of attention to important areas can be increased through attention mechanism adjustment, thereby further enhancing the detection accuracy and robustness of the algorithm. sex.
具体的，步骤S6：模型评估与优化，包括以下详细步骤：Specifically, step S6: model evaluation and optimization includes the following detailed steps:
步骤S6.1选择标准的评估指标，通常包括精确率、召回率、F1-score。这些指标能够全面评估模型在不同场景下的性能表现，包括准确率、查全率和查准率等关键指标。对训练好的多尺度目标检测模型，使用测试集或交叉验证集中的数据进行评估。根据模型输出的检测结果和真实标签进行比对，计算各个评估指标的数值。Step S6.1 Select standard evaluation indicators, usually including precision, recall, and F1-score. These indicators can comprehensively evaluate the performance of the model in different scenarios, including key indicators such as accuracy, recall rate, and precision rate. Evaluate the trained multi-scale object detection model using data from the test set or cross-validation set. Based on the comparison between the detection results output by the model and the real labels, the values of each evaluation indicator are calculated.
步骤S6.2评估结果分析：分析模型在不同评估指标上的表现，特别关注F1-score。F1-score较高时，意味着模型在查准率和查全率上均有较好的表现。Step S6.2 Analysis of evaluation results: Analyze the performance of the model on different evaluation indicators, paying special attention to F1-score. When the F1-score is high, it means that the model has better performance in both precision and recall.
步骤S6.3模型优化：根据评估结果，对模型进行优化。Step S6.3 Model optimization: Optimize the model based on the evaluation results.
a.调整模型的超参数：例如调整学习率、网络深度、宽度等参数，以找到更好的模型配置。a. Adjust the hyperparameters of the model: for example, adjust parameters such as learning rate, network depth, and width to find a better model configuration.
b.网络结构改进：可以尝试引入更复杂的网络结构，或使用预训练模型来提取更高质量的特征。b. Network structure improvement: You can try to introduce a more complex network structure, or use a pre-trained model to extract higher quality features.
c.数据增强方式调整：优化数据增强策略，增加样本的多样性和数量，增强模型的泛化能力。c. Adjust the data enhancement method: optimize the data enhancement strategy, increase the diversity and quantity of samples, and enhance the generalization ability of the model.
步骤S6.3迭代优化：对模型进行改进后，再次进行训练和评估。如此迭代，直到达到满意的性能指标。Step S6.3 Iterative optimization: After improving the model, train and evaluate again. Iterate in this way until satisfactory performance indicators are achieved.
以上的内容和要素是在步骤S6中进行模型评估与优化时需要考虑和实施的。通过持续的评估、改进和优化，可以不断提升多尺度目标检测与注意力机制的性能，使其更适用于输电线路巡检图像处理，实现高效、准确的缺陷检测。The above content and elements need to be considered and implemented when performing model evaluation and optimization in step S6. Through continuous evaluation, improvement and optimization, the performance of multi-scale target detection and attention mechanisms can be continuously improved, making it more suitable for transmission line inspection image processing and achieving efficient and accurate defect detection.
本发明的有益效果：本发明结合多尺度目标检测与注意力机制，针对传统目标检测算法在处理输电线路巡检图像时可能出现的尺度差异导致漏检或误检的问题，提出了一种创新性的基于深度卷积神经网络的输电线路巡检图像检测方法。通过引入多尺度目标检测，使得模型能够同时适应不同尺度的输电设备缺陷，从而有效地提高了检测的准确性和鲁棒性。同时，结合注意力机制，增强了模型对重要区域的关注，进一步提高了缺陷检测的精度。该发明在实际应用中可以大幅降低漏检和误检的风险，显著提高了输电线路巡检的自动化程度，节约大量人力和时间成本，有助于改善电力系统的可靠性和安全性，对电力行业带来显著的经济和社会效益。Beneficial effects of the present invention: The present invention combines multi-scale target detection and attention mechanism, and proposes an innovative method to solve the problem that the scale difference that may occur in the traditional target detection algorithm when processing transmission line inspection images leads to missed detection or false detection. A comprehensive transmission line inspection image detection method based on deep convolutional neural network. By introducing multi-scale target detection, the model can simultaneously adapt to transmission equipment defects of different scales, thus effectively improving the accuracy and robustness of detection. At the same time, combined with the attention mechanism, the model's focus on important areas is enhanced, further improving the accuracy of defect detection. In practical applications, this invention can significantly reduce the risk of missed detections and false detections, significantly improve the automation of transmission line inspections, save a lot of manpower and time costs, and help improve the reliability and safety of the power system. The industry has brought significant economic and social benefits.
附图说明Description of drawings
图1为本发明的主体方法的步骤流程图；Figure 1 is a step flow chart of the main method of the present invention;
图2为本发明的步骤S1数据预处理流程图；Figure 2 is a flow chart of data preprocessing in step S1 of the present invention;
图3为本发明多尺度目标检测模型(PANet)架构图；Figure 3 is an architecture diagram of the multi-scale target detection model (PANet) of the present invention;
图4为本发明步骤S2的多尺度特征融合流程图；Figure 4 is a flow chart of multi-scale feature fusion in step S2 of the present invention;
图5为本发明步骤S3注意力机制流程图；Figure 5 is a flow chart of the attention mechanism in step S3 of the present invention;
图6为本发明步骤S4模型训练与优化流程图。Figure 6 is a flow chart of model training and optimization in step S4 of the present invention.
图7为本发明步骤S5多尺度融合与后处理流程图。Figure 7 is a flow chart of multi-scale fusion and post-processing in step S5 of the present invention.
图8为本发明步骤S6模型评估与优化流程图。Figure 8 is a flow chart of model evaluation and optimization in step S6 of the present invention.
图9为本发明实施例中不同技术方案检测效果对比图。Figure 9 is a comparison chart of detection effects of different technical solutions in the embodiment of the present invention.
具体实施方式Detailed ways
下面将结合本发明实施例中的附图，对本发明实施例中的技术方案进行清楚、完整地描述。基于本发明中的实施例，本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例，都属于本发明保护的范围，以下结合实施例具体说明。The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without making creative efforts belong to the scope of protection of the present invention, and will be described in detail below with reference to the embodiments.
一种基于深度卷积神经网络的输电线路巡检图像检测方法，包括如下步骤：A transmission line inspection image detection method based on deep convolutional neural network, including the following steps:
步骤S1：数据预处理Step S1: Data preprocessing
S1.1收集含标注信息的输电线路巡检图像数据：从实际输电线路巡检任务中获取大量带有缺陷标注的航拍图像数据，形成训练集和验证集。标注信息包括缺陷的位置和类别。S1.1 Collect transmission line inspection image data containing annotated information: Obtain a large amount of aerial image data with defect annotations from actual transmission line inspection tasks to form a training set and a verification set. The annotation information includes the location and category of the defect.
S1.2图像增强：采用图像处理技术对图像进行增强，增加图像的对比度、亮度，提高图像质量和清晰度，例如进行对比度增强。常用的图像增强方法包括直方图均衡化、自适应直方图均衡化等。S1.2 Image enhancement: Use image processing technology to enhance the image, increase the contrast and brightness of the image, and improve the image quality and clarity, such as contrast enhancement. Commonly used image enhancement methods include histogram equalization, adaptive histogram equalization, etc.
S1.3大小归一化：由于不同的输电线路巡检图像可能具有不同的分辨率和尺寸，为了保证多尺度目标检测模型的稳定性和效率，需要将图像统一缩放到固定的大小。一种常见的方法是将图像按比例缩放或填充到指定的高度和宽度。S1.3 Size normalization: Since different transmission line inspection images may have different resolutions and sizes, in order to ensure the stability and efficiency of the multi-scale target detection model, the images need to be uniformly scaled to a fixed size. A common approach is to scale or pad the image to a specified height and width.
步骤S2：多尺度目标检测模型设计Step S2: Multi-scale target detection model design
S2.1提取初始特征图：使用预训练的骨干网络(如ResNet或VGG)对输电线路巡检图像进行特征提取，得到初始特征图。骨干网络是深度卷积神经网络的前几层，用于提取图像的低级特征。S2.1 Extract the initial feature map: Use a pre-trained backbone network (such as ResNet or VGG) to extract features from the transmission line inspection image to obtain the initial feature map. The backbone network is the first few layers of a deep convolutional neural network and is used to extract low-level features of images.
S2.2多尺度卷积和池化操作：在初始特征图的基础上，通过不同大小的卷积核或池化操作，生成多尺度特征图。这样可以使得模型能够同时获取来自输入图像的不同尺度信息，增强对不同大小目标的检测能力。S2.2 Multi-scale convolution and pooling operations: Based on the initial feature map, multi-scale feature maps are generated through convolution kernels of different sizes or pooling operations. This allows the model to simultaneously obtain information of different scales from the input image and enhance the detection ability of targets of different sizes.
S2.3路径聚合：引入路径聚合模块，将多尺度特征图进行融合，得到最终的多尺度特征图。路径聚合模块通过聚合不同分支的特征图来获得全局感知能力，将来自不同分支的特征图进行逐元素求和或拼接，从而实现不同尺度特征的融合。S2.3 Path aggregation: Introduce the path aggregation module to fuse multi-scale feature maps to obtain the final multi-scale feature map. The path aggregation module obtains global perception capabilities by aggregating feature maps of different branches, and sums or splices feature maps from different branches element by element to achieve the fusion of features at different scales.
步骤S3：注意力机制引入Step S3: Introduction of attention mechanism
S3.1通道注意力模块：引入通道注意力模块，增强模型对不同通道间特征的关注。通道注意力模块主要包括两个部分：全局平均池化和多层感知机(MLP)。通过对特征图进行全局平均池化操作，得到全局特征向量，然后通过MLP学习通道注意力的权重，对特征图的每个通道进行加权，强化对重要通道的关注。S3.1 Channel attention module: The channel attention module is introduced to enhance the model’s attention to features between different channels. The channel attention module mainly consists of two parts: global average pooling and multi-layer perceptron (MLP). By performing a global average pooling operation on the feature map, the global feature vector is obtained, and then the weight of the channel attention is learned through MLP to weight each channel of the feature map to strengthen attention to important channels.
S3.2空间注意力模块：引入空间注意力模块，增强模型对不同空间位置特征的关注。空间注意力模块也包括全局池化和MLP，但是全局池化的操作是采用全局最大池化。通过学习空间注意力的权重，对特征图的每个空间位置进行加权，增强对重要区域的关注。S3.2 Spatial attention module: The spatial attention module is introduced to enhance the model's attention to the characteristics of different spatial locations. The spatial attention module also includes global pooling and MLP, but the global pooling operation uses global maximum pooling. By learning the weight of spatial attention, each spatial position of the feature map is weighted to enhance attention to important areas.
S3.3注意力融合：将通道注意力和空间注意力融合，得到最终的注意力图。可以采用元素级别的乘法或加法来融合通道注意力和空间注意力，增强对重要区域的关注程度。S3.3 Attention fusion: Fusion of channel attention and spatial attention to obtain the final attention map. Element-level multiplication or addition can be used to fuse channel attention and spatial attention to enhance the focus on important areas.
步骤S4：模型训练与优化Step S4: Model training and optimization
S4.1损失函数：定义损失函数，采用多尺度融合后的特征图进行目标检测，并与真实标签进行比对，计算损失值。常见的损失函数包括交叉熵损失和均方误差损失。S4.1 Loss function: Define the loss function, use the multi-scale fused feature map for target detection, compare it with the real label, and calculate the loss value. Common loss functions include cross-entropy loss and mean square error loss.
S4.2批量梯度下降：使用批量梯度下降法更新模型的参数，最小化损失函数。在每个批次中，模型根据训练数据计算梯度并更新参数，从而使得损失值逐渐降低，提高模型的准确性。S4.2 Batch gradient descent: Use the batch gradient descent method to update the parameters of the model and minimize the loss function. In each batch, the model calculates the gradient and updates the parameters based on the training data, so that the loss value gradually decreases and the accuracy of the model is improved.
S4.3学习率调整：采用定期衰减策略，降低学习率，有助于模型稳定收敛。学习率可以根据训练过程中的损失变化进行自适应调整。S4.3 Learning rate adjustment: Adopt a regular attenuation strategy to reduce the learning rate, which helps the model converge stably. The learning rate can be adaptively adjusted according to changes in loss during training.
S4.4正则化：添加L2正则化，防止模型过拟合。L2正则化是在损失函数中添加参数平方的惩罚项，限制模型参数的大小，减少过拟合风险。S4.4 Regularization: Add L2 regularization to prevent model overfitting. L2 regularization adds a penalty term of the square of the parameter to the loss function to limit the size of the model parameters and reduce the risk of overfitting.
步骤S5：多尺度融合与后处理Step S5: Multi-scale fusion and post-processing
S5.1多尺度融合：在预测阶段，多尺度目标检测模型会得到不同尺度下的目标检测结果。采用加权融合方法，将不同尺度检测结果进行加权融合，得到融合后的综合得分。S5.1 Multi-scale fusion: In the prediction stage, the multi-scale target detection model will obtain target detection results at different scales. Using the weighted fusion method, the detection results of different scales are weighted and fused to obtain a comprehensive score after fusion.
S5.2非极大值抑制：应用非极大值抑制算法，抑制重叠检测结果，保留得分最高的检测框。通过计算候选框的重叠面积，去除重叠区域内其他低得分的框，得到最终的缺陷检测框。S5.2 Non-maximum suppression: Apply the non-maximum suppression algorithm to suppress overlapping detection results and retain the detection frame with the highest score. By calculating the overlapping area of the candidate frames and removing other low-scoring frames in the overlapping area, the final defect detection frame is obtained.
步骤S6：模型评估与优化Step S6: Model evaluation and optimization
S6.1评估指标选择：选择标准的评估指标，例如精确率、召回率和F1-score，用于全面评估模型的性能。根据模型输出的检测结果和真实标签进行比对，计算各个评估指标的数值。S6.1 Evaluation index selection: Select standard evaluation indexes, such as precision, recall and F1-score, to comprehensively evaluate the performance of the model. Based on the comparison between the detection results output by the model and the real labels, the values of each evaluation indicator are calculated.
S6.2评估结果分析：分析模型在不同评估指标上的表现，特别关注F1-score。F1-score较高时，意味着模型在查准率和查全率上均有较好的表现。S6.2 Analysis of evaluation results: Analyze the performance of the model on different evaluation indicators, paying special attention to F1-score. When the F1-score is high, it means that the model has better performance in both precision and recall.
S6.3模型优化：根据评估结果，对模型进行优化。可以调整模型的超参数、网络结构或数据增强方式。也可以考虑采用更复杂的网络结构，或使用预训练模型来提取更高质量的特征。另外，可以优化数据增强策略，增加样本的多样性和数量，增强模型的泛化能力。S6.3 Model optimization: Optimize the model based on the evaluation results. You can adjust the model's hyperparameters, network structure, or data augmentation methods. You can also consider adopting a more complex network structure or using a pre-trained model to extract higher quality features. In addition, the data enhancement strategy can be optimized to increase the diversity and quantity of samples and enhance the generalization ability of the model.
S6.4迭代优化：对模型进行改进后，再次进行训练和评估。如此迭代，直到模型达到预期的性能。S6.4 Iterative optimization: After improving the model, train and evaluate again. Iterate like this until the model reaches the expected performance.
以下对本发明的有效性进行验证：The effectiveness of the present invention is verified as follows:
本部分验证基于深度卷积神经网络和注意力机制的多尺度目标检测技术在输电无人机巡检图片数据集上的有效性。使用一个无人机输电线路巡检的数据集，并将该数据集应用于本技术方案和其他两种常见的目标检测方法，以比较它们的性能表现。This section verifies the effectiveness of multi-scale target detection technology based on deep convolutional neural networks and attention mechanisms on the power transmission drone inspection image data set. A data set of drone power transmission line inspections is used, and the data set is applied to this technical solution and two other common target detection methods to compare their performance.
构建了一个包含1000张输电线路巡检无人机航拍图片的数据集，图像分辨率为6000×4000像素。每张图片中包含了不同尺度的输电设备缺陷，包括杆塔、绝缘子、导线等。数据集的标签信息包括每个缺陷的位置(边界框)和类别(如杆塔、绝缘子等)。A data set containing 1,000 aerial images of transmission line inspection drones was constructed, with an image resolution of 6000 × 4000 pixels. Each picture contains defects in power transmission equipment at different scales, including towers, insulators, wires, etc. The label information of the dataset includes the location (bounding box) and category (such as pole, insulator, etc.) of each defect.
将使用以下几种方法进行比较：The following methods will be used for comparison:
1.本专利技术方案：基于深度卷积神经网络和注意力机制的多尺度目标检测方法。1. This patented technical solution: multi-scale target detection method based on deep convolutional neural network and attention mechanism.
2.YOLOv3(You Only Look Once)：一种流行的单阶段目标检测算法。使用普通的深度卷积神经网络进行目标检测，不引入多尺度和注意力机制。2.YOLOv3 (You Only Look Once): A popular single-stage target detection algorithm. Use ordinary deep convolutional neural networks for target detection without introducing multi-scale and attention mechanisms.
3.Faster R-CNN：一种常用的两阶段目标检测算法。使用普通的深度卷积神经网络进行目标检测，不引入多尺度和注意力机制。3.Faster R-CNN: A commonly used two-stage target detection algorithm. Use ordinary deep convolutional neural networks for target detection without introducing multi-scale and attention mechanisms.
数据预处理：对数据集进行图像增强，包括调整亮度、对比度、随机裁剪和翻转等，以扩充数据集，并保证数据的多样性。Data preprocessing: Image enhancement of the data set, including adjusting brightness, contrast, random cropping and flipping, etc., to expand the data set and ensure data diversity.
模型训练：使用不同的方法对数据集进行训练，分别训练本专利技术方案、YOLOv3(单尺度目标检测)和FasterR-CNN(单尺度目标检测)。Model training: Use different methods to train the data set, and train this patented technical solution, YOLOv3 (single-scale target detection) and FasterR-CNN (single-scale target detection) respectively.
模型评估：使用测试集对训练好的模型进行评估，计算精确率、召回率、F1等评估指标，以评估各个方法的性能。Model evaluation: Use the test set to evaluate the trained model, and calculate evaluation indicators such as precision, recall, and F1 to evaluate the performance of each method.
结合图9所示，本实施例对比了YOLOv3(单尺度目标检测)和FasterR-CNN(单尺度目标检测)对于小尺度目标(带销插销丢失插销)的检测结果。本发明提出的算法检出丢失插销2个，正常10个(完全正确)。Faster R-CNN(单尺度目标检测)检出丢失插销4个，正常8个(2个误检)。YOLOv3(单尺度目标检测)检出丢失插销2个，正常8个(2个漏检)。As shown in Figure 9, this embodiment compares the detection results of YOLOv3 (single-scale target detection) and FasterR-CNN (single-scale target detection) for small-scale targets (pins with pins and missing pins). The algorithm proposed by the present invention detects 2 missing pins and 10 normal pins (completely correct). Faster R-CNN (single-scale object detection) detected 4 missing pins and 8 normal pins (2 false detections). YOLOv3 (single-scale target detection) detected 2 missing pins and 8 normal ones (2 missed pins).
将整个测试集的结果进行比较，结果如下：Comparing the results of the entire test set, the results are as follows:
与传统的单尺度目标检测方法相比，本专利技术方案在检测不同尺度的输电设备缺陷时表现更为优异。通过引入多尺度目标检测模型，能够更好地捕捉不同尺度目标的特征，从而提高了检测的准确性和鲁棒性。相比于YOLO和Faster R-CNN这两种流行的目标检测方法，本专利技术方案引入了注意力机制，进一步提升了对重要区域的关注程度。同时，在多尺度融合方面也有优势，使得模型能够同时获取来自输入图像的不同尺度信息，增强了对不同大小目标的检测能力。Compared with traditional single-scale target detection methods, this patented technical solution performs better in detecting defects in power transmission equipment of different scales. By introducing a multi-scale target detection model, the characteristics of targets at different scales can be better captured, thereby improving the accuracy and robustness of detection. Compared with YOLO and Faster R-CNN, two popular target detection methods, this patented technical solution introduces an attention mechanism to further enhance the degree of attention to important areas. At the same time, it also has advantages in multi-scale fusion, allowing the model to obtain different scale information from the input image at the same time, enhancing the detection ability of targets of different sizes.
综上所述，本专利技术方案在输电无人机巡检图片数据集上表现出更好的检测性能，相较于传统方法和其他流行目标检测算法，在准确性、鲁棒性和检测效率上均具有优势，因此更为有效。To sum up, this patented technical solution shows better detection performance on the power transmission drone inspection picture data set. Compared with traditional methods and other popular target detection algorithms, it has better accuracy, robustness and detection efficiency. Both have advantages and are therefore more effective.
对于本领域技术人员而言，显然本发明不限于上述示范性实施例的细节，而且在不背离本发明的精神或基本特征的情况下，能够以其他的具体形式实现本发明。因此，无论从哪一点来看，均应将实施例看作是示范性的，而且是非限制性的，本发明的范围由所附权利要求而不是上述说明限定，因此旨在将落在权利要求的得同要件的含义和范围内的所有变化囊括在本发明内。不应将权利要求中的任何附图标记视为限制所涉及的权利要求。It is obvious to those skilled in the art that the present invention is not limited to the details of the above-described exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be regarded as illustrative and non-restrictive from any point of view, and the scope of the present invention is defined by the appended claims rather than the above description, and it is therefore intended that all claims falling within the claims All changes within the meaning and scope of the equivalent requirements are included in the present invention. Any reference signs in the claims shall not be construed as limiting the claim in question.
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311319251.1ACN117541535A (en) | 2023-10-12 | 2023-10-12 | A transmission line inspection image detection method based on deep convolutional neural network |
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311319251.1ACN117541535A (en) | 2023-10-12 | 2023-10-12 | A transmission line inspection image detection method based on deep convolutional neural network |
| Publication Number | Publication Date |
|---|---|
| CN117541535Atrue CN117541535A (en) | 2024-02-09 |
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311319251.1AWithdrawnCN117541535A (en) | 2023-10-12 | 2023-10-12 | A transmission line inspection image detection method based on deep convolutional neural network |
| Country | Link |
|---|---|
| CN (1) | CN117541535A (en) |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN118470577A (en)* | 2024-07-09 | 2024-08-09 | 陕西陕煤黄陵矿业有限公司 | Inspection scene identification method and system based on big data |
| CN118781128A (en)* | 2024-09-13 | 2024-10-15 | 绍兴布吉针织有限公司 | A knitted fabric defect detection method and system based on image processing |
| CN119048495A (en)* | 2024-10-29 | 2024-11-29 | 南京信息工程大学 | Multi-scale feature fusion 3D printing defect detection method based on improvement YOLOv8 |
| CN119206342A (en)* | 2024-09-20 | 2024-12-27 | 山东阳光新材料科技有限公司 | A polyester fiber protective net abnormality monitoring system and monitoring method |
| CN119229097A (en)* | 2024-11-28 | 2024-12-31 | 中科方寸知微(南京)科技有限公司 | Autonomous inspection wire detection method and system for UAV based on Bezier curve modeling |
| CN119251641A (en)* | 2024-08-13 | 2025-01-03 | 国网湖北省电力有限公司经济技术研究院 | Transmission line reliability prediction method, system and equipment based on SENet and EffNet |
| CN119672078A (en)* | 2024-12-05 | 2025-03-21 | 中科方寸知微(南京)科技有限公司 | Wire registration method and system between UAV aerial images based on deep learning |
| CN119671981A (en)* | 2024-11-29 | 2025-03-21 | 苏州医朵云健康股份有限公司 | Drug logistics management device and method based on machine vision |
| CN119994875A (en)* | 2025-01-16 | 2025-05-13 | 长沙理工大学 | A short-term load forecasting method based on multi-scale spatiotemporal DCNN feature extraction |
| CN120355940A (en)* | 2025-06-18 | 2025-07-22 | 天津长芦汉沽盐场有限责任公司 | Power transmission line sag detection method and system based on light neural network model |
| CN120471997A (en)* | 2025-04-28 | 2025-08-12 | 华北电力大学（保定） | Insulator defect detection and positioning method and system based on cross-scale feature fusion |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN118470577A (en)* | 2024-07-09 | 2024-08-09 | 陕西陕煤黄陵矿业有限公司 | Inspection scene identification method and system based on big data |
| CN119251641A (en)* | 2024-08-13 | 2025-01-03 | 国网湖北省电力有限公司经济技术研究院 | Transmission line reliability prediction method, system and equipment based on SENet and EffNet |
| CN118781128A (en)* | 2024-09-13 | 2024-10-15 | 绍兴布吉针织有限公司 | A knitted fabric defect detection method and system based on image processing |
| CN119206342A (en)* | 2024-09-20 | 2024-12-27 | 山东阳光新材料科技有限公司 | A polyester fiber protective net abnormality monitoring system and monitoring method |
| CN119048495A (en)* | 2024-10-29 | 2024-11-29 | 南京信息工程大学 | Multi-scale feature fusion 3D printing defect detection method based on improvement YOLOv8 |
| CN119229097B (en)* | 2024-11-28 | 2025-05-16 | 中科方寸知微(南京)科技有限公司 | Unmanned aerial vehicle autonomous inspection wire detection method and system based on Bezier curve modeling |
| CN119229097A (en)* | 2024-11-28 | 2024-12-31 | 中科方寸知微(南京)科技有限公司 | Autonomous inspection wire detection method and system for UAV based on Bezier curve modeling |
| CN119671981A (en)* | 2024-11-29 | 2025-03-21 | 苏州医朵云健康股份有限公司 | Drug logistics management device and method based on machine vision |
| CN119672078A (en)* | 2024-12-05 | 2025-03-21 | 中科方寸知微(南京)科技有限公司 | Wire registration method and system between UAV aerial images based on deep learning |
| CN119672078B (en)* | 2024-12-05 | 2025-07-22 | 中科方寸知微(南京)科技有限公司 | Method and system for registering electric wires between aerial images of unmanned aerial vehicle based on deep learning |
| CN119994875A (en)* | 2025-01-16 | 2025-05-13 | 长沙理工大学 | A short-term load forecasting method based on multi-scale spatiotemporal DCNN feature extraction |
| CN120471997A (en)* | 2025-04-28 | 2025-08-12 | 华北电力大学（保定） | Insulator defect detection and positioning method and system based on cross-scale feature fusion |
| CN120355940A (en)* | 2025-06-18 | 2025-07-22 | 天津长芦汉沽盐场有限责任公司 | Power transmission line sag detection method and system based on light neural network model |
| CN120355940B (en)* | 2025-06-18 | 2025-09-30 | 天津长芦汉沽盐场有限责任公司 | Power transmission line sag detection method and system based on light neural network model |
| Publication | Publication Date | Title |
|---|---|---|
| CN117541535A (en) | A transmission line inspection image detection method based on deep convolutional neural network | |
| US11308714B1 (en) | Artificial intelligence system for identifying and assessing attributes of a property shown in aerial imagery | |
| CN116206185A (en) | Lightweight small target detection method based on improved YOLOv7 | |
| CN113065558A (en) | Lightweight small target detection method combined with attention mechanism | |
| CN113592786B (en) | A deep learning-based ocean mesoscale eddy detection method | |
| CN114998566B (en) | An interpretable multi-scale infrared small target detection network design method | |
| CN114998603B (en) | An underwater target detection method based on deep multi-scale feature factor fusion | |
| CN112464911A (en) | Improved YOLOv 3-tiny-based traffic sign detection and identification method | |
| CN111507371A (en) | method and apparatus | |
| CN112288700A (en) | Rail defect detection method | |
| CN108133235A (en) | A kind of pedestrian detection method based on neural network Analysis On Multi-scale Features figure | |
| CN111340019A (en) | Detection method of granary pests based on Faster R-CNN | |
| CN113313678A (en) | Automatic sperm morphology analysis method based on multi-scale feature fusion | |
| CN114170642A (en) | Image detection processing method, device, equipment and storage medium | |
| CN118506263A (en) | A foreign body detection method for power transmission lines in complex environments based on deep learning | |
| CN117635904A (en) | Dynamic self-adaptive feature-aware credible low-speed unmanned aerial vehicle detection method | |
| CN111667461B (en) | A Method for Detection of Abnormal Objects on Transmission Lines | |
| CN118351435A (en) | A method and device for detecting target in UAV remote sensing images based on lightweight model LTE-Det | |
| CN119048516B (en) | Defect solder ball detection method based on improved YOLO v8 network | |
| Liu et al. | ViTR-Net: An unsupervised lightweight transformer network for cable surface defect detection and adaptive classification | |
| CN114202649B (en) | A target localization method and system based on multi-scale attention constraints | |
| CN120031839A (en) | Photovoltaic panel surface dirt detection and analysis system applied to UAV platform | |
| CN118967550B (en) | Lightweight insulator defect detection method based on improved YOLOv network | |
| CN119478351A (en) | A method for target detection based on improved YOLOv8-DES model | |
| CN114842192A (en) | Aero-engine blade damage identification model, damage identification method and system |
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| WW01 | Invention patent application withdrawn after publication | Application publication date:20240209 | |
| WW01 | Invention patent application withdrawn after publication |