Improved YOLO-based electrical equipment defect image detection method and related equipment

Abstract

The invention discloses an improved YOLO-based electrical equipment defect image detection method and related equipment, and the method comprises the steps: inputting an electrical equipment image into a lightweight target detection model, and outputting defect category and position information; the model is improved based on a YOLO11 architecture, and images are processed through a feature extraction network, a fusion network and a detection head in sequence. A residual error enhancement re-parameterization convolution module is embedded into the feature extraction network, multi-scale features are extracted and fused through multiple branches during training, and re-parameterization is performed into a single branch during reasoning; a self-adaptive down-sampling module is adopted to replace a stride convolution down-sampling layer, and a dual-path structure reduces the resolution and retains information at the same time; the detection head is a lightweight scale decoupling detection head, decouples target classification and bounding box regression tasks, and adopts a detail enhancement structure. The invention aims to reduce the model volume and the calculation overhead on the premise of ensuring the detection precision, realize the real-time detection of the airship airborne equipment, construct a fine defect feature retention mechanism, reduce the small target omission ratio and improve the adaptability of the model to a complex dynamic environment.

Description

Technical Field

[0001] This invention belongs to the field of power equipment and facility inspection technology, specifically relating to a power equipment defect image detection method and related equipment based on an improved YOLO. Background Technology

[0002] Regular inspections of power equipment and facilities (such as wind turbine blades, transmission cables, and power poles) are crucial for ensuring the safe and stable operation of the power grid. With the widespread adoption of aerial inspection platforms such as drones and tethered airships, computer vision-based automatic defect detection technology has been widely applied. Among these, the YOLO series of models is often chosen as the basic framework for inspection tasks due to its good balance between speed and accuracy.

[0003] Applying existing YOLO models directly to power equipment defect detection, especially when deployed on edge devices such as tethered airships with limited computing resources, faces the following problems: (1) To ensure detection accuracy, mainstream YOLO models usually adopt deep network layers and complex feature pyramid structures, resulting in a large number of model parameters and floating-point operations (FLOPs). For example, even the lightweight YOLOv8n model has FLOPs as high as 38G, making it difficult to achieve high frame rate real-time processing on edge computing platforms such as Jetson Nano, which cannot meet the instant response requirements of airship inspection. (2) Defects such as surface cracks on wind turbine blades and broken insulators often appear small in aerial images, such as less than 32×32 pixels, with low contrast and weak texture. Existing models generally use stride convolution or pooling for downsampling, which easily loses key fine-grained texture information. At the same time, feature pyramids mostly adopt a top-down unidirectional fusion path, which does not make full use of the underlying detailed features, resulting in a high rate of missed detection for small target defects. (3) Aerial photography by airships is significantly affected by changes in lighting, weather conditions, and background interference (such as clouds and tower shadows). Traditional anchor frame mechanisms and fixed loss functions have limited adjustment capabilities when facing such dynamic scenarios, which can easily lead to background misjudgment or inaccurate positioning, resulting in a simultaneous increase in false detection rate and false negative rate in complex environments. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention provides a power equipment defect image detection method and related equipment based on an improved YOLO. The purpose is to reduce the model size and computational overhead while ensuring detection accuracy, enabling real-time detection on airship-borne equipment with limited computing power; to construct an effective mechanism for preserving subtle defect features, reducing the false negative rate of small targets; and to improve the model's adaptability to dynamic environments such as complex backgrounds and varying lighting conditions in airship aerial images, reducing false positives and false negatives.

[0005] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution: According to a first aspect of the present invention, a method for detecting defects in power equipment based on an improved YOLO is provided, characterized in that it includes: Acquire images of the power equipment to be inspected; The image of the power equipment is input into a pre-trained lightweight target detection model, which outputs the category and location information of defects in the image of the power equipment. The lightweight object detection model is based on an improved YOLO11 architecture. It processes the input image sequentially through a feature extraction network, a feature fusion network, and a detection head. The feature extraction network embeds a residual-enhanced reparameterized convolutional module, which is used to extract and fuse multi-scale contextual features through a multi-branch structure during the training phase and reparameterize it into a single-branch structure during the inference phase. The feature extraction network uses an adaptive downsampling module to replace the stride convolutional downsampling layer, which is used to reduce the resolution of the feature map while preserving information through a dual-path structure. The detection head is a lightweight scale-decoupled detection head, which is used to decouple the object classification task from the bounding box regression task and uses a detail enhancement structure for processing.

[0006] In one possible implementation of the first aspect, the residual-enhanced reparameterized convolution module includes multiple feature processing branches executed in parallel during the training phase, including at least: The first branch performs a residual depthwise convolution operation; The second branch performs dilated convolution operations with different dilation rates; The outputs of the multiple branches are fused at the channel dimension and then weighted by a channel attention mechanism.

[0007] In one possible implementation of the first aspect, the residual-enhanced reparameterized convolution module performs global response normalization after performing channel attention weighting. The calculation method for global response normalization is as follows:

[0008] in, Let X be the L2 norm of the feature X in the spatial dimension. The average response of the channel. , These are learnable parameters.

[0009] In one possible implementation of the first aspect, the dual-path structure of the adaptive downsampling module includes: The first path involves downsampling through a 3×3 convolution operation with a stride of 2. The second approach involves downsampling through pooling operations; The outputs of the first path and the second path are concatenated along the channel dimension to form a fused downsampling feature.

[0010] In one possible implementation of the first aspect, the adaptive downsampling module further includes a channel recalibration mechanism to process the fused downsampling features, specifically including: Calculate channel attention weights Specifically:

[0011] The channel attention weights are multiplied channel by channel by the fused downsampling features to obtain the final downsampling output. Specifically:

[0012] in, For global average pooling, It is the ReLU activation function. For the Sigmoid function, and For the weights of the fully connected layer, This indicates multiplication by channel.

[0013] In one possible implementation of the first aspect, the lightweight scale-decoupled detection head includes independent classification and regression branches, and shares a detail-enhancing convolutional unit; the detail-enhancing convolutional unit enhances edge and texture features by fusing the results of multi-directional differential convolutions, and its output features... The calculation method is as follows:

[0014] in, , , , , These represent the results of central difference, horizontal difference, vertical difference, anti-angle difference, and standard convolution operations, respectively. Indicates normalization, This represents a non-linear activation function.

[0015] In one possible implementation of the first aspect, the regression branch of the lightweight scale-decoupled detection head uses distributed focal loss to perform continuous value regression of the bounding box, and the calculation method for its predicted bounding box is as follows:

[0016] in, For the characteristics of the regression branch, The distribution focus loss function, This is a learnable channel scaling factor.

[0017] In one possible implementation of the first aspect, the method is deployed on a tethered airship equipped with edge computing devices for real-time inspection and defect detection of power equipment and facilities.

[0018] According to a second aspect of the present invention, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the aforementioned method for detecting defect images of power equipment based on an improved YOLO.

[0019] According to a third aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program, which, when executed by a processor, implements the aforementioned method for detecting defect images of power equipment based on an improved YOLO.

[0020] According to a fourth aspect of the present invention, a computer program product is provided that, when executed by a processor, implements the aforementioned method for detecting defect images of power equipment based on an improved YOLO.

[0021] Compared with the prior art, the present invention has at least the following beneficial effects: This invention provides a power equipment defect image detection method based on an improved YOLO. By introducing a residual-enhanced reparameterized convolutional module, it fully learns features during training using a multi-branch structure, and then uses structural reparameterization to effectively convert it into a single branch during inference. This enhances the model's feature representation capability without increasing the computational burden of the inference stage. Simultaneously, the design of an adaptive downsampling module and a lightweight scale-decoupled detection head further optimizes the computational process. These improvements work synergistically to reduce the number of parameters and computational complexity of the final lightweight target detection model, enabling it to run efficiently on edge computing devices with limited computing power, such as tethered airships, and meeting the high frame rate real-time processing requirements of power equipment inspection.

[0022] Specifically, for subtle defects such as cracks, the adaptive downsampling module, through a dual-path structure of convolution and pooling, preserves crucial fine-grained texture and structural information while performing spatial downsampling, mitigating the detail loss problem caused by traditional downsampling methods. Combined with the detail enhancement structure used in the detection head, it strengthens the processing of high-frequency features such as defect edges and irregular shapes, making the model more sensitive to low-contrast, small-sized defects in images. This results in higher recall even in complex backgrounds, effectively reducing the risk of missed detections of small defects. The residual-enhanced reparameterized convolution module integrates multi-scale contextual features during training, improving the model's understanding of illumination changes and background interference. The lightweight scale-decoupled detection head decouples classification and regression tasks, reducing interference between tasks and improving localization accuracy. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the specific embodiments of the present invention, the drawings used in the description of the specific embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0024] Figure 1 This is a flowchart of a power equipment defect image detection method based on an improved YOLO according to the present invention; Figure 2 This is a schematic diagram of the RAL-YOLO11 network structure; Figure 3 This is a schematic diagram of the C3k2-RCB module structure; Figure 4 This is a schematic diagram of the ADown module structure. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] It should be noted that the power equipment defect image detection method based on the improved YOLO provided by this invention can be implemented on a computer device, server, or embedded system equipped with a processor and memory, such as the onboard computing unit of a tethered airship. When the processor executes the computer program instructions stored in the memory, it implements the following method.

[0027] This invention first constructs a lightweight target detection model based on RAL-YOLO11. (Reference) Figure 2 The network structure diagram shown follows the basic framework of YOLO11, consisting of an input layer, a feature extraction network (backbone network), a feature fusion network (neck network), and a detection head. The improvement of this invention lies in the targeted reconstruction of the feature extraction network and the detection head.

[0028] The feature extraction network is responsible for extracting abstract features layer by layer from the input image. This invention embeds a Residual Enhanced Reparameterized Convolutional Module (RepConvBlock, RCB) into the feature extraction network and replaces the original stride convolutional downsampling layer with an Adaptive Downsampling Module (ADown). During training, the RCB module utilizes a multi-branch structure to enhance feature representation capabilities; during inference, it merges into a single branch through structural reparameterization to ensure efficiency. The ADown module, through a dual-path design, reduces the spatial resolution of the feature map while preserving as much fine-grained texture information as possible, which is crucial for defect detection.

[0029] The feature fusion network is responsible for fusing features from different levels of the backbone network to take into account both semantic and detailed information.

[0030] The detection head is responsible for predicting the category and precise location of defects based on the fused features. This invention designs a Lightweight Shared Detail Enhanced Convolutional Detection Head (LSDECD), which decouples the classification task from the bounding box regression task and introduces a detail enhancement structure to improve the localization and recognition capabilities of small-scale, weak-edge defects such as cracks and peeling.

[0031] like Figure 1 As shown, this invention provides a method for detecting defects in power equipment images based on an improved YOLO, specifically including the following steps: S101. Obtain an image of the power equipment to be inspected.

[0032] S102. Input the image of the power equipment into the pre-trained lightweight target detection model, and output the category and location information of the defects in the image of the power equipment.

[0033] The lightweight object detection model is based on an improved YOLO11 architecture. It processes the input image sequentially through a feature extraction network, a feature fusion network, and a detection head. The feature extraction network embeds a residual-enhanced reparameterized convolutional module, which is used to extract and fuse multi-scale contextual features through a multi-branch structure during the training phase and reparameterize it into a single-branch structure during the inference phase. The feature extraction network uses an adaptive downsampling module instead of stride convolution, which is used to reduce the resolution of the feature map while preserving information through a dual-path structure. The detection head is a lightweight scale-decoupled detection head, which is used to decouple the object classification task from the bounding box regression task and uses a detail enhancement structure for processing.

[0034] In detail, during the offline phase, an aerial image dataset containing defects such as wind turbine blade cracks, foreign objects in power transmission cables, and corrosion on power towers was collected and labeled. This dataset was then used to train the constructed RAL-YOLO11 model. During training, the RCB module operated in its multi-branch configuration, optimizing all parameters through backpropagation. After training, a one-time structural reparameterization operation was performed: all parallel convolutional branches in the RCB module were fused into an equivalent single-branch convolutional layer based on their parameters. The reparameterized RAL-YOLO11 model file, ready for efficient inference, was then deployed to the onboard computer of a tethered airship equipped with edge computing devices.

[0035] During the online inspection phase, the tethered airship flies along the inspection route, its onboard optical camera continuously capturing images of power equipment and facilities (such as wind turbine blades, transmission cables, and power poles). After acquiring the images of the power equipment to be inspected, the onboard computer performs resizing and normalization preprocessing before inputting them into the deployed RAL-YOLO11 model. The RAL-YOLO11 model processes the input images via forward propagation. This involves a feature extraction network embedding a single-branch morphology RCB module and an ADown module, a feature fusion network, and an LSDECD detection head. The model output includes defect categories such as cracks, foreign objects, and corrosion, as well as predicted bounding box locations.

[0036] In this embodiment, the residual augmentation reparameterized convolution module includes multiple feature processing branches executed in parallel during the training phase, including at least: a first branch that performs residual depthwise convolution operations; a second branch that performs dilated convolution operations with different dilation rates; and the outputs of the multiple branches are fused in the channel dimension and then weighted by a channel attention mechanism.

[0037] Specifically, in combination Figure 3 As shown, with input feature map X For example, the RCB module executes the following branches in parallel during training: The first branch is the residual depthwise convolution branch, which performs a residual depthwise convolution (ResDWConv) operation. This operation first performs a channel-wise depthwise convolution (DWConv) on the input feature map X, and then adds the result to the original input feature map X element-wise, that is:

[0038] Among them, DWConv( () indicates a channel-wise convolution operation. By stacking residuals, the model can simultaneously preserve the original spatial structure and the local texture extracted by convolution, thereby improving the recognizability of small defect regions.

[0039] The purpose of the first branch operation is to enhance the module's sensitivity to local textures, while preserving the original spatial structure information through residual connections, effectively reducing computational complexity and facilitating the identification of small defect areas.

[0040] The second branch is the multi-scale dilated convolution branch, which performs multi-scale dilated reparameterized convolutions (DRBs) with different dilation rates. This branch consists of a set of parallel depthwise convolutions, each with a different dilation rate. For example, 3x3 convolution kernels with dilation rates of 1, 3, and 5 can be set. During training, the outputs of multiple dilated convolution branches are summed, and the expression is:

[0041] in, Represents the basic convolutional kernel, used to extract the core local context; each For core size void ratio Deep convolutional kernels are used to capture long-range dependencies at different receptive field scales; Indicates the convolution operation; This corresponds to the Batch Normalization layer, used to maintain the numerical stability of the output of each branch. Through multi-scale combination, the module can effectively enhance the semantic representation of defects in complex backgrounds.

[0042] The purpose of the second branch is to simulate the receptive field of a large convolutional kernel and capture long-range contextual dependencies without increasing the number of parameters.

[0043] During the inference phase, all convolutional branches are integrated into a single equivalent large convolutional kernel through BN fusion and dilated kernel unrolling operations, achieving structural reparameterization and ensuring no additional computational overhead during inference.

[0044] Preferably, the outputs of all the above branches are concatenated and fused along the channel dimension. The fused features are then fed into the channel attention mechanism module. To highlight key features and suppress invalid information, this implementation introduces an SE (Squeeze-and-Excitation) module, which dynamically adjusts the feature response through adaptive weight learning along the channel dimension.

[0045] in, ( ) represents global average pooling; W2 represents the weights of the fully connected layer; δ( ) is the GELU activation function; σ( () is the Sigmoid activation function. This mechanism can adaptively emphasize channels sensitive to key areas such as cracks and erosion, thereby improving detection accuracy.

[0046] Preferably, after performing channel attention weighting, the residual enhancement reparameterized convolution module also performs global response normalization. The calculation method for global response normalization is as follows:

[0047] in, Let X be the L2 norm of the feature X in the spatial dimension. The average response of the channel. , These are learnable parameters.

[0048] In other words, in the RCB module, the features after channel attention weighting are not directly output, but are further processed by Global Response Normalization (GRN). The specific process is as follows: For the input feature X, first, its L2 norm in the spatial dimension (H, W) is calculated, resulting in a vector of channel dimensions. Then, the mean of this vector across all channels is calculated. Next, the L2 norm of each channel is divided by this mean to obtain a coefficient representing the relative response intensity of each channel. This coefficient is multiplied by the learnable parameter γ, then incremented by 1, and finally multiplied channel-by-channel with the original feature X, plus the learnable bias β.

[0049] Unlike traditional batch normalization, GRN does not rely on batch statistics, but instead performs channel-to-channel response normalization independently for each sample. This makes it more robust to scenarios with small batch training or unstable batch statistics, thus improving the model's generalization ability.

[0050] To enhance the modeling capabilities for complex textures and multi-scale defects in power equipment and facility defect detection tasks, this implementation introduces an RCB module into the YOLO11 backbone network. In traditional convolutional neural networks, lightweight structures typically rely on small convolutional kernels for local feature modeling. However, when faced with large receptive fields and long-range dependencies, local convolutions struggle to fully capture the global context, limiting feature representation capabilities. The RCB module achieves efficient feature fusion by integrating mechanisms such as dilated convolution for multi-scale feature extraction, channel attention enhancement, and global response normalization.

[0051] Preferably, in the output stage, the RCB module introduces LayerScale to assign a learnable scaling factor to the residual branches, thereby stabilizing the training of deep networks.

[0052] in, It is a learnable scalar used to control the contribution of the residual path; This represents the nonlinear transformation result of the projection branch. By limiting the magnitude of the residual perturbation, network training becomes more stable.

[0053] Meanwhile, the DropPath mechanism is introduced, which randomly discards some residual paths with probability during training. This is equivalent to implicit model ensemble, which can effectively reduce overfitting and enhance model robustness.

[0054] In one implementation, the dual-path structure of the adaptive downsampling module includes: a first path, which performs downsampling through a 3×3 convolution operation with a stride of 2; a second path, which performs downsampling through a pooling operation; and the outputs of the first path and the second path are concatenated along the channel dimension to form a fused downsampling feature.

[0055] Specifically, such as Figure 4 As shown, ADown employs a dual-branch dynamic downsampling structure, including a convolutional enhancement branch and a pooling structure preservation branch. It models complementary features using different computational methods, ensuring that the downsampled features possess both semantic expressiveness and structural information retention. Specifically, the input features are first divided into two equal quantum sets along the channel dimension; that is, the ADown module uses a dual-path structure to downsample the input features. The first path is the convolutional enhancement path, which uses a 3x3 convolutional layer with a stride of 2 for processing. This path is for the extraction and compression of semantic information.

[0056] The second path is a pooling structure preservation path. First, it uses average pooling (AvgPool) to achieve smooth downsampling, preserving global statistical features and complete edge contours. Then, it uses 1×1 convolution for channel compression and feature alignment, matching the output of the first path in the number of channels. Finally, the outputs of the first and second paths are concatenated along the channel dimension to form a fused downsampling feature. Specifically:

[0057] Where F1 and F2 represent the two channel branches of the input feature, namely the first path and the second path, respectively. ( ) represents the normalization and activation operations. This indicates channel splicing.

[0058] The design of the ADown module ensures that the downsampled features are rich in both semantic information extracted by the convolutional path and structural information retained by the pooling path. This helps to alleviate the loss of fine-grained defect information during single convolutional downsampling, balancing semantic refinement and structural fidelity, and making the downsampled features more expressive of multi-scale defects.

[0059] Preferably, the adaptive downsampling module further includes a channel recalibration mechanism to process the fused downsampling features, specifically including: calculating channel attention weights. The channel attention weights are multiplied channel by channel by the fused downsampling features to obtain the final downsampling output. .in, For global average pooling, It is the ReLU activation function. For the Sigmoid function, and For the weights of the fully connected layer, This indicates multiplication by channel.

[0060] Specifically, to further reduce the loss of key texture information during downsampling, ADown introduces a channel adaptive recalibration mechanism, which generates channel attention weights through global average pooling and a two-layer perceptron, thereby enhancing important features and suppressing redundant information.

[0061] In convolutional neural networks, downsampling operations not only affect the spatial resolution of feature maps but also determine the level of detail the network can perceive. Traditional stride convolution, due to its fixed nature, often inevitably loses local edge and fine-grained texture information while reducing spatial dimensions. Defects in power equipment, such as wind turbine blades, are characterized by weak texture, low contrast, and small scale, making them easily weakened during the downsampling stage, leading to problems like missed detections and localization misalignment. Therefore, the ADown module was adopted, employing a fusion strategy of multi-path structured downsampling and global context-driven channel recalibration to enhance texture fidelity and defect sensitivity after downsampling while maintaining inference efficiency.

[0062] Unlike traditional detector heads that share convolutional layers, LSDECD employs two independent paths for classification and regression to reduce gradient interference and enhance task relevance. The overall structure can be represented as: in, For input features, DWConv is a depthwise separable convolution used to capture directional edge structures.

[0063] In one implementation, the lightweight scale-decoupled detection head includes independent classification and regression branches, sharing a detail-enhancing convolutional unit (DEConv); this detail-enhancing convolutional unit enhances edge and texture features by fusing the results of multi-directional differential convolutions, and its output features... The calculation method is as follows:

[0064] in, , , , , These represent the results of central difference, horizontal difference, vertical difference, anti-angle difference, and standard convolution operations, respectively. Indicates normalization, This represents a non-linear activation function.

[0065] Specifically, the LSDECD detection head receives multi-scale feature maps from a feature fusion network as input. First, the input features undergo preliminary adjustment through a basic convolutional-normalization (Conv-GN) layer. Then, they enter two DEConv units, which explicitly enhance the extraction of high-frequency information such as defect edges, corners, and texture abrupt changes. DEConv extracts local edge and texture variations through a multi-directional differential convolutional structure, enhancing high-frequency information such as cracks, paint peeling, and contamination boundaries, ensuring high discriminability of small targets against complex backgrounds. The outputs of each branch are summed and fused, then normalized and activated by a non-linear activation function to obtain the enhanced features.

[0066] The DEConv unit applies multiple orientation-sensitive differential convolution kernels and one standard convolution kernel in parallel. The convolution operations include: central differential convolution to enhance the difference between the center point and surrounding points, horizontal differential convolution to enhance horizontal edges, vertical differential convolution to enhance vertical edges, anti-diagonal differential convolution to enhance diagonal edges, and the standard convolution to provide basic contextual information. The outputs of the five convolution operations are then element-wise summed. The summed result is normalized and non-linearly activated to obtain the final detail enhancement features. Subsequently, the network structure decoupled into two independent branches: a classification branch and a sub-branch. As input, after passing through several convolutional layers, the final output is the probability that each anchor point belongs to each defect category; the regression branch takes another path. As input, it passes through several convolutional layers and is responsible for predicting the offset, width, and height of the bounding box.

[0067] By using a design that combines enhanced detail with task decoupling, classification and regression tasks can be optimized separately based on richer detailed features, reducing interference between tasks. This is particularly beneficial for improving the localization and classification accuracy of small and ambiguous defect targets.

[0068] Preferably, the regression branch of the lightweight scale-decoupled detection head uses distributed focal loss to perform continuous value regression of the bounding box, and the calculation method for its predicted bounding box is as follows:

[0069] in, For the characteristics of the regression branch, The distribution focus loss function, This is a learnable channel scaling factor.

[0070] In other words, in the regression branch of the LSDECD detection head, in order to more accurately locate the defect boundary, especially for small targets with blurred boundaries, this invention uses distribution focal loss (DFL) to perform continuous value regression of the bounding box.

[0071] Features of the regression branch output After passing through a convolutional layer, its output is fed into the DFL calculation module. Simultaneously, a learnable channel scaling factor is introduced. Multiplying this by the output of DFL yields the final bounding box location prediction. Scaling factor During training, backpropagation enables automatic learning, allowing the network to adaptively adjust the contribution strength of different feature channels to the final localization result, ensuring more accurate boundary prediction for small targets. DFL allows the model to more accurately align the true boundaries of defects when dealing with cracks or spalling with unclear edges.

[0072] To further reduce computational complexity, LSDECD employs a weight-sharing mechanism in some shallow convolutional layers, reusing similar convolutional kernels between classification and regression branches to achieve parameter sharing and maintain feature consistency. This mechanism reduces computation by approximately 30% without sacrificing detection accuracy, thus balancing lightweight design and high performance.

[0073] In object detection tasks, the feature representation capability of the detection head directly affects localization accuracy and classification performance. Traditional YOLO series detection heads, while maintaining lightweight design and high frame rates, often suffer from insufficient boundary modeling for small or textured targets, leading to significant localization errors. This implementation uses the LSDECD detection head, which enhances edge and local texture information through differential convolution to improve the localization accuracy of small targets. Features across all scales share a detail enhancement module, reducing the number of parameters and enhancing cross-scale consistency. Multi-branch convolution is used during training to enhance capabilities, while inference is fused into a single convolution, reducing computational costs.

[0074] This invention introduces a residual-enhanced reparameterized convolutional module into the backbone network. During the training phase, it integrates multi-branch structures to enhance the feature representation capability for weak textures and small targets, and merges them into a single branch through structural reparameterization during the inference phase without increasing additional computational overhead. Secondly, it adopts an adaptive downsampling module to replace the traditional stride convolution, and combines a dual-path mechanism of convolution and pooling with a channel recalibration strategy to effectively retain fine-grained defect information such as cracks and peeling while reducing resolution. Finally, it designs a lightweight scale-decoupled detection head to decouple classification and regression tasks, and introduces multi-directional detail-enhanced convolution and distributed focus loss to improve the localization accuracy and recall rate of dense regions and multi-scale defects. The overall solution maintains the model's lightweight nature (39% reduction in parameters and 23% reduction in GFLOPs) while significantly improving detection accuracy (4.4% improvement in mAP@0.5–0.95) and robustness in complex scenarios, and can be efficiently deployed on tethered airships.

[0075] In another embodiment of the present invention, a computer device is provided, comprising a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions in the computer storage medium to achieve a corresponding method flow or corresponding function. The processor described in this embodiment of the present invention can be used in the operation of a power equipment defect image detection method based on an improved YOLO.

[0076] In another embodiment of the present invention, a storage medium is provided, specifically a computer-readable storage medium (Memory), which is a memory device in a computer device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be Random Access Memory (RAM) or non-volatile memory, such as at least one disk storage device. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the above embodiment regarding a method for detecting defect images of power equipment based on an improved YOLO.

[0077] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, optical storage, etc.) containing computer-usable program code.

[0078] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0079] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0080] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0081] This invention also provides a computer program product, which is used to execute any of the above-described methods for detecting power equipment defects based on improved YOLO. Since the computer program product provided by this invention belongs to the same inventive concept as the above-described method for detecting power equipment defects based on improved YOLO, it possesses all the advantages of the above-described method. Therefore, the beneficial effects of the computer program product provided by this invention will not be elaborated upon here.

[0082] In this invention, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to a specific feature, structure, material, or characteristic described in connection with that embodiment or example, which is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0083] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit them. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the scope of the technology disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention.

Claims

1. A method for detecting defects in power equipment images based on an improved YOLO, characterized in that, include: Acquire images of the power equipment to be inspected; The image of the power equipment is input into a pre-trained lightweight target detection model, which outputs the category and location information of defects in the image of the power equipment. The lightweight target detection model is based on the improved YOLO11 architecture. It processes the input image sequentially through a feature extraction network, a feature fusion network, and a detection head. The feature extraction network embeds a residual-enhanced reparameterized convolution module, which is used to extract and fuse multi-scale contextual features through a multi-branch structure during the training phase, and reparameterize it into a single-branch structure during the inference phase. The feature extraction network uses an adaptive downsampling module to replace the stride convolutional downsampling layer, which is used to reduce the resolution of the feature map while preserving information through a dual-path structure; the detection head is a lightweight scale-decoupled detection head, which is used to decouple the target classification task from the bounding box regression task and is processed using a detail enhancement structure.

2. The method for detecting defects in power equipment images based on an improved YOLO according to claim 1, characterized in that, The residual-enhanced reparameterized convolution module includes multiple feature processing branches executed in parallel during the training phase, including at least: The first branch performs a residual depthwise convolution operation; The second branch performs dilated convolution operations with different dilation rates; The outputs of the multiple branches are fused at the channel dimension and then weighted by a channel attention mechanism.

3. The method for detecting defects in power equipment based on an improved YOLO according to claim 2, characterized in that, After performing channel attention weighting, the residual-enhanced reparameterized convolution module also performs global response normalization. The calculation method for global response normalization is as follows: in, Let X be the L2 norm of the feature X in the spatial dimension. The average response of the channel. , These are learnable parameters.

4. The method for detecting defects in power equipment based on an improved YOLO according to claim 1, characterized in that, The dual-path structure of the adaptive downsampling module includes: The first path involves downsampling through a 3×3 convolution operation with a stride of 2. The second approach involves downsampling through pooling operations; The outputs of the first path and the second path are concatenated along the channel dimension to form a fused downsampling feature.

5. The method for detecting defects in power equipment based on an improved YOLO according to claim 4, characterized in that, The adaptive downsampling module also includes a channel recalibration mechanism to process the fused downsampling features, specifically including: Calculate channel attention weights Specifically: The channel attention weights are multiplied channel by channel by the fused downsampling features to obtain the final downsampling output. Specifically: in, For global average pooling, It is the ReLU activation function. For the Sigmoid function, and For the weights of the fully connected layer, This indicates multiplication by channel.

6. The method for detecting defects in power equipment images based on an improved YOLO according to claim 1, characterized in that, The lightweight scale-decoupled detection head includes independent classification and regression branches, and shares a detail enhancement convolutional unit. This detail enhancement convolutional unit enhances edge and texture features by fusing the results of multi-directional differential convolutions, and its output features... The calculation method is as follows: in, , , , , These represent the results of central difference, horizontal difference, vertical difference, anti-angle difference, and standard convolution operations, respectively. Indicates normalization. This represents a non-linear activation function.

7. The method for detecting defects in power equipment images based on an improved YOLO according to claim 6, characterized in that, The regression branch of the lightweight scale-decoupled detection head uses distributed focal loss to perform continuous value regression of the bounding box. The calculation method for its predicted bounding box is as follows: in, For the characteristics of the regression branch, For the distribution focus loss function, This is a learnable channel scaling factor.

8. The method for detecting defects in power equipment images based on an improved YOLO according to claim 1, characterized in that, The method is deployed on a tethered airship equipped with edge computing devices for real-time inspection and defect detection of power equipment and facilities.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements a power equipment defect image detection method based on an improved YOLO as described in any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements a power equipment defect image detection method based on an improved YOLO as described in any one of claims 1 to 8.

Images