Multi-weather robust target detection method and device for electric power inspection and storage medium

Abstract

The invention discloses a multi-weather robust target detection method and device for electric power inspection and a storage medium, and relates to the technical field of image processing, and the method comprises the steps: obtaining a target image of an electric power system; inputting the target image into a pre-trained target detection model to obtain a detection result output by the target detection model; wherein the target detection model is a model which is obtained by training a preset neural network model by adopting a pre-labeled sample target image and is used for carrying out target detection, and the neural network model processes the sample target image to obtain global style features and local semantic features; training is carried out by minimizing the orthogonal loss of the global style features and the local semantic features, so that the target detection precision is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, and in particular to a multi-weather robust target detection method, device and storage medium for power inspection. BACKGROUND

[0002] In the power system, target detection as the core basic task in power inspection aims to realize accurate positioning and identification of key targets in natural images. Thanks to the rapid development of deep learning technology, significant progress has been made in the field of target detection in the past few decades. However, existing deep learning methods usually rely on the assumption that the distribution of training data and test data is consistent, which is often difficult to meet in practical applications. In the changing real environment, the image distribution will change significantly due to weather conditions, lighting conditions, imaging device differences and target appearance changes, etc., which will cause the performance of the deep learning-based detector to degrade seriously in the cross-domain scene, especially in the safety-critical scene, which limits its actual deployment and application.

[0003] Most of the generalization methods focus on aligning the feature space or extracting domain-invariant features, but visual features are extremely sensitive to low-level information such as light, texture, and resolution. When encountering extreme weather (such as rain, fog, snow) or severe sensor noise interference, the performance of the model is still difficult to guarantee, especially in real inspection scenarios, the degradation factors often preferentially affect the edge structure and low-frequency components of the image, thereby significantly weakening the target detection accuracy. SUMMARY

[0004] The present application aims to at least partially solve one of the technical problems in the related art. To this end, one object of the present application is to provide a multi-weather robust target detection method, device and storage medium for power inspection to improve the target detection accuracy.

[0005] According to a first aspect of an embodiment of the present application, a multi-weather robust target detection method for power inspection is provided, the method comprising: obtaining a target image of a power system; inputting the target image into a pre-trained target detection model to obtain a detection result output by the target detection model; wherein the target detection model is a model for target detection obtained by training a pre-set neural network model using pre-labeled sample target images, wherein the neural network model processes the sample target images to obtain global style features and local semantic features, and is trained by minimizing the orthogonal loss of the global style features and the local semantic features.

[0006] Optionally, before the neural network model processes the sample target images, the method further comprises: perform image segmentation on the sample target image to obtain a target region and a background region of the sample target image; perform local enhancement transformation on the target region and global transformation on the background region to obtain an enhanced sample target image.

[0007] Optionally, the neural network model is trained in the following manner: determine a total loss based on the orthogonal loss and a detection loss of a detection result output by the neural network model; adjust model parameters of the neural network model according to the total loss until a preset training completion condition is reached.

[0008] Optionally, the orthogonal loss of the global style feature and the local semantic feature is generated in the following manner :

[0009] wherein, denotes a batch size of the sample target image; and denote a global style feature and a local semantic feature of an i-th sample target image, respectively.

[0010] Optionally, the detection loss includes a classification loss and a bounding box regression loss; the classification loss and the bounding box regression loss are obtained based on a detection result for the local semantic feature.

[0011] Optionally, the total loss further includes a cross-scale consistency loss, which is obtained according to a bounding box prediction result of a multi-scale feature map of the local semantic feature.

[0012] Optionally, the total loss further includes a cross-style consistency loss, which is obtained based on a class probability distribution of a different style image of the target image and a class probability distribution corresponding to the local semantic feature.

[0013] Optionally, the global style feature is obtained by processing a global statistical feature of the sample target image based on a mask attention mechanism.

[0014] According to a second aspect of an embodiment of the present application, a multi-weather robust target detection device for power inspection is provided, and the device comprises: an image acquisition device configured to acquire a target image of a power system; a target detection device configured to input the target image into a pre-trained target detection model to obtain a detection result output by the target detection model; ​The target detection model is obtained by training a preset neural network model by using pre-labeled sample target images, wherein the neural network model processes the sample target images to obtain global style features and local semantic features, and is trained by minimizing the orthogonal loss of the global style features and the local semantic features.

[0015] According to a third aspect of the embodiments of the present application, a computer readable storage medium is provided, and the computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the power inspection multi-weather robust target detection method.

[0016] In the scheme provided by the embodiments of the present application, the global style features are extracted from the sample target images in the process of training the model, and the influence caused by the environment, such as weather change, will be reflected in the global style features. By minimizing the orthogonal loss of the global style features and the local semantic features, the space of the style and the semantic is ensured to be independent of each other, and the global style features are prevented from interfering with the semantic judgment. In this way, the model can detect the local semantic features more accurately, and the trained model can improve the accuracy of target detection.

[0017] Additional aspects and advantages of the present application will be given in part in the following description, become apparent in part from the following description, or be learned by practice of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 is a flowchart of a power inspection multi-weather robust target detection method provided by the embodiments of the present application; Figure 2 is a flowchart of a first model training method provided by the embodiments of the present application; Figure 3 is a flowchart of a second model training method provided by the embodiments of the present application; Figure 4 is a flowchart of a third model training method provided by the embodiments of the present application; Figure 5 is a flowchart of a fourth model training method provided by the embodiments of the present application; Figure 6 is a structural diagram of a power inspection multi-weather robust target detection device provided by the embodiments of the present application. DETAILED DESCRIPTION

[0019] Embodiments of the present application are described below in detail, examples of which are shown in the drawings, wherein the same or similar notations represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by reference to the drawings are exemplary and are intended to explain the present application, and cannot be understood as a limitation of the present application.

[0020] A power inspection multi-weather robust target detection method, a storage medium and an electronic device are described below with reference to the accompanying drawings.

[0021] In one embodiment of the present application, referring to Figure 1 , a power inspection multi-weather robust target detection method is provided, the method comprising the following steps S101-S102.

[0022] S101: obtaining a target image of a power system; S102: inputting the target image into a pre-trained target detection model to obtain a detection result output by the target detection model; The target detection model is a model for target detection obtained by training a preset neural network model using pre-labeled sample target images, wherein the neural network model processes the sample target images to obtain global style features and local semantic features, and is trained by minimizing the orthogonal loss of the global style features and the local semantic features.

[0023] The target image and the sample target image can be images obtained by photographing a scene where the power system exists using a photographing device, which can be a mobile phone, a camera, a camera, etc., and the present application does not limit this.

[0024] For the sample target image, the targets existing in the image can be pre-labeled, including power equipment bodies such as towers, insulators, cross arms, etc., and can also include environmental risk targets such as bird nests and flying wires (such as kites and steel wires wrapped on the line). When labeling, the position and category of the target can be labeled.

[0025] In one embodiment of the present application, the global style features are obtained by processing the global statistical features of the sample target image based on a mask attention mechanism.

[0026] As shown in Figure 3 , in the target detection model, a semantic branch is provided, which extracts local semantic features using a deep convolution or a Transformer structure. The input of the semantic branch can be the original sample target image, or the enhanced image in the following embodiments .

[0027] The specific extraction process can be as follows: first, the feature map of the sample target image is The feature map is processed into (H*W) tokens of length C to form a sequence:

[0028] Wherein, R represents a sample target image set; H and W respectively represent the height and width of the sample target image, and C represents the output channel number; Then, Swin (Shifted Window) / ViT (Vision Transformer)-w and other local window self-attention are used to perform self-attention within the window and shift-attention across the window, and the token sequence is output . Then, the tokens in each s*s window are averaged and pooled, where s is a preset length, to obtain the pooled vector (j=1, 2, …, n) to form a new semantic feature map :

[0029] Statistical aggregation is performed on the high-level feature map corresponding to the entire sample target image to obtain statistical features such as channel mean, variance, and Gram matrix, which are divided into a group of statistical tokens. Then, a mask attention mechanism is used to introduce a learnable mask M in the 0-1 range to the scoring matrix of multi-head self-attention, so as to selectively shield or strengthen the correlation between different statistical tokens. The weighted result output by the soft-max normalization of the mask-modulated attention weight is the global style feature.

[0030] As shown in Figure 4 , the enhanced image is input into the style branch, and the statistical feature is obtained by modeling the global statistical feature. The global style feature is obtained by learnable mask attention, and the predicted mask, that is, the above-mentioned learnable mask, is output. The global style feature and the local semantic feature calculate the orthogonal loss , and the predicted mask is used to calculate the mask loss . For specific calculation methods, see subsequent embodiments, which are not described in detail here.

[0031] In this way, not only the modeling ability of long-range dependence of self-attention is retained, but also the non-style information is filtered through the mask, so as to obtain a discriminative style representation that is orthogonal to the semantic feature and independent of the spatial position. The use of high-level feature maps retains the abstract ability of convolutional features, and the use of statistics eliminates the position information, so that the style representation is more compact and easier to be orthogonal to the semantic representation.

[0032] By setting the above style branch and semantic branch, the original image features in the target image are decomposed into a semantic-related subspace and a domain / style-related subspace, and are represented by local semantic features and global style features, respectively. The core purpose is to let the subsequent target detection network only rely on high-order semantic information directly related to class discrimination, and eliminate low-order style information that drifts with scene, weather, and imaging device changes.

[0033] According to the global style features and the local semantic features, target detection is performed: as shown in Figure 3 The local semantic features are input into the target detection network in the model, the class of the target in the target image is predicted by classification, the class output by the model is the class with the maximum probability in the class probability distribution, and the position of the target in the target image is output by the bounding box regression.

[0034] In this way, the class and position of the detected target can be included in the detection result.

[0035] The above decoupling process can bring three quantifiable changes: Domain discriminability significantly decreases: the accuracy of the local semantic features input into the domain classifier decreases from 86.3% to 21.4% (≈random), indicating that the domain information has been successfully eliminated; Intra-class compactness is improved: t-SNE (t-distributed Stochastic Neighbor Embedding) visualization shows that the uncoupled features are mixed in a "stretching" shape in different domains; after decoupling, samples of the same class are clustered into tight clusters, and the Silhouette coefficient increases from 0.27 to 0.53; Cross-domain detection robustness is improved: in target detection, the cross-domain meanAveragePrecision (cross-domain mAP) is increased by more than 4.8% by using local semantic features instead of original features, and the false detection rate of small targets and adverse weather scenes is reduced by 11.2%.

[0036] These results show that the decoupled semantic features are more "pure", retaining discriminative semantics and eliminating noise related to domain, scale, illumination, and other contexts, thereby enhancing the generalization ability.

[0037] In the scheme provided by the embodiment of the application, global style features are extracted from a sample target image in the process of training the model, so that the influence caused by the environment, for example, weather changes, can be reflected in the global style features; by minimizing the orthogonal loss of the global style features and the local semantic features, it is ensured that the space of the style and the semantic are independent of each other, and the global style features are prevented from interfering with semantic judgment, so that the model can detect the local semantic features more accurately, and the trained model can improve the accuracy of target detection.

[0038] In an embodiment of the application, the sample target image input into the neural network for training can be subjected to data enhancement, specifically including: performing image segmentation on the sample target image to obtain a target region and a background region of the sample target image; performing local enhancement transformation on the target region and global transformation on the background region to obtain an enhanced sample target image.

[0039] Wherein, an instance segmentation model (such as Mask R-CNN or Segment Anything) can be used to generate a binary mask of each target ( ) represents that a pixel belongs to the th target), so as to separate the target region and the background region.

[0040] The target region is represented as:

[0041] (element-wise multiplication to extract target pixels); The background region is represented as: (the remaining region after removing all targets); wherein, represents a sample target image.

[0042] The local enhancement transformation on the target region includes: luminance adjustment (brightness, contrast, hue), geometric perturbation (±5° rotation, small range translation scaling), and texture disturbance (Gaussian noise, local sharpening).

[0043] For the background region, global transformation is used to simulate bad weather, including: fog and haze: large-scale Gaussian blur , wherein is a Gaussian kernel with a variance .

[0044] rain and snow: superimposed particle noise and added motion blur.

[0045] Highlight: global brightness / contrast boost or local high-light overlay.

[0046] Superimpose the enhanced target region and the background region according to the original spatial position to obtain an enhanced sample target image .

[0047] wherein,

[0048] A(·) represents a local enhancement transformation, and G(·) represents a global transformation. Joint mask (assuming no target overlap, ). The data enhancement process is uniformly represented in matrix form as:

[0049] In the above formulas, the letters represent the same parameters have the same meaning.

[0050] The image is segmented into a target region and a background region by using an instance segmentation model, and targeted enhancement is applied respectively, and finally an enhanced image is generated by fusion, simulating target detection in different environments and different styles. Under the limited training data set lacking of extreme weather data, high-precision target detection can still be realized. This method has good computing efficiency, low resource occupation, and is suitable for edge computing devices and actual engineering deployment, meeting the needs of real-time and convenient deployment.

[0051] By explicitly decoupling the style and semantic features of the data enhanced image, the performance degradation problem of target detection under complex weather and variable imaging conditions can be reduced, and the generalization performance of the model can be improved.

[0052] As shown in Figure 2 , the original image I can be a sample target image or a target image input into the model after training for detection. The original image I is input into an instance segmentation model to generate a target mask , and the target region and the background region are separated in the data enhancement module.

[0053] For the target region , target-oriented enhancement is performed, that is, a local enhancement transformation is used, and photometric / geometric / texture transformation is performed to obtain an enhanced target region ; for the background region , global weather simulation is performed, that is, a global transformation is used, including simulating haze / rain / snow / strong light, to obtain an enhanced background region ; after fusing the enhanced image, an enhanced image is obtained.

[0054] In one embodiment of the present application, the neural network model is trained in the following manner: The total loss is determined based on the orthogonal loss and a detection loss of a detection result output by the neural network model; The model parameters of the neural network model are adjusted according to the total loss until a preset training completion condition is reached.

[0055] The manner of adjusting the model parameters can be back propagation, iteration of the model for a preset number of times, or convergence, and the preset training completion condition is reached.

[0056] In one embodiment of the present application, the orthogonal loss of the global style feature and the local semantic feature is generated in the following manner: :

[0057] wherein, represents a batch size of the sample target image; and respectively represent the global style feature and the local semantic feature of the i-th sample target image.

[0058] In one embodiment of the present application, the detection loss includes a classification loss and a bounding box regression loss; the classification loss and the bounding box regression loss are obtained based on a detection result for the local semantic feature.

[0059] As shown in FIG. 1, the local semantic feature is processed through an RPN (Region Proposal Network) network and an ROI (Region of Interest) head to obtain a detection result including a classification prediction and a bounding box regression. Figure 3

[0060] The ROI head is a module responsible for processing the candidate regions generated by the RPN network, which includes a classification subnetwork and a regression subnetwork. The classification subnetwork classifies the target in the candidate region, and the regression subnetwork finely adjusts the target position to optimize the accuracy of the bounding box coordinates.

[0061] In this case, the classification loss is obtained by calculating the cross-entropy , and the bounding box regression loss is calculated as follows: The bounding box regression loss is calculated using a Smooth L1 loss function:

[0062]

[0063] ​​​in, Represents the number of candidate regions in the target sample image; C is the number of target categories; is the classification prediction, indicating the The candidate region belongs to Probability of the class; Represents the number of positive samples, where a positive sample is a candidate area whose intersection-over-union ratio with any real bounding box is greater than the preset intersection-over-union ratio threshold and is marked as a foreground; because only positive samples need to be regressed, Chinese Perform normalization. is the Smooth L1 loss function; represents the predicted bounding box, represents the ground-truth bounding box. Indicates the A valid label for each candidate region in category C (including background).

[0064] The Smooth L1 loss function is used when the error is small (less than the preset threshold ), uses L2 penalty to maintain smoothness when the error is large, degenerates to L1 to suppress outliers, and is more robust to bounding box regression than pure L2.

[0065] The predicted bounding box is the bounding box representing the target location predicted by the model during the bounding box regression process; the true bounding box is the pre-annotated bounding box.

[0066] In one embodiment of the present invention, the total loss further includes a cross-scale consistency loss, which is obtained based on the bounding box prediction results of the multi-scale feature map of the local semantic features.

[0067] Cross-scale consistency loss It is used to constrain the consistency of detection results of the same target on feature maps of different scales. It is defined as follows:

[0068] in, Indicates the total number of target instances involved in calculating the cross-scale consistency loss; 、 For the same goal, and Predicted bounding box on scale feature map; Indicates calculating the intersection-over-union ratio.

[0069] like Figure 3The target detection network shown extracts multi-scale feature maps from local semantic features. The extraction method can be a feature pyramid network (FPN, Feature Pyramid Networks), which combines high-level features and low-level features in local semantic features to construct multi-scale feature maps. Cross-scale bounding box prediction obtains the predicted bounding boxes on each scale feature map, so as to calculate the cross-scale consistency loss based on the intersection over union of the predicted bounding boxes.

[0070] In the convolutional network, the low-level features are the pixel-level details such as texture and edge extracted by the first few layers, and the high-level features are the abstract semantic information such as object shape and category obtained by deep aggregation.

[0071] In the power inspection scene, "context" not only refers to background semantics, that is, the appearance distribution of the target is affected by the information such as spatial geometry, semantic co-occurrence, scene globality and scale hierarchy, but also implies the key condition of scale hierarchy: the appearance and geometric information of the same target corresponding to the FPN feature map at different shooting distances or resolutions are significantly different. If the network does not explicitly align these differences, when the scene migration causes the scale distribution to drift, it often produces positioning drift and missed detection - this is the "context dependence" emphasized in this work.

[0072] To weaken this dependence, a cross-scale consistency loss is introduced in the training stage to directly constrain the predicted boxes of the same target on different scale feature maps to be consistent, so as to maximize the intersection over union by minimizing the loss function, and explicitly align the geometric representation under different scale contexts, and internalize the "scale" context as an optimizable geometric consistency constraint in the feature space.

[0073] In an embodiment of the present application, the total loss further includes a cross-style consistency loss, which is obtained based on the class probability distribution of the target image and the class probability distribution corresponding to the local semantic features of the different style images.

[0074] Let the predicted class distribution of different style images of the same scene generated by data augmentation be and As shown in Figure 5 , different style images are generated by the original image I, Different style images refer to enhanced images generated by different enhancement transformation methods such as photometric adjustment and geometric perturbation. For example, for the same original image I, a photometrically adjusted image X for I is generated, and a geometrically perturbed image Y for I is generated, then X and Y are different style images of I.

[0075] In the aforementioned embodiment, any of the local enhancement transformation and global transformation methods can be used to generate images of different styles, or a combination of any several of the local enhancement transformation and global transformation methods can be used for generation, and the embodiment of the present invention is not limited to this.

[0076] The predicted category distribution refers to the probability vector formed by the predicted probability of belonging to each category in the C categories after normalization in the classification prediction, such as Figure 5 As shown, it can be obtained based on the detection network, and the detection network can be a model that has been trained in the prior art, such as the YOLO model.

[0077] Figure 3 In the category probability distribution, the original prediction can be obtained, that is, the predicted category of the sample target image, or the probability vector of the category distribution can be used to represent the probability of belonging to different categories.

[0078] Defining cross-style consistency loss In the form of KL divergence:

[0079] in, is the number of sample target images, is the number of target categories, and Respectively represent the target in different style images of the i-th sample target image belongs to the category The predicted probability of , and the predicted category of the sample target image.

[0080] This loss constrains the model to maintain consistent predictions under style changes, thereby improving the robustness of the detection model to image style changes.

[0081] In one embodiment, in order to further clarify the model's attention to the target area, a learnable mask attention supervision is proposed. The target detection model outputs the predicted mask of the target sample image and calculates the mask loss with the real target mask. , perform binary cross entropy supervision:

[0082] in, represents the set of pixel locations in the sample target image; Represents the true mask pixel value, which is the pixel value of the pre-annotated mask; Represents the predicted mask pixel value, which is the pixel value of the predicted mask output by the target detection model.

[0083] In this case, binary cross entropy loss can be added to the total loss for training.

[0084] In one embodiment, to reduce the interference of low-level style changes on high-level semantic features, a statistical difference loss of low-level convolutional features can be added to the total loss as an auxiliary loss.

[0085] Statistical difference loss is defined as follows:

[0086] where, through global statistical feature modeling shown in formula (1), the statistical mean and variance of two different style images of the same sample target image in low-level convolutional features are respectively Figure 4 and , that is, the low-level style statistical features, so as to calculate the statistical difference loss .

[0087] In summary, the total loss can be set as:

[0088] The loss weight coefficients , , , , , , are hyperparameters, and the meanings of other parameters are consistent with the parameters of the same letter in the foregoing embodiments, which are adjusted according to actual needs to balance the contributions of various optimization objectives. As shown in formula (2), the total loss is the sum of the cross-style consistency loss, the mask loss, the orthogonal loss, the statistical difference loss, the classification loss, the regression loss, and the cross-scale consistency loss. The parameters are updated through the back propagation of the total loss to complete the model training process. Figure 5

[0089] Corresponding to the method embodiments described above, in one embodiment, referring to formula (3), a power inspection multi-weather robust target detection device is also provided, and the device comprises: Figure 6 An image acquisition device 601 is configured to acquire a target image of a power system. A target detection device 602 is configured to input the target image into a pre-trained target detection model to obtain a detection result output by the target detection model. The target detection model is a model for target detection obtained by training a preset neural network model using pre-labeled sample target images, wherein the neural network model processes the sample target images to obtain global style features and local semantic features, and is trained by minimizing the orthogonal loss of the global style features and the local semantic features.

[0090] ​​​According to a third aspect of the embodiments of the present application, a computer readable storage medium is provided, having stored thereon a computer program, which, when executed by a processor, implements the power inspection multi-weather robust target detection method according to any of the above embodiments.

[0091] It should be noted that the logic and / or steps represented in the flowcharts and / or described herein can be considered as a sequence of executable instructions stored in a computer readable medium, which can be executed by an instruction execution system, apparatus or device, such as a computer-based system, a processor-based system, or other system that can fetch the instructions from the instruction execution system, apparatus or device and execute the instructions, or a combination of the above. For the purposes of this specification, a "computer readable medium" can be any apparatus that can contain, store, communicate, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus or device. The computer readable medium can be a computer readable storage medium or a computer readable signal medium. The computer readable storage medium can include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or a propagation medium. The computer readable signal medium can include, but is not limited to, a computer readable medium that can be a propagated data signal with the computer readable program code embodied therein, for example, in baseband or as part of a carrier wave of a data stream. The computer readable program code can be transmitted on a computer readable medium such as a computer readable storage medium or a computer readable signal medium. The computer readable medium can be any available medium or a combination of media that can be accessed by a general purpose or special purpose computer system. The computer readable medium can be a computer readable storage medium or a computer readable signal medium. By way of example, and not limitation, the computer readable medium can comprise: an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device; a computer readable storage medium, comprising: a magnetic storage device, an optical storage device, a solid state memory device, etc.; or a computer readable signal medium, comprising: a propagated data signal with the computer readable program code embodied therein, for example, in baseband or as part of a carrier wave of a data stream transmitted over a transmission medium (wireless or wired), such as a coherent optical or electromagnetic carrier wave traveling through space, between a source and a destination on the computer readable signal medium. The program code can be transmitted directly, via a cable, wire, or other physical transmission medium, or via a computer readable signal medium capable of storing the computer readable program code for a period of time and transmitting it over a

[0092] It should be understood that portions of the present application can be implemented in hardware, software, firmware, or combinations thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented in hardware, and in another embodiment, any of the following technologies known in the art or their combinations can be used: discrete logic circuitry having logic gates for implementing logic functions on data signals, application specific integrated circuits having appropriate combinational logic gates, programmable gate arrays (PGA), field programmable gate arrays (FPGA), etc.

[0093] In the description of the present application, the description of the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the description of the present application, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0094] In the description of the present application, it should be understood that the orientation or positional relationship indicated by the terms "center", "longitudinal", "transverse", "length", "width", "thickness", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", "clockwise", "counterclockwise", "axial", "radial", "circumferential" and the like is based on the orientation or positional relationship shown in the drawings, and is only for the purpose of facilitating the description of the present application and simplifying the description, and does not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation of the present application.

[0095] In addition, the terms "first", "second" are only for the purpose of description, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the technical features indicated. Therefore, the features defined with "first", "second" can explicitly or implicitly include at least one of the features. In the description of the present application, the meaning of "a plurality of" is at least two, for example, two, three, etc., unless otherwise specifically limited.

[0096] In the present application, unless otherwise specifically defined and limited, the terms "mounting", "connection", "connection", "fixing" and the like should be understood in a broad sense, for example, it can be fixed connection, or detachable connection, or integral; it can be mechanical connection, or electrical connection; it can be directly connected, or indirectly connected through an intermediate medium, it can be the internal communication of two elements or the interaction relationship between two elements, unless otherwise specifically limited. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0097] In the present application, unless otherwise explicitly specified and limited, a first feature is "on" or "under" a second feature can mean that the first and second features are in direct contact, or the first and second features are in indirect contact through an intermediate medium. Moreover, the first feature is "over", "above" and "on top of" the second feature can mean that the first feature is directly above or obliquely above the second feature, or simply means that the first feature is horizontally higher than the second feature. The first feature is "under", "below" and "underneath" the second feature can mean that the first feature is directly below or obliquely below the second feature, or simply means that the first feature is horizontally lower than the second feature.

[0098] Although the embodiments of the present application have been shown and described above, it is to be understood that the above-described embodiments are exemplary only, and are not to be construed as limiting the present application, and that changes, modifications, substitutions and variations can be made thereto by those skilled in the art without departing from the scope of the present application.

Claims

1. A multi-weather robust target detection method for power inspection, characterized in that: The method comprises: Acquire target images of the power system; Inputting the target image into a pre-trained target detection model to obtain a detection result output by the target detection model; Among them, the target detection model is: a model for target detection obtained by training a preset neural network model using pre-labeled sample target images, wherein the neural network model processes the sample target images to obtain global style features and local semantic features, and trains by minimizing the orthogonal loss of the global style features and the local semantic features.

2. The method according to claim 1, characterized in that Before the neural network model processes the sample target image, the method further includes: Performing image segmentation on the sample target image to obtain a target area and a background area of ​​the sample target image; A local enhancement transformation is performed on the target area, and a global transformation is performed on the background area to obtain an enhanced sample target image.

3. The method according to claim 1, characterized in that The neural network model is trained as follows: Determine a total loss based on the orthogonal loss and the detection loss of the detection result output by the neural network model; The model parameters of the neural network model are adjusted according to the total loss until a preset training completion condition is reached.

4. The method according to claim 3, characterized in that The orthogonal loss of the global style feature and the local semantic feature is generated as follows : in, represents the batch size of the sample target image; and Respectively represent The global style features and local semantic features of sample target images.

5. The method according to claim 3, characterized in that The detection loss includes a classification loss and a bounding box regression loss; the classification loss and the bounding box regression loss are obtained based on the detection results of the local semantic features.

6. The method according to claim 3, characterized in that The total loss also includes a cross-scale consistency loss, which is obtained based on the bounding box prediction results of the multi-scale feature map of the local semantic features.

7. The method according to claim 3, characterized in that The total loss also includes a cross-style consistency loss, which is obtained based on the category probability distribution of different style images of the target image and the category probability distribution corresponding to the local semantic feature.

8. The method according to claim 1, characterized in that The global style feature is obtained by processing the global statistical features of the sample target image based on a mask attention mechanism.

9. A multi-weather robust target detection device for power inspection, characterized in that: The device comprises: An image acquisition device, used for acquiring a target image of the power system; An object detection device, configured to input the object image into a pre-trained object detection model to obtain a detection result output by the object detection model; Among them, the target detection model is: a model for target detection obtained by training a preset neural network model using pre-labeled sample target images, wherein the neural network model processes the sample target images to obtain global style features and local semantic features, and trains by minimizing the orthogonal loss of the global style features and the local semantic features.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the multi-weather robust target detection method for power inspection according to any one of claims 1 to 8 is implemented.

Images