A method and apparatus for enhancing images of power plant inspections based on semantic consistency and reinforcement learning

Abstract

This application relates to a method and apparatus for enhancing power line inspection images based on semantic consistency and reinforcement learning. The method includes: acquiring an original power line inspection image; inputting the original power line inspection image into a semantic segmentation network to extract noise components from the original power line inspection image; collecting component-carrying information of the noise components to determine the information loss degree of the denoised original power line inspection image; collecting occlusion information of the noise components to determine the artifact introduction degree of the denoised original power line inspection image; determining associated images of the original power line inspection image and calculating the linkage observation degree between the denoised original power line inspection image and the associated images; determining the denoising degree of the denoised original power line inspection image based on the information loss degree, artifact introduction degree, and linkage observation degree; and enhancing the original power line inspection image according to the denoising degree. This method can improve the reliability of image enhancement processing.

Description

Technical Field

[0001] This application relates to the field of image enhancement technology, and in particular to a method and apparatus for enhancing images of power generator inspections based on semantic consistency and reinforcement learning. Background Technology

[0002] With the development of smart grids, high-quality drone inspection images are crucial for timely detection of equipment hazards. Power drone inspection images refer to the optimization of images of power equipment collected by drones or robots using image processing technology, which is a key means to improve the accuracy of defect identification and ensure the safe operation of the power grid.

[0003] In related technologies, existing image enhancement methods often adopt a "one-size-fits-all" strategy, performing global denoising on the entire image without selectively judging the characteristics of noise. This approach does not fully consider the similarity between some texture details and noise in power scenes, which can easily lead to the accidental deletion of critical equipment edge features or micro-crack information while removing noise. This results in the loss of effective information or the introduction of artificial artifacts, which reduces the reliability of subsequent defect detection and leads to inaccurate final detection results. Summary of the Invention

[0004] Therefore, it is necessary to provide a method, apparatus, computer device, computer-readable storage medium, and computer program product for enhancing power turbine inspection images that can distinguish between noise and non-noise information based on semantic guidance and evaluate the suitability of denoising based on semantic consistency and reinforcement learning, in order to address the above-mentioned technical problems.

[0005] Firstly, this application provides a method for enhancing power machine inspection images based on semantic consistency and reinforcement learning, including:

[0006] The original power line inspection images are obtained and input into a semantic segmentation network to extract noise components from the original power line inspection images.

[0007] Collect the component-carrying information of the noise components, and determine the information loss degree of the original power inspection image after denoising based on the component-carrying information;

[0008] Obstruction information of noise components is collected, and the artifact introduction degree of the original power inspection image after denoising is determined based on the obstruction information; the artifact introduction degree is used to quantify the risk of introducing new artifacts after denoising the original power inspection image.

[0009] The associated images of the original power inspection images are identified, and the linkage observation degree between the denoised original power inspection images and the associated images is calculated; the linkage observation degree is used to quantitatively characterize the consistency and information complementarity between the denoised original power inspection images and the associated images.

[0010] Based on the information missing degree, the artifact introduction degree, and the linkage observation degree, the denoising degree of the original power inspection image after denoising is determined, and the original power inspection image is enhanced according to the denoising degree.

[0011] In one embodiment, the component-carrying information includes defect information of the original power line inspection image; the step of collecting the component-carrying information of the noise components and determining the information loss degree of the denoised original power line inspection image based on the component-carrying information includes:

[0012] Extract the non-noise components from the original power line inspection images before and after denoising;

[0013] The non-noise parts of the original power inspection images before and after denoising are compared to determine the non-noise similarity of the non-noise components before and after denoising.

[0014] The topological structures of the original power inspection images before and after denoising are compared to determine the structural similarity of the topological structures.

[0015] Based on the defect information, determine the degree of assistance the noise component provides in identifying the defect information;

[0016] Based on the non-noise similarity, the structural similarity, and the recognition assistance, the information loss degree of the original power inspection image after denoising is determined.

[0017] In one embodiment, determining the degree of assistance of the noise component in identifying the defect information based on the defect information includes:

[0018] In cases where defect information is not identified only under noise components, it is determined whether the noise components participated in the identification in the historical identification data of the defect information;

[0019] When noise components are involved in identification, it is determined whether the noise components enhance the features of defect information;

[0020] In the case of enhancing defect information with noise components, the average recognition accuracy under noise components is taken as the noise accuracy, and the average recognition accuracy under non-noise components is taken as the non-noise accuracy.

[0021] When the noise accuracy is greater than the non-noise accuracy, the noise assistance degree of the defect information is determined based on the difference between the identification participation rate of noise components, the non-noise accuracy, and the noise accuracy in historical identification data.

[0022] In the absence of features that enhance defect information in the noise component, the information aiding degree of the noise component is determined based on the component-carrying information of the noise component.

[0023] Based on the noise assistance degree and the information assistance degree, the identification assistance degree of the noise component for the defect information is determined.

[0024] In one embodiment, determining the information assistance degree of the noise component based on the component-carrying information of the noise component includes:

[0025] The defect information in the component-carrying information is marked as an application defect;

[0026] Determine the percentage of defects in the application defects, the average application rate of component-carrying information in defect identification, and the average time difference in defect information identification before and after the application of component-carrying information;

[0027] The information assistance degree of the noise component is calculated based on the proportion of the number of defects, the average application rate, and the average duration difference.

[0028] In one embodiment, the step of acquiring occlusion information of noise components and determining the artifact introduction degree of the original power inspection image after denoising based on the occlusion information includes:

[0029] Determine the occlusion region in the occlusion information and calculate the distribution uniformity of noise components in the occlusion region;

[0030] Determine the content that is obscured corresponding to the obscured area, and calculate the difficulty of restoring the content of the obscured area;

[0031] The historical artifact introduction probability of the noise component is obtained, and the artifact introduction degree of the original power inspection image after denoising is calculated based on the historical artifact introduction probability, the distribution uniformity, and the content restoration difficulty.

[0032] In one embodiment, determining the associated image of the original power line inspection image and calculating the correlation observation degree between the denoised original power line inspection image and the associated image includes:

[0033] Calculate the feature matching degree between the original power inspection image after denoising and the associated image after denoising;

[0034] Calculate the first feature preservation degree between the original power inspection image after denoising and the second feature preservation degree between the associated image and the denoised associated image; wherein, the feature preservation degree is used to quantify the degree of information preservation between the denoised image and the image before denoising;

[0035] When noise components are compared to obtain component-carrying information, the comparison requirement rate of component-carrying information is determined.

[0036] Based on the feature matching degree, the first feature preservation degree, the second feature preservation degree, and the comparison demand rate, the linkage observation degree is determined.

[0037] Secondly, this application also provides a power motor inspection image enhancement device based on semantic consistency and reinforcement learning, comprising:

[0038] The extraction module is used to acquire the original power inspection image, input the original power inspection image into the semantic segmentation network, and extract the noise components in the original power inspection image;

[0039] The determination module is used to collect the component-carrying information of the noise components and determine the information missing degree of the original power inspection image after denoising based on the component-carrying information.

[0040] The determining module is also used to collect occlusion information of noise components and determine the artifact introduction degree of the original power inspection image after denoising based on the occlusion information; the artifact introduction degree is used to quantify the risk of introducing new artifacts after denoising the original power inspection image.

[0041] The determining module is further configured to determine the associated images of the original power inspection image and calculate the linkage observation degree between the denoised original power inspection image and the associated images; the linkage observation degree is used to quantitatively characterize the consistency and information complementarity availability between the denoised original power inspection image and the associated images.

[0042] An enhancement module is used to determine the denoising degree of the original power inspection image after denoising based on the information missing degree, the artifact introduction degree, and the linkage observation degree, and to perform enhancement processing on the original power inspection image according to the denoising degree.

[0043] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0044] The original power line inspection images are obtained and input into a semantic segmentation network to extract noise components from the original power line inspection images.

[0045] Collect the component-carrying information of the noise components, and determine the information loss degree of the original power inspection image after denoising based on the component-carrying information;

[0046] Obstruction information of noise components is collected, and the artifact introduction degree of the original power inspection image after denoising is determined based on the obstruction information; the artifact introduction degree is used to quantify the risk of introducing new artifacts after denoising the original power inspection image.

[0047] The associated images of the original power inspection images are identified, and the linkage observation degree between the denoised original power inspection images and the associated images is calculated; the linkage observation degree is used to quantitatively characterize the consistency and information complementarity between the denoised original power inspection images and the associated images.

[0048] Based on the information missing degree, the artifact introduction degree, and the linkage observation degree, the denoising degree of the original power inspection image after denoising is determined, and the original power inspection image is enhanced according to the denoising degree.

[0049] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0050] The original power line inspection images are obtained and input into a semantic segmentation network to extract noise components from the original power line inspection images.

[0051] Collect the component-carrying information of the noise components, and determine the information loss degree of the original power inspection image after denoising based on the component-carrying information;

[0052] Obstruction information of noise components is collected, and the artifact introduction degree of the original power inspection image after denoising is determined based on the obstruction information; the artifact introduction degree is used to quantify the risk of introducing new artifacts after denoising the original power inspection image.

[0053] The associated images of the original power inspection images are identified, and the linkage observation degree between the denoised original power inspection images and the associated images is calculated; the linkage observation degree is used to quantitatively characterize the consistency and information complementarity between the denoised original power inspection images and the associated images.

[0054] Based on the information missing degree, the artifact introduction degree, and the linkage observation degree, the denoising degree of the original power inspection image after denoising is determined, and the original power inspection image is enhanced according to the denoising degree.

[0055] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:

[0056] The original power line inspection images are obtained and input into a semantic segmentation network to extract noise components from the original power line inspection images.

[0057] Collect the component-carrying information of the noise components, and determine the information loss degree of the original power inspection image after denoising based on the component-carrying information;

[0058] Obstruction information of noise components is collected, and the artifact introduction degree of the original power inspection image after denoising is determined based on the obstruction information; the artifact introduction degree is used to quantify the risk of introducing new artifacts after denoising the original power inspection image.

[0059] The associated images of the original power inspection images are identified, and the linkage observation degree between the denoised original power inspection images and the associated images is calculated; the linkage observation degree is used to quantitatively characterize the consistency and information complementarity between the denoised original power inspection images and the associated images.

[0060] Based on the information missing degree, the artifact introduction degree, and the linkage observation degree, the denoising degree of the original power inspection image after denoising is determined, and the original power inspection image is enhanced according to the denoising degree.

[0061] The aforementioned method, apparatus, computer equipment, computer-readable storage medium, and computer program product for enhancing power grid inspection images based on semantic consistency and reinforcement learning first acquire the original power grid inspection image and input it into a semantic segmentation network to extract noise components from the original power grid inspection image. Next, the component-carrying information of the noise components is collected, and the information loss degree of the denoised original power grid inspection image is determined based on this information. Then, the occlusion information of the noise components is collected, and the artifact introduction degree of the denoised original power grid inspection image is determined based on this information. Finally, the associated images of the original power grid inspection image are identified, and the linkage observation degree between the denoised original power grid inspection image and the associated images is calculated. Based on the information loss degree, artifact introduction degree, and linkage observation degree, the denoising degree of the denoised original power grid inspection image is determined, and the original power grid inspection image is enhanced based on the denoising degree. In this way, the semantic segmentation network accurately separates noise components from non-noise components and appropriately introduces information loss and artifact introduction, quantifying the risks of effective information loss and artifact generation that may occur during the denoising process. By combining the observational accuracy with spatiotemporally correlated images to evaluate the cross-view consistency after denoising, the denoising score is dynamically calculated by integrating all three factors. This mechanism effectively avoids the problems of defect feature erasure or artificial artifact introduction caused by the "one-size-fits-all" strategy of traditional denoising methods, significantly improves the structural integrity and defect identifiability of power generator inspection images in complex noisy environments, and enhances the reliability of power generator inspection image enhancement. Attached Figure Description

[0062] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0063] Figure 1 This is an application environment diagram of a power turbine inspection image enhancement method based on semantic consistency and reinforcement learning in one embodiment;

[0064] Figure 2 This is a flowchart illustrating a power generator inspection image enhancement method based on semantic consistency and reinforcement learning in one embodiment.

[0065] Figure 3 This is a structural block diagram of a power generator inspection image enhancement device based on semantic consistency and reinforcement learning in one embodiment.

[0066] Figure 4 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0067] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0068] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0069] The image enhancement method for power generator inspection based on semantic consistency and reinforcement learning provided in this application can be applied to, for example... Figure 1In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104 or located on the cloud or other network servers. Terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, drones, low-altitude aircraft, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, projection devices, etc. Portable wearable devices can include smartwatches, smart bracelets, head-mounted devices, etc. Head-mounted devices can be virtual reality (VR) devices, augmented reality (AR) devices, smart glasses, etc. Server 104 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0070] In one exemplary embodiment, such as Figure 2 As shown, a method for enhancing images of electric motor inspections based on semantic consistency and reinforcement learning is presented, and this method is applied to... Figure 1 Taking terminal 102 as an example, the explanation includes the following steps 202 to 210. Wherein:

[0071] Step 202: Obtain the original power inspection image, input the original power inspection image into the semantic segmentation network, and extract the noise components in the original power inspection image.

[0072] For example, raw power inspection images are obtained by drones or inspection robots as input images and fed into a trained semantic segmentation network (such as an improved U-Net or DeepLabV3+ architecture) to perform pixel-level parsing of the input images and extract noise and non-noise components from the raw power inspection images.

[0073] In some embodiments, the semantic segmentation network extracts multi-scale features to divide the original power inspection image into noise components that represent interference and non-noise components that represent effective structures. The noise components include, but are not limited to, cloud cover, shadows caused by uneven lighting, sensor thermal noise, and background clutter. The non-noise components include, but are not limited to, the main body of power equipment such as conductors, insulators, towers, and fittings, as well as their key defect features.

[0074] In some embodiments, the semantic segmentation network outputs a noise mask and a non-noise mask with semantic labels, wherein the noise mask includes noise information such as noise category.

[0075] Step 204: Collect the component-carrying information of the noise components, and determine the information missing degree of the original power inspection image after denoising based on the component-carrying information.

[0076] Among these, the information carried by the components are physical or geometric cues inherent in the noise itself that can be utilized by computer vision algorithms. For example, the concentration gradient of fog noise carries the depth (distance) information of the scene, the direction of motion blur carries the relative velocity vector of objects, and the tilt angle of rain lines carries wind direction information.

[0077] Optionally, the component-carrying information of the noise components is collected, and the information loss degree of the original power inspection image after denoising is determined based on the component-carrying information.

[0078] Step 206: Collect occlusion information of noise components and determine the artifact introduction degree of the original power inspection image after denoising based on the occlusion information.

[0079] Among them, the artifact introduction degree is used to assess the risk of introducing new artificial distortions (artifacts) after denoising the original power turbine inspection image.

[0080] In some embodiments, artifact introduction measures the likelihood that, during the removal of original noise (such as rain, fog, occlusion), non-realistic visual artifacts (such as edge breaks, block effects, false textures, etc.) are mistakenly generated in the image due to improper algorithm processing.

[0081] For example, occlusion information of noise components is collected, and the artifact introduction degree of the original power inspection image after denoising is determined based on the occlusion information.

[0082] Step 208: Determine the associated images of the original power inspection images and calculate the correlation observation degree between the original power inspection images and the associated images after denoising.

[0083] Among them, the linkage observation degree is used to evaluate the comprehensive collaborative quality between the original power inspection image after denoising and its associated images with spatiotemporal correspondence in terms of feature consistency, information preservation, and availability of comparative information. It refers to the quantitative index determined based on the feature matching degree, feature preservation degree, and comparison demand rate between the original power inspection image after denoising and its associated images with spatiotemporal correspondence, which is used to characterize the cross-view collaborative consistency and information complementarity availability between the original power inspection image after denoising and its associated images.

[0084] In some embodiments, the linkage observation degree measures whether, after denoising the original power inspection image, the image and related images (such as previous and next frames of a video, or another perspective image from binocular stereo vision) can form a logically consistent, information-complementary, and cross-validation linkage observation relationship.

[0085] Optionally, the associated images of the original power inspection images are determined, and the correlation observation degree between the denoised original power inspection images and the associated images is calculated.

[0086] Step 210: Based on the information missing degree, artifact introduction degree, and linkage observation degree, determine the denoising degree of the original power inspection image after denoising, and perform enhancement processing on the original power inspection image according to the denoising degree.

[0087] For example, the information missing degree, artifact introduction degree, and linkage observation degree are normalized and weighted summation is performed to obtain the denoising degree D of the original power inspection image after denoising. The denoising degree D is mapped to the dynamic parameters of the filtering algorithm (e.g., the smoothing parameter of guided filtering or the spatial variance of bilateral filtering). Using the mapped dynamic parameters, the original power inspection image is subjected to pixel-by-pixel adaptive filtering (such as guided filtering) to generate an enhanced power training image that suppresses noise and retains key defect features.

[0088] Specifically, when the denoising level is greater than a preset value, the dynamic parameter is determined to be smaller to perform strong denoising; when the denoising level is less than a preset value, the dynamic parameter is determined to be larger to preserve image details.

[0089] In some embodiments, the mapping method is linear mapping or piecewise linear mapping, the denoising value range is set to [Dmin, Dmax] (e.g., [0, 1]), and the filtering parameters (e.g., smoothing parameters) are set. The effective range of ) is [ min, [max], establish an inverse proportional mapping relationship = max-( max- min)×D, thus, when D=1 (extremely high noise reduction requirement), = min (strongest smoothing), when D=0 (no noise reduction required), = max(weakest smoothing / original image preservation).

[0090] In some embodiments, the adaptive filtering process calculates specific smoothing parameters on a pixel or local block basis, based on the corresponding denoising degree. For example, when using guided filtering, the regularization parameter in the formula... It is no longer a constant, but a function related to the image position (x, y). (x, y) = f(D(x, y)). During the filter's operation, the corresponding filter is applied to each pixel. (x, y), thereby achieving an adaptive effect in the noise processing process.

[0091] The aforementioned power line inspection image enhancement method based on semantic consistency and reinforcement learning acquires the original power line inspection image, inputs it into a semantic segmentation network to extract noise components, collects component-carrying information of the noise components, and determines the information loss degree of the denoised original power line inspection image based on this information. It also collects occlusion information of the noise components and determines the artifact introduction degree of the denoised original power line inspection image based on this information. Furthermore, it identifies associated images of the original power line inspection image and calculates the linkage observation degree between the denoised original power line inspection image and associated images. Based on the information loss degree, artifact introduction degree, and linkage observation degree, it determines the denoising degree of the denoised original power line inspection image and performs enhancement processing based on the denoising degree. In this way, the semantic segmentation network accurately separates noise and non-noise components and introduces information loss and artifact introduction, respectively quantifying the risks of effective information loss and artifact generation that may occur during the denoising process. The linkage observation degree is combined with spatiotemporally associated images to evaluate the cross-view consistency after denoising, and finally, the denoising degree is dynamically calculated by integrating all three factors. This mechanism effectively avoids the problems of defect feature erasure or artificial artifact introduction caused by the "one-size-fits-all" strategy of traditional denoising methods, significantly improves the structural integrity and defect identifiability of power generator inspection images in complex noise environments, and enhances the reliability of power generator inspection image enhancement.

[0092] In an exemplary embodiment, the component-carrying information includes defect information of the original power inspection image; collecting component-carrying information of noise components and determining the information missing degree of the original power inspection image after denoising based on the component-carrying information includes: extracting non-noise parts from the original power inspection image before and after denoising; comparing the non-noise parts of the original power inspection images before and after denoising to determine the non-noise similarity of the non-noise components before and after denoising; comparing the topological structures of the original power inspection images before and after denoising to determine the structural similarity of the topological structures; determining the recognition assistance degree of noise components for defect information based on defect information; and determining the information missing degree of the original power inspection image after denoising based on non-noise similarity, structural similarity, and recognition assistance degree.

[0093] In practical implementation, noise reduction techniques are used to identify noise components. These techniques are applied to the original power line inspection images to obtain denoised original power line inspection images. The non-noise portions of the original power line inspection images before and after denoising are compared to determine the non-noise similarity of the non-noise components. The topological structures of the original power line inspection images before and after denoising are compared to determine the structural similarity of the topological structures. Based on defect information, the degree of assistance of noise components in defect information recognition is determined. A weighted fusion model is constructed, and the non-noise similarity, structural similarity, and recognition assistance are normalized and then linearly weighted to obtain the information missing degree.

[0094] In some embodiments, the denoising technique is a set of algorithms adapted to the current noise type from a pre-set library in the system, including but not limited to traditional filtering algorithms based on non-local means, deep learning-based DnCNN networks, or generative adversarial network (GAN) repair models. The system automatically matches the corresponding denoising algorithm according to the noise type (such as rain / fog, Gaussian noise, or motion blur) identified by the semantic guidance module.

[0095] In some embodiments, the non-noise component is the main body region of the power equipment (e.g., conductors, insulators, tower hardware, etc.) extracted through semantic segmentation. The deep feature vectors of the main body region of the power equipment before and after denoising are extracted, and the cosine similarity or structural similarity index between the two is calculated as the non-noise similarity.

[0096] Among them, the deep feature vector is extracted by a pre-trained deep convolutional neural network. The total non-noise component region of the image is input into the network, and after multiple layers of convolution, pooling and nonlinear activation, the pixel-level spatial information is transformed into a high-dimensional abstract feature representation (i.e., deep feature vector) to characterize texture, shape and semantic information.

[0097] In some embodiments, the topology refers to the geometric connections and spatial skeleton of power equipment, such as the continuity and connectivity of conductors and the node arrangement order of insulator strings. A thinning algorithm is used to extract the equipment skeleton and construct a graph model. The structural changes are quantified by calculating the edit distance or the maximum common subgraph ratio of the graphs before and after denoising.

[0098] In some embodiments, the component-carrying information includes defect information, which is information such as defect type and defect identification steps from a fault heatmap output by a pre-trained defect detection network or a historical defect annotation library.

[0099] In the above embodiments, by introducing historical image comparison and multi-dimensional similarity evaluation mechanisms, the system effectively solves the problem of missed defects caused by blind smoothing in traditional denoising methods. Simultaneously, by combining the defect distribution analysis of the current image, the system identifies that noisy regions have a high degree of assistance in crack identification. Therefore, the calculated high "information missing degree" is immediately fed back to the enhancement module, prompting the system to adopt a conservative enhancement strategy or switch to a detail-preserving algorithm in that region. This removes background interference while completely preserving millimeter-level crack features. This adaptive mechanism based on quantitative evaluation significantly improves the detection rate of minute defects in complex environments, avoids misjudgments caused by excessive denoising, and provides a more reliable data foundation for power grid safety diagnosis.

[0100] In an exemplary embodiment, determining the degree of assistance of noise components in the identification of defect information based on defect information includes: when the defect information is not only identified under noise components, determining whether noise components participate in the identification in the historical identification data of the defect information; when noise components participate in the identification, determining whether noise components enhance the features of the defect information; when noise components enhance the features of the defect information, taking the average identification accuracy under noise components as the noise accuracy and the average identification accuracy under non-noise components as the non-noise accuracy; when the noise accuracy is greater than the non-noise accuracy, determining the degree of noise assistance of the defect information based on the difference between the identification participation rate of noise components, the non-noise accuracy, and the noise accuracy in the historical identification data; when noise components do not enhance the features of the defect information, determining the degree of information assistance of noise components based on the component-carrying information of the noise components; and determining the degree of assistance of noise components in the identification of defect information based on the degree of noise assistance and the degree of information assistance.

[0101] In practice, it is determined whether the defect information is identified only under noise components. If it is not identified only under noise components, it is determined whether noise components were involved in the historical identification of the defect information. Identification only under noise components means that when the noise components in the input image are removed or suppressed, the original defect features disappear, become blurred to the point of being undetectable, or the confidence level drops below the threshold.

[0102] Specifically, taking the detection of micro-stress cracks on the surface of power transmission lines as an example, under certain lighting conditions, the reflectivity of micro-cracks on the conductor surface is extremely low. Their gray-level gradient change in a clean image is less than the perception threshold of the detection network, resulting in the cracks being completely invisible (i.e., missed detection) when directly detecting the denoised, clear image. However, when Gaussian noise of a specific particle size or inherent shot noise in the sensor exists in the image, these random noise points happen to fall at the depressions of the crack, enhancing the local contrast between the crack edge and the surrounding normal metal surface through random scattering effects, or the high-frequency components of the noise "activate" the edge-sensitive convolutional kernels in the detection network. In this case, if the system runs on the noisy original image, it can detect the crack with high confidence. However, once denoising is performed, the crack features are immediately obliterated. This situation is judged as "identification only under noise components." If identification is only under noise components, the recognition assistance is at its maximum value.

[0103] If noise components are involved, it is determined whether the noise components enhance the features of the defect information. If they enhance the features of the defect information, the average recognition accuracy under the noise components is collected and recorded as the noise accuracy, and the average recognition accuracy under the non-noise components is used as the non-noise accuracy.

[0104] Among them, the enhancement feature is that the texture, frequency or contrast characteristics of noise unexpectedly resonate with the defect features (such as the high-frequency components of crack edges), which improves the confidence score of the detection network; the average recognition accuracy is obtained based on historical labeled datasets or validation sets, by inputting a subset of images containing noise components (or retaining noise) into the defect detection network, and statistically analyzing the proportion of correctly detected defects to the total number of defects.

[0105] By comparing the detection results of the sub-image containing only noise regions and the denoised sub-image on the validation set, the mean precision (mAP) or recall of detecting this type of defect by the sub-image containing only noise regions is calculated and defined as noise precision. If the noise component is not involved, it means that the noise component has no effect on defect identification, and therefore the identification assistance is at its minimum.

[0106] In some embodiments, if the noise accuracy is not greater than the non-noise accuracy, it indicates that the noise component does not play a significant role in the identification, and the identification assistance is determined to be the minimum value (e.g., 0).

[0107] If the noise accuracy is greater than the non-noise accuracy, the participation rate of the noise component in the historical identification is extracted, and the difference between the non-noise accuracy and the noise accuracy is weighted and summed to obtain the noise assistance degree of the defect information.

[0108] The identification participation rate is the proportion of the frequency in which noisy areas in historical data are marked as critical contribution areas of defects.

[0109] In some embodiments, when the noise component does not have features that enhance the defect information, the information aid degree of the noise component is determined based on the component-carrying information of the noise component; based on the noise aid degree and the information aid degree, a segmented weighting strategy is used to generate the final recognition aid degree.

[0110] In the above embodiments, noise enhancement effects are identified through historical data analysis, and high noise assistance values ​​are calculated. This guides the system to retain some specific noise in that area or employ special feature fusion strategies, rather than simply filtering it out. Conversely, for ordinary raindrop noise, if analysis reveals that it does not enhance rust detection but carries location context information, the system assigns a lower information assistance value, allowing for conventional denoising. This dynamic evaluation mechanism based on measured accuracy comparison ensures that, under extreme weather or complex backgrounds, it can maximize interference removal while also keenly capturing hidden defects that only become apparent under specific noise conditions, significantly improving the robustness and detection accuracy of the intelligent inspection system for minor faults.

[0111] In an exemplary embodiment, determining the information assistance degree of a noise component based on the component-carrying information includes: marking defect information in the component-carrying information as application defects; determining the defect quantity ratio of application defects, the average application rate of component-carrying information in defect identification, and the average time difference in the identification of defect information before and after the application of component-carrying information; and calculating the information assistance degree of the noise component based on the defect quantity ratio, the average application rate, and the average time difference.

[0112] In practice, defect information in component-carrying information is marked as application defects; the proportion of application defects, the average application rate of component-carrying information in defect identification, and the average time difference in the identification of defect information before and after the application of component-carrying information are statistically analyzed; based on the proportion of defect numbers, the average application rate, and the average time difference, a weighted sum is performed to calculate the information assistance degree of noise components.

[0113] In this process, features are extracted using a pre-trained physical attribute decoder, and a historical defect database is retrieved to identify defect types that explicitly rely on the aforementioned clues for accurate location or classification during the identification process. These are defined as "application defects." For example, "missing distant insulators" in the scenario of "using fog density to determine whether distant insulator strings are missing" is marked as an application defect.

[0114] Among them, the defect quantity ratio refers to the ratio of the number of images marked as application defects in the current batch to the total number of defects, and the average application rate refers to the ratio of the number of times the application defect carries information in the identification component to the total number of identifications.

[0115] In some embodiments, two sets of experimental data are recorded: one set is the recognition time while retaining the original noise components, and the other set is the recognition time after artificially masking or removing the information carried by the components (e.g., setting the fog to a uniform grayscale and erasing depth cues). The "average time difference" reflects the contribution of noise information to recognition efficiency. If the average time difference is positive, it indicates that using noise cues can significantly accelerate convergence or reduce the search space. If the average time difference is negative, it indicates that noise information causes interference and increases the computational burden.

[0116] In the above embodiments, the information assistance degree is calculated by weighted summation method, and environmental noise, which is traditionally regarded as interference, is transformed into an auxiliary signal that enhances diagnostic capabilities, which significantly improves the efficiency and accuracy of inspection under complex weather conditions.

[0117] In an exemplary embodiment, collecting occlusion information of noise components and determining the artifact introduction degree of the original power inspection image after denoising based on the occlusion information includes: determining the occlusion area in the occlusion information and calculating the distribution uniformity of noise components in the occlusion area; determining the occlusion content corresponding to the occlusion area and calculating the content restoration difficulty of the occlusion content; obtaining the historical artifact introduction probability of noise components, and calculating the artifact introduction degree of the original power inspection image after denoising based on the historical artifact introduction probability, distribution uniformity, and content restoration difficulty.

[0118] In practice, the occlusion information is used to determine the occlusion area in the occlusion information and to calculate the distribution uniformity of noise components in the occlusion area; the occlusion content corresponding to the occlusion area is determined and the content restoration difficulty of the occlusion content is calculated; the historical artifact introduction probability of the noise component is obtained, and based on the historical artifact introduction probability, distribution uniformity and content restoration difficulty, the artifact introduction degree of the original power inspection image after denoising is calculated.

[0119] In some embodiments, the occlusion information includes information such as the area, depth, and distance of the noise component occlusion. The gray-level variance, spectral entropy, or local standard deviation of the noise signal within the occlusion area are statistically analyzed to quantify the distribution uniformity.

[0120] If the noise is highly uniformly diffused within the occluded area (such as uniform fog), the distribution uniformity is high; if the noise exhibits drastic non-uniform abrupt changes (such as mottled raindrops or localized snow patches), the distribution uniformity is low; the more uneven the distribution, the greater the uncertainty the denoising algorithm faces in restoring the underlying texture, and the more likely it is to produce artifacts such as block effects or ringing effects.

[0121] In some embodiments, determining the occluded content corresponding to the occluded area and calculating the content restoration difficulty of the occluded content includes: calculating the local signal-to-noise ratio of each pixel in the occluded area, where the local signal-to-noise ratio is the ratio of the signal strength variance to the noise strength variance in the neighborhood of the current pixel.

[0122] Local signal-to-noise ratio (SNR) is an indicator that quantifies the effective image texture (signal) relative to the intensity of interference noise within a small space.

[0123] Specifically, a K×K sliding window is constructed centered on the current pixel (x, y). Within this window, the variance of pixel grayscale values ​​is calculated as the signal intensity variance, representing the texture richness of the region. Simultaneously, the variance of high-frequency components within this window is extracted using a high-frequency filter or a pre-trained noise estimation model as the noise intensity variance. A high local signal-to-noise ratio indicates clear texture and dominant signal at that location. A low local signal-to-noise ratio indicates that noise has obscured underlying details.

[0124] The local signal-to-noise ratio is compared with a preset noise saliency threshold. Pixels with a local signal-to-noise ratio lower than the noise saliency threshold are marked as noise foreground pixels, generating a binary mask that represents the spatial distribution of noise.

[0125] The noise saliency threshold is a dynamic or static cutoff value used to define identifiable regions from severely occluded regions. All pixels in the image that meet this threshold are assigned a value of 1 (white) and marked as noisy foreground pixels; the remaining pixels are assigned a value of 0 (black).

[0126] The connected component labeling algorithm is used to count the number of connected components and the area of ​​the largest connected component in the binary mask, and to detect whether there are noisy connected paths that run from one side boundary of the image to the opposite side boundary to determine the penetration crossing marker.

[0127] The connected component labeling algorithm is used to identify interconnected white regions (i.e., connected noise blocks) in the mask. It counts the total number of independent connected components and the total number of pixels occupied by the largest connected component. The penetration crossing flag is a Boolean variable (0 or 1) determined by traversing the pixel coordinates of the largest connected component and checking whether it simultaneously touches the left / right or top / bottom boundaries of the image. If such a penetration path exists, it indicates that the noise forms a continuous physical barrier (like a thick rain curtain or ice wall), completely blocking the lateral or vertical continuity of visual information.

[0128] The topological blocking degree is calculated based on the number of connected components, the area of ​​the largest connected component, and the penetration crossing indicator.

[0129] The topological blocking degree is calculated using a weighted summation logic. The topological blocking degree is highest when noise forms a huge single connected component and creates a penetrating barrier, which means that the spatial structure of the underlying content is severely fragmented, making restoration extremely difficult.

[0130] Defect features are extracted from defect information, noise features of noise components are collected, feature aliasing degree is determined, and content restoration difficulty is obtained by combining topological blocking degree.

[0131] In some embodiments, the steps of extracting defect features from defect information, collecting noise features of noise components, and determining the feature aliasing degree include: constructing a frequency domain template of the target to be detected and calculating the power spectral density of the current noise components.

[0132] The target to be detected refers to a specific type of defect (such as broken strands in conductors, cracks in insulators, or corrosion of hardware). The frequency domain template is a standardized spectral distribution map generated by performing Fast Fourier Transform (FFT) on a large number of standard sample images of this type of defect, extracting their amplitude spectrum, and then averaging or performing principal component analysis. It characterizes the energy distribution law of this type of defect in the frequency domain (for example, cracks usually manifest as high-frequency components in a specific direction).

[0133] Meanwhile, the system performs the same two-dimensional Fourier transform on the "noise component" (i.e. the occluded region extracted in the previous step) in the currently acquired noisy image to calculate its "power spectral density". This step transforms the texture features in the spatial domain into the energy distribution in the frequency domain so as to quantify the similarity between the two from the perspective of the spectrum.

[0134] The power spectral density and the frequency domain template are weighted and overlapped to obtain the frequency domain overlap energy value.

[0135] To more accurately match key features, a weighted matrix is ​​introduced. This matrix is ​​set according to the sensitivity of defect identification, assigning higher weights to frequency bands with significant defect features (such as specific high-frequency bands corresponding to cracks) and lower weights to general low frequencies of the background. Subsequently, the weighted power spectral density is multiplied pixel by pixel with the frequency domain template and integrated (summed) over the entire frequency domain.

[0136] The frequency domain overlap energy value and the total energy of the frequency domain template are normalized to obtain the characteristic aliasing degree.

[0137] In the above embodiments, by introducing artifact introduction, the algorithm effectively avoids misjudging the false textures generated by denoising as cracks or looseness, which significantly improves the reliability and safety of defect detection in extreme environments.

[0138] In an exemplary embodiment, determining the associated images of the original power inspection image and calculating the linkage observation degree between the denoised original power inspection image and the associated images includes: calculating the feature matching degree between the denoised original power inspection image and the denoised associated images; calculating a first feature preservation degree between the denoised original power inspection image and the original power inspection image, and a second feature preservation degree between the associated images and the denoised associated images; wherein the feature preservation degree is used to quantify the degree of information retention between the denoised images and the undenoised images; determining the comparison demand rate of the component-carried information when the noise components obtain component-carried information through comparison; and determining the linkage observation degree based on the feature matching degree, the first feature preservation degree, the second feature preservation degree, and the comparison demand rate.

[0139] In practice, associated images that have a spatiotemporal correspondence with the original power inspection images are obtained. Denoising processing is performed on the original power inspection images (input objects) and associated images respectively to obtain denoised input objects and denoised associated objects, while retaining the original input images and original associated images.

[0140] In this context, associated images refer to image data that are temporally continuous (e.g., consecutive frames in a video sequence) or spatially complementary (e.g., another perspective in binocular stereo vision, or infrared images in different bands) to the current input image to be processed. The spatiotemporal correspondence is ensured at the pixel level by establishing a homography matrix using optical flow or feature point matching (e.g., SIFT / ORB). The aforementioned denoising algorithms are then used to process these two images separately, generating denoised input objects and denoised associated objects. Simultaneously, the unprocessed original input image and the original associated image are cached in memory as ground truth values ​​for subsequent evaluation of information loss.

[0141] Calculate the feature matching degree between the denoised input object and the denoised associated object.

[0142] Calculate the feature preservation degree of the denoised input object and the denoised associated object relative to their respective original images.

[0143] Feature preservation is used to evaluate the degree to which the processed image retains information relative to the unprocessed image.

[0144] In some embodiments, feature preservation measures how much of the effective features (especially key information such as high-frequency texture details and edge structures) in the original image are preserved after denoising, or whether the denoising process leads to oversmoothing and information loss.

[0145] Determine whether the noise component carries information through comparison. If it does, collect the comparison rate of the component carrying information.

[0146] The determination of whether the information is obtained through comparison refers to the system analyzing the differences between the input image and the associated images to see if the useful information can be extracted from the noise using stereo parallax or multi-frame difference. The comparison demand rate refers to the ratio of the number of image comparisons required to obtain the component-carrying information to the total number of associated images.

[0147] The correlation observation degree is obtained by weighted summation of the comparison demand rate, feature matching degree, and feature preservation degree.

[0148] In some embodiments, if the component-carrying information cannot be obtained through comparison, the "comparison demand rate" will not play a role in calculating the "linkage observation degree" (it can be regarded as 0 or not included).

[0149] In some embodiments, the step of calculating the feature matching degree between the denoised input object and the denoised associated object includes: mapping the denoised input object to the imaging view of the associated image with a spatiotemporal correspondence, and constructing a projection overlap region.

[0150] The denoising input object refers to the current frame image (or current viewpoint image) after preliminary denoising processing, and the associated image refers to the denoised image that is temporally adjacent (such as the previous frame of a video sequence) or spatially complementary (such as the right eye image of a stereo camera).

[0151] The mapping process relies on a pre-computed homography matrix or essential matrix, obtained through feature point matching and RANSAC robust estimation, which describes the geometric transformation relationship between two images. The pixel coordinates of the denoised input object are transformed to the coordinate system of the associated image using this matrix, yielding the transformed coordinates. Subsequently, the actual overlapping pixel range between the two images after transformation is identified, and non-overlapping boundaries caused by viewpoint differences or object motion are removed. This generated common effective region is the "projected overlap region." Within the projected overlap region, the edge contour features of the denoised input object and the associated denoised object are extracted, and the contour overlap rate is calculated.

[0152] Edge contour features refer to the skeletal information of the object's structure, and binarized edge maps are typically extracted using the Canny or Laplacian operators. Edge maps from two denoised images are extracted within the overlapping projection regions. To quantify the degree of similarity between the two, the ratio of the intersection to the union of the edges after morphological dilation is calculated, i.e., the Jaccard similarity coefficient.

[0153] Extract the local texture descriptors of the denoised input object and the local texture descriptors of the denoised associated object, and calculate the texture cross-correlation coefficient.

[0154] The local texture descriptor is a mathematical representation of the microscopic gray-level distribution of an image, and can be extracted using LBP (Local Binary Pattern) or SIFT feature extraction algorithms. Several local windows are divided within the projected overlapping region, and a descriptor vector is calculated for each window. Then, the linear correlation between all corresponding windows is calculated using the Pearson correlation coefficient formula, and the average value is taken to obtain the texture cross-correlation coefficient. The feature matching degree is generated based on a weighted average of the contour overlap rate and the texture cross-correlation coefficient.

[0155] In some embodiments, the step of calculating the feature preservation degree of the denoised input object and the denoised associated object relative to their respective original images includes: constructing a first difference map between the original input image and the denoised input object, and a second difference map between the original associated image and the denoised associated object.

[0156] The first difference image and the second difference image are residual images obtained by pixel-level subtraction.

[0157] The proportion of high-frequency texture energy in the first difference map is extracted as the noise retention rate at the input end, and the proportion of high-frequency texture energy in the second difference map is extracted as the noise retention rate at the associated end.

[0158] A frequency domain transformation (such as Fourier or wavelet transform) is performed on the difference map to decompose the image into different frequency components. The high-frequency texture energy ratio is defined as the ratio of the energy of the components above a certain cutoff frequency in the frequency domain to the total energy. If this ratio exceeds a preset energy threshold, it indicates that the denoising process has preserved more high-frequency details. If the ratio is lower than the preset energy threshold, it suggests over-denoising. This ratio is defined as the input noise retention rate and the correlation noise retention rate, respectively, to quantify the aggressiveness of the denoising operations at both ends in the frequency domain.

[0159] Calculate the relative deviation between the noise retention rate at the input end and the noise retention rate at the associated end.

[0160] If the relative deviation value is less than the preset consistency threshold, the reciprocal of the relative deviation value is normalized and used as the feature preservation degree.

[0161] A consistency threshold is set. When the relative deviation is less than the threshold, the denoising effect at both ends is considered to be highly consistent. At this time, the feature preservation degree is directly obtained by mapping the inverse of the deviation, ensuring that the score approaches 1 as the deviation decreases.

[0162] If the relative deviation value is greater than or equal to the consistency threshold, the rate of change of structural similarity of the main body region of the object before and after denoising is calculated. The rate of change of structural similarity is multiplied by the penalty factor of the relative deviation value to obtain the corrected feature preservation degree.

[0163] In the above embodiments, the linkage observation degree is calculated using the entropy weight method. Higher feature matching degree and preservation degree indicate better denoising quality and stronger consistency. Conversely, a higher comparison demand rate indicates greater difficulty in obtaining the information carried by the denoised components themselves after denoising, warning the system against aggressive denoising to avoid losing crucial scene information carried in the noise. Feature matching degree acts as a checkpoint for spatiotemporal consistency, ensuring logical coherence in structure, texture, and semantics between the denoised input object and associated objects, preventing temporal jitter or stereoscopic parallax distortion caused by excessive denoising in a single frame, thus ensuring a smooth transition of the observation perspective. Feature preservation degree acts as a gatekeeper for information integrity, quantifying the degree to which the denoising process retains subtle details of the original defects, avoiding the erasure of key fault features due to blind smoothing, and ensuring the authenticity and reliability of the observed content.

[0164] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0165] Based on the same inventive concept, this application also provides an image enhancement device for power generator inspection based on semantic consistency and reinforcement learning, used to implement the above-mentioned image enhancement method for power generator inspection based on semantic consistency and reinforcement learning. The solution provided by this device is similar to the implementation described in the above method. Therefore, the specific limitations of one or more embodiments of the image enhancement device for power generator inspection based on semantic consistency and reinforcement learning provided below can be found in the limitations of the image enhancement method for power generator inspection based on semantic consistency and reinforcement learning described above, and will not be repeated here.

[0166] In one exemplary embodiment, such as Figure 3 As shown, a device for enhancing images of electric motor inspections based on semantic consistency and reinforcement learning is provided, comprising: an extraction module 301, a determination module 302, and an enhancement module 303, wherein:

[0167] The extraction module 301 is used to acquire the original power inspection image, input the original power inspection image into the semantic segmentation network, and extract the noise components in the original power inspection image.

[0168] The determination module 302 is used to collect the component-carrying information of the noise components and determine the information missing degree of the original power inspection image after denoising based on the component-carrying information.

[0169] The determining module 302 is further configured to collect occlusion information of noise components and determine the artifact introduction degree of the original power inspection image after denoising based on the occlusion information; the artifact introduction degree is used to quantify the risk of introducing new artifacts after denoising the original power inspection image.

[0170] The determining module 302 is further configured to determine the associated images of the original power inspection image and calculate the linkage observation degree between the denoised original power inspection image and the associated images; the linkage observation degree is used to quantitatively characterize the consistency and information complementarity availability between the denoised original power inspection image and the associated images.

[0171] The enhancement module 303 is used to determine the denoising degree of the original power inspection image after denoising based on the information missing degree, the artifact introduction degree, and the linkage observation degree, and to perform enhancement processing on the original power inspection image according to the denoising degree.

[0172] In some exemplary embodiments, the determining module 302 described above is further configured to:

[0173] Extract the non-noise components from the original power line inspection images before and after denoising;

[0174] The non-noise parts of the original power inspection images before and after denoising are compared to determine the non-noise similarity of the non-noise components before and after denoising.

[0175] The topological structures of the original power inspection images before and after denoising are compared to determine the structural similarity of the topological structures.

[0176] Based on defect information, determine the degree to which noise components aid in the identification of defect information;

[0177] Based on non-noise similarity, structural similarity, and recognition assistance, the information loss degree of the original power inspection image after denoising is determined.

[0178] In some exemplary embodiments, the determining module 302 described above is further configured to:

[0179] In cases where defect information is not only identified under noise components, determine whether noise components are involved in the identification of historical identification data of defect information.

[0180] When noise components are involved in identification, it is determined whether the noise components enhance the features of defect information;

[0181] In the case of enhancing defect information with noise components, the average recognition accuracy under noise components is taken as the noise accuracy, and the average recognition accuracy under non-noise components is taken as the non-noise accuracy.

[0182] When the noise accuracy is greater than the non-noise accuracy, the noise assistance degree of the defect information is determined based on the difference between the identification participation rate of noise components, the non-noise accuracy, and the noise accuracy in historical identification data.

[0183] In the absence of features that enhance defect information in the noise component, the information aiding degree of the noise component is determined based on the component-carrying information of the noise component.

[0184] Based on noise assistance and information assistance, the degree of assistance of noise components in identifying defect information is determined.

[0185] In some exemplary embodiments, the determining module 302 described above is further configured to:

[0186] Mark the defect information in the component-carrying information as an application defect;

[0187] Determine the percentage of defects in the application, the average application rate of component-carrying information in defect identification, and the average time difference in defect information identification before and after the application of component-carrying information;

[0188] The information aids of noise components are calculated based on the proportion of defective items, average application rate, and average duration difference.

[0189] In some exemplary embodiments, the determining module 302 described above is further configured to:

[0190] Identify the occlusion region in the occlusion information and calculate the uniformity of noise component distribution within the occlusion region;

[0191] Determine the content that is obscured corresponding to the obscured area, and calculate the difficulty of restoring the content of the obscured area;

[0192] The historical artifact introduction probability of the noise component is obtained, and the artifact introduction degree of the original power inspection image after denoising is calculated based on the historical artifact introduction probability, distribution uniformity and content restoration difficulty.

[0193] In some exemplary embodiments, the above-described apparatus further includes a computing module for:

[0194] Calculate the feature matching degree between the original power inspection image after denoising and the associated image after denoising;

[0195] Calculate the first feature preservation degree between the original power inspection image after denoising and the second feature preservation degree between the associated image and the denoised associated image; wherein, the feature preservation degree is used to quantify the degree of information preservation between the denoised image and the image before denoising;

[0196] When noise components are compared to obtain component-carrying information, the comparison requirement rate of component-carrying information is determined.

[0197] The linkage observation degree is determined based on feature matching degree, first feature preservation degree, second feature preservation degree, and comparison demand rate.

[0198] The modules in the aforementioned image enhancement device for power generator inspection based on semantic consistency and reinforcement learning can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0199] In one exemplary embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 4 As shown, the computer device includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, Near Field Communication (NFC), or other technologies. When the computer program is executed by the processor, it implements a power motor inspection enhancement method based on semantic consistency and reinforcement learning.

[0200] The display unit of this computer device is used to form a visually visible image and can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of this computer device can be a touch layer covering the display screen, or buttons, a trackball, or a touchpad set on the casing of the computer device, or an external keyboard, touchpad, or mouse, etc.

[0201] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0202] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0203] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0204] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0205] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0206] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0207] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0208] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for enhancing images of electric motor inspections based on semantic consistency and reinforcement learning, characterized in that, The method includes: The original power line inspection images are obtained and input into a semantic segmentation network to extract noise components from the original power line inspection images. Collect the component-carrying information of the noise components, and determine the information loss degree of the original power inspection image after denoising based on the component-carrying information; Obstruction information of noise components is collected, and the artifact introduction degree of the original power inspection image after denoising is determined based on the obstruction information; the artifact introduction degree is used to quantify the risk of introducing new artifacts after denoising the original power inspection image. The associated images of the original power inspection images are identified, and the linkage observation degree between the denoised original power inspection images and the associated images is calculated; the linkage observation degree is used to quantitatively characterize the consistency and information complementarity between the denoised original power inspection images and the associated images. Based on the information missing degree, the artifact introduction degree, and the linkage observation degree, the denoising degree of the original power inspection image after denoising is determined, and the original power inspection image is enhanced according to the denoising degree.

2. The method according to claim 1, characterized in that, The component-carrying information includes defect information in the original power line inspection image; the process of collecting the component-carrying information of the noise components and determining the information loss degree of the original power line inspection image after denoising based on the component-carrying information includes: Extract the non-noise components from the original power line inspection images before and after denoising; The non-noise components of the original power inspection images before and after denoising are compared to determine the non-noise similarity of the non-noise components before and after denoising. The topological structures of the original power inspection images before and after denoising are compared to determine the structural similarity of the topological structures. Based on the defect information, determine the degree of assistance the noise component provides in identifying the defect information; Based on the non-noise similarity, the structural similarity, and the recognition assistance, the information loss degree of the original power inspection image after denoising is determined.

3. The method according to claim 2, characterized in that, The step of determining the degree of assistance of the noise component in identifying the defect information based on the defect information includes: In cases where defect information is not identified only under noise components, it is determined whether the noise components participated in the identification in the historical identification data of the defect information; When noise components are involved in identification, it is determined whether the noise components enhance the features of defect information; In the case of enhancing defect information with noise components, the average recognition accuracy under noise components is taken as the noise accuracy, and the average recognition accuracy under non-noise components is taken as the non-noise accuracy. When the noise accuracy is greater than the non-noise accuracy, the difference between the non-noise accuracy and the noise accuracy is calculated. Based on the identification participation rate of noise components in historical identification data and the difference, the noise assistance degree of the defect information is determined. In the absence of features that enhance defect information in the noise component, the information aiding degree of the noise component is determined based on the component-carrying information of the noise component. Based on the noise assistance degree and the information assistance degree, the identification assistance degree of the noise component for the defect information is determined.

4. The method according to claim 3, characterized in that, The determination of the information assistance degree of noise components based on the component-carrying information of noise components includes: The defect information in the component-carrying information is marked as an application defect; Determine the percentage of defects in the application defects, the average application rate of component-carrying information in defect identification, and the average time difference in defect information identification before and after the application of component-carrying information; The information assistance degree of the noise component is calculated based on the proportion of the number of defects, the average application rate, and the average duration difference.

5. The method according to claim 1, characterized in that, The process of acquiring occlusion information of noise components and determining the artifact introduction degree of the original power inspection image after denoising based on the occlusion information includes: Determine the occlusion region in the occlusion information and calculate the distribution uniformity of noise components in the occlusion region; Determine the content that is obscured corresponding to the obscured area, and calculate the difficulty of restoring the content of the obscured area; The historical artifact introduction probability of the noise component is obtained, and the artifact introduction degree of the original power inspection image after denoising is calculated based on the historical artifact introduction probability, the distribution uniformity, and the content restoration difficulty.

6. The method according to claim 1, characterized in that, The process of determining the associated images of the original power line inspection images and calculating the correlation observation degree between the denoised original power line inspection images and the associated images includes: Calculate the feature matching degree between the original power inspection image after denoising and the associated image after denoising; Calculate the first feature preservation degree between the original power inspection image after denoising and the second feature preservation degree between the associated image and the denoised associated image; wherein, the feature preservation degree is used to quantify the degree of information preservation between the denoised image and the image before denoising; In the case of obtaining component-carrying information by comparing the noise components of the original power inspection image and the associated image, the comparison requirement rate of the component-carrying information is determined. Based on the feature matching degree, the first feature preservation degree, the second feature preservation degree, and the comparison demand rate, the linkage observation degree is determined.

7. A power motor inspection image enhancement device based on semantic consistency and reinforcement learning, characterized in that, The device includes: The extraction module is used to acquire the original power inspection image, input the original power inspection image into the semantic segmentation network, and extract the noise components in the original power inspection image; The determination module is used to collect the component-carrying information of the noise components and determine the information missing degree of the original power inspection image after denoising based on the component-carrying information. The determining module is also used to collect occlusion information of noise components and determine the artifact introduction degree of the original power inspection image after denoising based on the occlusion information; the artifact introduction degree is used to quantify the risk of introducing new artifacts after denoising the original power inspection image. The determining module is further configured to determine the associated images of the original power inspection image and calculate the linkage observation degree between the denoised original power inspection image and the associated images; the linkage observation degree is used to quantitatively characterize the consistency and information complementarity availability between the denoised original power inspection image and the associated images. An enhancement module is used to determine the denoising degree of the original power inspection image after denoising based on the information missing degree, the artifact introduction degree, and the linkage observation degree, and to perform enhancement processing on the original power inspection image according to the denoising degree.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images