CN-121981957-A - Power transmission line visual monitoring image quality assessment method based on AI large model

CN121981957ACN 121981957 ACN121981957 ACN 121981957ACN-121981957-A

Abstract

The invention relates to a transmission line visual monitoring image quality evaluation method based on an AI large model, which belongs to the technical field of image processing, and comprises the steps that an edge device uploads an acquired monitoring image to a cloud end, the cloud end AI large model classifies environmental interference, and a key region mask and a dynamic threshold value are generated; the method comprises the steps of receiving a cloud end monitoring image, acquiring a key area mask and a dynamic threshold value, caching the key area mask and the dynamic threshold value issued by the cloud end at an edge equipment end, extracting global features from the real-time collected monitoring image by adopting a local model to obtain an environmental score, positioning a key area by utilizing the cached key area mask, performing channel-space two-stage weighted strengthening on the key area features to obtain a key area score, fusing the environmental score and the key area score to obtain an evaluation score of the real-time image, and judging whether the quality of the monitoring image is qualified or not by comparing the evaluation score with the cached dynamic threshold value. The method is suitable for the scene of the power transmission line and gives consideration to real-time performance and accuracy.

Inventors

JIANG ANFENG
ZHAO YINGYING
TIAN YUE
SI WENRONG
LIU ZHAOJIE
WAN XIN
ZHANG MING

Assignees

国网上海市电力公司
华东电力试验研究院有限公司
西安艾飞能源数字技术有限公司

Dates

Publication Date: 20260505
Application Date: 20251226

Claims (10)

1. The transmission line visual monitoring image quality evaluation method based on the AI large model is characterized by comprising the following steps of: the edge equipment end uploads the acquired monitoring image to the cloud end, and the cloud end AI large model classifies the environmental interference and generates a key region mask and a dynamic threshold value, wherein the dynamic threshold value is obtained by the joint calculation of the environment type probability and the low-frequency energy duty ratio of the image and is used for evaluating the quality of the monitoring image; the edge equipment side caches the key area mask and the dynamic threshold value issued by the cloud; the method comprises the steps of (1) extracting global features from a monitored image acquired in real time to obtain an environmental score by a local model arranged in an edge equipment end, positioning a key region by using a cached key region mask, performing channel-space two-stage weighted reinforcement on the key region features to obtain a key region score, and fusing the environmental score and the key region score to obtain an evaluation score of the real-time image; and the edge equipment end judges whether the quality of the monitored image is qualified or not by comparing the evaluation score with the cached dynamic threshold value.
2. The AI-large-model-based transmission line visual monitoring image quality evaluation method according to claim 1, characterized in that, The AI large model is a Qwen-VL-7B visual language large model subjected to LoRA fine adjustment, fine adjustment data are transmission line labeling images, fine adjustment objects are visual coding layers, fine adjustment prompting words are defined as analyzing transmission line images, only focusing on insulators/wires and outputting environment types and confidence degrees, the input monitoring images are processed by the fine adjustment large model to obtain environment classification comprising rain and snow/haze/normal, and key area masks and dynamic thresholds are generated.
3. The AI-large-model-based transmission line visual monitoring image quality evaluation method according to claim 2, characterized in that, The AI large model processes the input monitoring image according to a preset irregular form standard and a preset visibility requirement to generate a key area mask; The irregular form standard comprises preset insulator umbrella skirt curvature, wire sag height difference and fitting connection point curvature radius, wherein the generated key region mask comprises only an umbrella skirt bending part for an insulator region mask, only a wire sag lowest point part for a wire region mask and only a wire clamp/damper connection region for a fitting region mask; In the 'visibility' requirement, the visibility of the key region is graded and scored according to the average gradient of the edge and the continuous quantification result, so as to quantify the visibility of the key region under the condition of rain and snow/fog environment interference.
4. The AI-large-model-based transmission line visual monitoring image quality evaluation method according to claim 2, characterized in that, The calculation formula of the dynamic threshold value is as follows: Threshold=base_threshold+ΔT rain ·I(r_s_prob≥θ r_s )+ΔT freq ·I(l_f_ratio≥θ freq ) The method comprises the steps of taking Threshold as a dynamic evaluation Threshold, taking base_threshold as a reference Threshold, taking the base_threshold as a basic Threshold under an environment-free interference scene, taking DeltaT rain as a rain and snow environment Threshold increment, taking r_s_prob as a rain and snow probability, taking theta r_s as a rain and snow probability judgment Threshold, taking I (·) as an indication function, returning 1 when the condition in brackets is true, returning 0 when the condition is false, taking DeltaT freq as a low-frequency interference Threshold increment, taking l_f_ratio as a low-frequency energy duty ratio, and taking theta freq as a low-frequency duty ratio judgment Threshold.
5. The AI large model based transmission line visual monitoring image quality assessment method as set forth in claim 4, wherein, The rain and snow probability calculation formula is as follows: r_s_prob=σ(W r_s ·feat Q +b r_s ) wherein sigma (·) is a Softmax activation function, W r_s 、b r_s is a classification weight and a bias term of the rain and snow environment type respectively, and feat Q is a visual feature vector output by the Qwen-VL-7B visual language large model.
6. The AI large model based transmission line visual monitoring image quality assessment method as set forth in claim 4, wherein, According to Haar wavelet transformation of the image, a multi-scale decomposition characteristic diagram and a low-frequency energy ratio l_f_ratio are obtained through wavelet frequency domain separation; Wherein Haar i,j is the frequency domain coefficient of the image point i, j after Haar transformation, H is the height of the image, W is the width of the image; For the low frequency region, the image point i, j is a "1" when in the low frequency region, otherwise is a "0".
7. The AI-large-model-based transmission line visual monitoring image quality assessment method according to any one of claims 1 to 6, characterized in that, A local model disposed in an edge device side, comprising: The preprocessing module is used for carrying out standardization and normalization processing on the monitoring image acquired in real time to obtain a normalized RGB image with a set size, carrying out Haar wavelet transformation on the normalized RGB image, and separating the normalized RGB image through a wavelet frequency domain to obtain a multi-scale decomposition characteristic image and a low-frequency energy duty ratio; The deformable convolution module is used for dynamically generating offset according to the cached key region mask and the component morphological parameters, and extracting the characteristics of the key region with irregular morphology in the normalized RGB image to obtain a key region characteristic map; the global feature extraction module is used for taking the enhancement feature map obtained by splicing the key region feature map and the multi-scale decomposition feature map as an input image, and extracting the global feature map and learning the global feature map by residual errors; The key region feature extraction module sequentially executes channel-space two-stage weighted reinforcement on the input key region feature map and the global feature map by adopting CBAM attentions to extract a key region dominant feature map after double weighting; The evaluation score is compared with a dynamic threshold value to judge whether the monitored image is qualified or not, and a real-time decision comprising the evaluation score and a judgment result is obtained; and the result output module is used for outputting the real-time decision, the key region mask map and the corresponding coordinates.
8. The AI large model based transmission line visual monitoring image quality evaluation method as set forth in claim 7, wherein, The deformable convolution formula in the deformable convolution module is: Δp i ＝k·max(0,shape_param-θ param ) Where Δp i represents the ith pixel offset of the deformable convolution kernel, k is the offset coefficient positively correlated to the component profile curvature, shape_param is the component morphology parameter, and θ param is the critical threshold for the morphology parameter.
9. The AI large model based transmission line visual monitoring image quality evaluation method as set forth in claim 7, wherein, In the key region feature extraction module, The channel attention weights for the CBAM attention mechanism are: the spatial attention weights of CBAM attention mechanisms are: spatial_weight=σ(7×7Conv(concat(avg_pool,max_pool))) The fusion characteristics obtained are: Wherein W 1 、W 2 is a channel attention weight matrix, b 1 、b 2 is a channel attention bias term, sigma (·) is a Softmax activation function, 7×7Conv is a 7×7 convolution kernel of the spatial attention convolution layer, and feat is a feature vector extracted by ResNet 18.
10. The AI large model based transmission line visual monitoring image quality evaluation method as set forth in claim 7, wherein, The scoring and decision-making module comprises a fusion scoring sub-module and a decision sub-module, wherein, The fusion evaluation sub-module is used for balancing the contribution of the key area and the environment by adopting a weighted integration strategy and outputting an evaluation score; Splicing the key region dominant feature map and the global feature map along the channel dimension, and inputting single neuron full-connection layer regression to obtain a key region score S_region corresponding to the key region dominant feature map and an environment score S_env corresponding to the global feature map; the final evaluation score obtained was s=α 1 ×S_region+α 2 ×s_env; The decision sub-module is used for comparing the evaluation score with a dynamic threshold value to obtain a judging result of whether the monitored image is qualified or not, wherein the evaluation score is qualified if the evaluation score is larger than the dynamic threshold value, and is unqualified if the evaluation score is not smaller than the dynamic threshold value; When the cloud end and the edge equipment end are well connected, comparing the latest issued dynamic threshold values of the large model; When the network connection between the cloud end and the edge equipment end is disconnected, searching for the median in a plurality of recently cached dynamic thresholds, and comparing.

Description

Power transmission line visual monitoring image quality assessment method based on AI large model Technical Field The invention relates to the technical field of image processing, in particular to a transmission line visual monitoring image quality assessment method based on an AI large model. Background With the intelligent development of the power system, the inspection of the power transmission line gradually relies on unmanned aerial vehicle, fixed cameras and other equipment to collect images, and fault detection of key components such as towers, insulators, wires and the like is realized through a follow-up defect identification algorithm. The image quality evaluation is a core front link, namely, a qualified image which has no environmental distortion (such as rain, snow and haze) and clear key areas is required to be screened from a mass of inspection images, otherwise, the false judgment rate of a defect recognition algorithm is increased (such as false judgment of haze blurring as part corrosion). In the current power inspection scene, image quality evaluation faces two major core challenges, namely complex environment distortion types (rain and snow coverage and haze with different concentrations), difficult accurate distinction of a general image quality evaluation method, and most of inspection equipment is edge equipment, and real-time performance and accuracy are required to be evaluated in a balanced manner. Specifically, the following drawbacks are included: 1. typical interferences such as rain, snow, haze and the like are not explicitly modeled, so that the accuracy rate of distortion type identification is less than 80%, and the evaluation result is severely fluctuated along with meteorological conditions. 2. Depending on global features or low-level edges, the semantic priori of insulator-wire-hardware is lacking, the omission ratio of key parts is more than 20%, and qualified erroneous judgment of clear background and fuzzy target appears. 3. The fixed judgment boundary of offline statistics is adopted, and cannot be adjusted in real time according to environment priori and image frequency domain energy, and the false erasure rate fluctuation is more than 10% when seasons alternate. 4. The method has no geometric modeling capability on irregular structures such as wire sag, insulator umbrella skirt special-shaped and the like, the positioning error of a complex scene is more than 3pix, and the omission ratio is more than 30%. 5. The threshold value is determined by relying on a ten-thousand-level labeling sample, periodic manual re-labeling is needed, online learning cannot be realized, updating capability is lost in a network breaking environment, and edge deployment elasticity is insufficient. Therefore, the development of an image quality evaluation scheme which is adaptive to the scene of the power transmission line and gives consideration to the instantaneity and the accuracy has important significance in improving the power inspection efficiency. Disclosure of Invention In view of the analysis, the invention aims to disclose a transmission line visual monitoring image quality evaluation method based on an AI large model, and solve the problems in the background technology. The invention discloses a transmission line visual monitoring image quality evaluation method based on an AI large model, which comprises the following steps: the edge equipment end uploads the acquired monitoring image to the cloud end, and the cloud end AI large model classifies the environmental interference and generates a key region mask and a dynamic threshold value, wherein the dynamic threshold value is obtained by the joint calculation of the environment type probability and the low-frequency energy duty ratio of the image and is used for evaluating the quality of the monitoring image; the edge equipment side caches the key area mask and the dynamic threshold value issued by the cloud; the method comprises the steps of (1) extracting global features from a monitored image acquired in real time to obtain an environmental score by a local model arranged in an edge equipment end, positioning a key region by using a cached key region mask, performing channel-space two-stage weighted reinforcement on the key region features to obtain a key region score, and fusing the environmental score and the key region score to obtain an evaluation score of the real-time image; and the edge equipment end judges whether the quality of the monitored image is qualified or not by comparing the evaluation score with the cached dynamic threshold value. Further, the AI large model is a Qwen-VL-7B visual language large model subjected to LoRA fine adjustment, fine adjustment data are transmission line labeling images, fine adjustment objects are visual coding layers, fine adjustment prompt words are defined as 'analyzing transmission line images, only paying attention to insulators/wires, outputting environment types and confid