Visual recognition of icing risk on overhead transmission lines based on improved fuzzy wavelet domain

• The proposed method maintains an F1 value above 0.95 and recognition accuracy above 97%, significantly outperforming compare methods in identifying icing risks on overhead transmission lines. • The method consistently achieves recognition times below 1 second per sample, ensuring real-time performance and responsiveness, in contrast to comparison methods that exceed 2 seconds. • By dynamically adjusting fuzzy membership function parameters based on statistical characteristics of wavelet sub-band images, the method extracts more representative features, enhancing recognition accuracy for varying icing levels. • The removal of key components—improved fuzzy wavelet domain feature extraction and adaptive fuzzification—results in significant performance drops, confirming the effectiveness and necessity of these designs in the proposed method. When extracting icing features of overhead transmission lines in the fuzzy wavelet domain, reliance on static fuzzy membership function parameters leads to a lower F1 value in identifying icing risks. To improve the visual recognition effect of icing risk on overhead transmission lines, a visual recognition method for icing risk on overhead transmission lines based on improved fuzzy wavelet domain is proposed. Visual image acquisition of overhead transmission lines is completed by a spherical camera, and to ensure image quality, image filtering and morphological unfolding are sequentially used for image preprocessing; By using an adaptive fuzzy improved fuzzy wavelet domain method, the ice covered image of overhead transmission lines is decomposed into two-dimensional discrete wavelet transform, and the fuzzy membership function parameters are dynamically adjusted based on the statistical characteristics of each sub-band image. Statistical and texture features are extracted and fused, and the feature vector is finally normalized, providing strong support for ice cover risk identification; Input the extracted icing features into a convolutional neural network, perform convolution and pooling processing, and integrate the results through a fully connected layer; The Softmax function ultimately outputs the probability of risk level, selects the category with the highest probability as the final recognition result, and presents it in a graphical interface created using Python to complete the visual recognition of icing risk on overhead transmission lines. The test results show that when using the proposed method for recognition, the F1 value can be maintained above 0.95, the recognition accuracy is above 97%, and the recognition time is always below 1 second. The misclassification rate is low, and the average absolute error (MAE) is reduced by 12% -30% compared to other models. Cohen's Kappa consistency coefficient remains stable above 0.78, and the average inference time is only 0.5-0.7 seconds, which has a visual recognition effect on the risk of icing on overhead transmission lines.