Transmission Line Defect Identification Using Saliency Map and Semantic-Embedded Feature Pyramid
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for transmission line defect identification face challenges due to the long tail effect of defective and non-defective samples, leading to model overfitting and poor performance in feature extraction.
Innovation Solution
A transmission line defect identification method based on a saliency map and a semantic-embedded feature pyramid, which uses a super-resolution generative adversarial network (SRGAN) for data augmentation, a saliency map for defect feature retention, and a deep semantic embedding (DSE)-based feature pyramid classification network for accurate defect classification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If data augmentation and resampling are performed to generate new samples for categories with insufficient samples, then the long tail effect is resolved, but model overfitting occurs and feature extraction model performance is lowered
Solution Approach 1:
The patent applies local quality by differentiating the treatment of different image regions through saliency map guidance. Instead of uniform data augmentation, the method identifies salient regions (defect areas) and applies augmentation strategies specifically to those regions, while preserving non-salient regions. This localized approach ensures that augmentation enhances defect visibility without introducing artifacts that would cause overfitting.
Solution Approach 2:
The patent implements preliminary action by pre-computing saliency maps and segmenting images into salient and non-salient regions before performing data augmentation. This preliminary segmentation allows the subsequent augmentation process to be guided by pre-identified defect regions, ensuring that synthetic data generation focuses on creating relevant defect variations rather than random modifications that could lead to overfitting.
2Quantity of substance
If random cutout algorithm is used to augment defect images, then data augmentation is performed, but salient features are cut out and defect feature learning is hindered
Solution Approach 1:
The patent applies local quality by using saliency maps to identify and protect defect regions from cutting operations. The augmentation process selectively applies cutout operations only to non-salient regions while preserving salient defect features. This ensures that data augmentation increases dataset quantity without removing critical defect information needed for model learning.
Solution Approach 2:
The patent converts the potentially harmful effect of random cutout operations into a beneficial process by using saliency maps to guide the cutout operation. Instead of randomly cutting any region, the method uses saliency information to deliberately preserve defect regions and only apply cutout to safe zones, transforming a harmful random process into a controlled beneficial augmentation strategy.
3Area of stationary object
If small target images are used for transmission line inspection, then inspection coverage is improved, but boundary clarity and semantic information are insufficient
Solution Approach 1:
The patent applies dimensionality change by transitioning from 2D image processing to 3D spatial reasoning through saliency map generation. The saliency map creates an additional dimensional representation that highlights defect regions, enabling the model to infer boundary information and semantic context from the enhanced spatial relationships in the augmented dimension.
Solution Approach 2:
The patent implements preliminary action by pre-computing saliency maps that enhance the semantic information of small target images before classification. This preliminary enhancement adds boundary clarity and contextual information to the 2D images, allowing the model to accurately identify and classify defects in small targets without requiring higher resolution input images.
Data Source
AI summary
The present disclosure provides a transmission line defect identification method based on a saliency map and a semantic-embedded feature pyramid, including the following steps: step 1: cleaning and classifying a dataset; step 2: generating a super-resolution image for a small target of a transmission line by using an Electric Line-Enhanced Super-Resolution Generative Adversarial Network (EL-ESRGAN) model; step 3: performing image saliency detection on the dataset by constructing a U2-Net; step 4: performing data augmentation on the dataset by using GridMask and random cutout algorithms based on a saliency map, and generating a classified dataset; and step 5: performing image classification on a normal set and a defect set by using a ResNet34 classification algorithm and a deep semantic embedding (DSE)-based feature pyramid classification network.