Self-Supervised Power Line Segmentation for UAV Embedded Systems
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional algorithms for power line segmentation are ineffective in complex scenes, and deep learning-based models are impractical for embedded devices in unmanned aerial vehicles due to the need for large amounts of labeled data and significant computing power.
Innovation Solution
A method utilizing self-supervised learning, including region separation, random combination, and image inpainting, generates a large amount of labeled data for training a real-time power line segmentation network (SaSnet) with a reduced model size, allowing deployment on embedded devices.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional line-based algorithms are used for power line segmentation, then the algorithm can be applied in simple scenes, but false detection and missed detection occur in complex scenes
Solution Approach 1:
The patent replaces traditional mechanical line-based algorithms with a deep learning-based segmentation model that uses automated feature extraction and classification. The model substitutes manual thresholding and geometric operations with neural network-based image processing, enabling accurate segmentation in complex scenes through learned patterns rather than rigid algorithmic rules.
2Measurement precision
If a deep learning-based segmentation model is used, then segmentation accuracy improves, but the model cannot be deployed on embedded devices due to large model size and computing power requirements
Solution Approach 1:
The patent segments the deep learning model into a lightweight architecture suitable for embedded deployment. This involves dividing the model into essential components that maintain accuracy while reducing computational burden, allowing the segmentation functionality to be distributed across different processing stages or hardware modules.
Solution Approach 2:
The patent changes key model parameters including reducing the number of layers, decreasing filter counts, using lighter activation functions, and optimizing weight initialization. These parameter adjustments significantly reduce model size and computing requirements while preserving segmentation accuracy for power line detection.
3Measurement precision
If a deep learning model is trained with large amounts of labeled data, then model performance improves, but training time and data preparation complexity increase
Solution Approach 1:
The patent performs preliminary actions by pre-processing and augmenting a small set of labeled power line images before training. This includes geometric transformations, color adjustments, and synthetic label generation that prepare the data in advance, reducing the need for extensive labeled datasets and accelerating the training process.
Solution Approach 2:
The patent implements self-service through self-supervised learning mechanisms where the model generates its own training labels from unlabeled data. The system automatically creates pseudo-labels and augments training samples without manual annotation, enabling the model to improve performance while minimizing external data preparation efforts.
Data Source
AI summary
A method for segmenting a power line image in real time based on self-supervised learning includes: inputting an input power line sample image and power line sample image mask set for the same batch of images into a region growing algorithm to obtain a single power line sub-image and single power line mask set; randomly extracting at least one single power line image pair for combination, and combining the single power line image pair with a random background picture to generate a power line random background fusion image and power line random background mask set; and carrying out random non-repetitive region growing to obtain image inpainting regions, forming a segmentation mask with the image inpainting regions, obtaining power line segmentation images through an image inpainting algorithm, inputting the power line segmentation images into a power line real-time segmentation network for training, and carrying out predicted segmentation.