UAV Image Change Detection by Separating Essential and State Features
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing change detection methods using twin convolutional neural networks for low-altitude UAV images struggle to differentiate between essential changes and state changes caused by lighting, shading, or color styles, leading to inaccurate predictions.
Innovation Solution
Decouple backbone features into essential and state parts, analyzing the importance of state features to eliminate their influence and focus on predicting real changes, using a cascade fusion approach with a trained change detection model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If twin convolutional neural networks are used to extract features from images at different times, then feature extraction capability is improved, but the model cannot distinguish essential changes from state changes caused by lighting, shading, or color styles
Solution Approach 1:
The patent segments the feature extraction process into two distinct pathways: a shared backbone network for extracting common features, and separate projection networks for capturing image-specific features. This segmentation allows the model to separately process and compare essential features while being invariant to state changes like lighting and color variations.
Solution Approach 2:
The patent extracts and removes state-related features (lighting, shading, color styles) from the feature representation before comparison. By taking out these confounding factors through the projection network that learns image-specific transformations, the model focuses only on essential changes in the target objects.
2Measurement precision
If all differences between two images are detected, then comprehensive change detection is achieved, but state changes (lighting, shading, color) interfere with predicting correct change maps
Solution Approach 1:
The patent converts the harmful effect of state changes into a beneficial learning signal. The projection network is trained to learn image-specific transformations that capture state variations, and this knowledge is then used to remove state changes from the feature representation, turning the previously harmful variability into a useful feature for improving change detection accuracy.
Solution Approach 2:
The patent applies different processing qualities to different parts of the feature space: the shared backbone network uses standardized feature extraction for consistency, while the projection networks apply localized, image-specific transformations to handle state changes. This local quality differentiation allows the model to treat essential features and state features differently.
Data Source
AI summary
The present invention provides a method and system for detecting changes in areas, comprising: acquiring images at different times of an area by low-altitude UAVs; extracting state features for each of backbone feature pairs; analyzing importance of each of state parts due to the state parts have different importance degrees to the whole image features under different scenes; eliminating the state features from complete features by virtue of differences in importance degrees, to obtain corresponding essential parts; determining at least one change in the area by analyzing the essential parts. The invention comprehensively considers the importance of the state parts in the input images, and the essential parts contains complete and pure information are obtained, which reflects an actual changed region.