Satellite Road Segmentation via Ground Truth Dilation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing road detection methods using convolutional neural networks, such as U-Net, face challenges in preserving the connectivity of road feature points in satellite images, leading to disconnection issues and reduced precision in segmentation maps.
Innovation Solution
The proposed method applies a dilation operation to the ground truth of satellite images during training and uses a U-Net-based convolutional neural network to generate an original segmentation map, followed by an erosion operation to produce a final segmentation map, with the size of the structuring element determined by precision, recall, and IoU values to improve connectivity and reduce false negatives.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a U-Net-based convolutional neural network is used for road region detection, then the segmentation performance is improved, but the connectivity of road feature points is lost resulting in disconnected output
Solution Approach 1:
The patent applies a dilation operation to the ground truth labels before training the U-Net model. This preliminary action expands the road region boundaries in the training data, enabling the model to learn and preserve connectivity patterns that would otherwise be lost during segmentation. The dilated ground truth serves as a buffer that maintains road continuity even when the model's segmentation boundaries are imprecise.
2Stability of the object's composition
If the ground truth is dilated during training, then the connectivity of road feature points is preserved, but the manufacturing precision of the segmentation map decreases
Solution Approach 1:
The dilation operation is applied to ground truth labels before model training to prevent connectivity loss. This preliminary expansion of road regions in training data allows the model to learn connectivity patterns that compensate for the imprecision introduced by dilation, ultimately producing segmentation maps that maintain both connectivity and acceptable precision.
Solution Approach 2:
The patent adjusts the dilation kernel size as a parameter to balance connectivity preservation and segmentation precision. By optimizing the dilation radius, the method finds a sweet spot where road connectivity is maintained while minimizing the loss of segmentation accuracy. This parameter tuning allows the system to adapt to different road characteristics and image resolutions.
Data Source
AI summary
Disclosed are convolutional neural network-based road detecting apparatus and method and a convolutional neural network-based road detecting method according to an exemplary embodiment of the present disclosure includes applying a dilation operation to a ground truth for a road image of a learning image, training an inference model which detects a road region from a satellite image based on the learning image and the ground truth to which the dilation operation is applied, and receiving a prediction target image and generating an original segmentation map in which a road region is detected from the prediction target image by means of the inference model.