Aerial Image Augmentation via Polygon Cropping to Prevent Overfitting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for increasing the number of aerial-view images for training a discriminator in ground-to-aerial cross-view matching lead to overfitting due to the use of visible edges in augmented images, which are not addressed in current techniques.
Innovation Solution
An image augmentation process involving cropping and rotation of aerial-view images to generate augmented images, where the target region is either circular or a regular polygon, preventing the discriminator from relying on edges to guess the original orientation, thereby increasing the dataset size without overfitting.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If image augmentation is performed by rotating original aerial-view images, then the number of training datasets is increased, but the discriminator overfits to the augmented training data due to visible edges
Solution Approach 1:
The patent extracts and removes the problematic edges from the aerial-view images through cropping operations. By cropping the images to eliminate visible edges before rotation, the method prevents the discriminator from learning spurious edge-based features, thus avoiding overfitting while still generating diverse training samples through subsequent rotation operations.
Solution Approach 2:
The patent segments the aerial-view images by cropping them into smaller regions with specific aspect ratios (e.g., 16:9, 9:16, 1:1) before rotation. This segmentation removes the global edge structure that causes overfitting while preserving local features, allowing the discriminator to learn robust features from multiple cropped and rotated versions of each image.
2Measurement precision
If the discriminator is trained with more aerial-view images, then the matching accuracy is improved, but the training time and computational resources increase
Solution Approach 1:
The patent applies periodic action through systematic rotation of cropped images at multiple predetermined angles (e.g., 0°, 45°, 90°, 135°, 180°, 225°, 270°, 315°). This generates multiple augmented training samples from each original image in a structured, periodic manner, efficiently expanding the training dataset without requiring proportional increases in training time, as all rotations can be computed from a single cropped image.
Solution Approach 2:
The patent creates multiple copies of each cropped aerial-view image by rotating them at different angles. These rotated copies serve as additional training samples, effectively multiplying the training dataset size without requiring additional data collection or processing time proportional to the number of copies generated, since rotation is a computationally efficient transformation.
Data Source
AI summary
An image augmentation apparatus (2000) acquires an original training dataset (40). The original training dataset (40) includes a ground-view image (42) and an original aerial-view image (44). The image augmentation apparatus (2000) performs an image augmentation including a cropping process and a rotation process on the original aerial-view image (44) to generate an augmented aerial-view image (54). In the cropping process, a target region with a circle shape or a regular polygon shape is cropped from the original aerial-view image (44). In the rotation process, the original aerial-view image (44) is rotated. An angle of the rotation is a multiple of a center angle of the target region when the target region has a regular polygon shape. The image augmentation apparatus (2000) outputs an augmented training dataset (50) including the ground-view image (52) and the augmented aerial-view image (54).