Image Rotation Using ML Tilt Prediction From Image Patches
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing methods struggle to accurately rotate digital images that are slightly or significantly tilted, affecting their aesthetics and orientation accuracy, especially when automatic or manual adjustments are used.
Innovation Solution
A machine learning-based approach is employed to generate image pairs for training, where target images are rotated randomly within defined constraints, and a classification loss is minimized to develop a model that predicts the optimal rotation, using feature vectors from divided image parts to improve orientation correction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional edge detection methods (Hough transform) are used for image rotation, then the process is computationally simpler, but the rotation accuracy deteriorates for images with slight or significant tilts
Solution Approach 1:
The patent replaces traditional mechanical edge detection algorithms (Hough transform) with a machine learning-based system. A neural network model is trained to directly predict rotation angles from image input, substituting the step-by-step mechanical edge detection process with a learned transformation that achieves higher accuracy for tilted images while maintaining computational efficiency through the trained model.
Solution Approach 2:
The patent changes the approach from detecting geometric parameters (edges, lines) to learning rotation parameters directly through machine learning. The system transforms the problem from finding edge orientations to predicting rotation angles as continuous values, allowing for more precise measurement of tilt angles in challenging image scenarios.
2Measurement precision
If more training data with various rotation angles is used, then the model's rotation prediction accuracy improves, but the training time and computational resources increase
Solution Approach 1:
The patent applies preliminary action by pre-training the machine learning model with a comprehensive dataset of images rotated at various angles before deployment. The model is trained in advance on synthetically generated image pairs with known rotation angles, so that when deployed, it can quickly predict rotations without requiring extensive real-time computation or additional training data collection.
3Adaptability or versatility
If the model is trained to handle all possible rotation angles, then the versatility improves, but the model complexity and training requirements increase
Solution Approach 1:
The patent creates a universal rotation detection model that can handle images with any rotation angle within a defined range. The neural network is designed with sufficient capacity to learn rotation patterns across the full spectrum of possible tilts, making it adaptable to diverse image orientations without requiring separate specialized models for different rotation scenarios.
Data Source
AI summary
Methods of training a machine learning model for image processing are described, as well as methods for image processing to determine a predicted rotation value for an image based on the trained machine learning model. A method of training includes utilising as a learning objective a reduction or minimisation of a classification loss. Training image pairs may be associated with classifications corresponding to rotation values for the purpose of determining the classification loss. An image may be divided into a plurality of images for determining an input to the machine learning model, for training or for image processing.