Image Rotation Angle Prediction for Accurate Tilt Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing methods struggle to accurately correct the orientation of tilted digital images, whether slightly or significantly tilted, which affects aesthetics and orientation accuracy.
Innovation Solution
A machine learning-based approach is employed to generate image pairs for training, where target images are rotated by random angles within defined constraints, and a classification loss is minimized to update the model, enabling precise image rotation predictions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional edge detection methods (Hough transform) are used to correct image orientation, then the processing can be performed without machine learning infrastructure, but the accuracy of orientation correction is insufficient for both slight and significant tilts
Solution Approach 1:
The patent replaces traditional mechanical/mathematical edge detection algorithms (Hough transform) with a machine learning-based orientation correction system. The ML model learns to directly predict rotation angles from image content, substituting the step-by-step mechanical edge detection and line fitting process with a learned transformation that achieves higher accuracy for both slight and significant tilts.
Solution Approach 2:
The patent changes the approach from detecting geometric parameters (edges, lines) to predicting a single critical parameter (rotation angle) directly. The ML model is trained to output the rotation angle that needs to be applied, transforming the problem from complex geometric analysis to parameter prediction, which simplifies the processing while improving accuracy.
2Measurement precision
If a machine learning model is trained to predict image rotation angles, then orientation accuracy is improved, but the training data generation and model training process becomes complex
Solution Approach 1:
The patent performs preliminary actions by pre-generating training data through systematic rotation of images and creating corresponding rotation angle labels. The training dataset is prepared in advance with known ground truth rotation angles, allowing the model to learn from predetermined examples. This preliminary data preparation simplifies the subsequent training process and ensures consistent training quality.
3Manufacturing precision
If images are processed to determine exact upright orientation, then aesthetic quality and orientation accuracy are improved, but processing time increases
Solution Approach 1:
The patent extracts only the essential information needed for orientation correction by training the ML model to predict a single rotation angle parameter. Instead of performing complete image analysis including edge detection, line fitting, and geometric transformation calculations, the system extracts the critical rotation angle directly through the trained model, significantly reducing processing time while maintaining precision.
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.