Image Orientation Correction Using Federated Tilt Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image tilt correction techniques face challenges in training machine learning models due to the limitations of synthetically created datasets with decreased image quality and user-labeled datasets that are restricted in scene diversity and tilt angle accuracy, leading to poor generalization and accuracy in correcting image tilts.
Innovation Solution
A method for generating a training dataset using federated learning, where client devices capture images and determine their tilt angles, storing metadata to create a diverse dataset that maintains privacy, and augmenting data to improve non-bias scores, while allowing model weights to be shared for updates.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If synthetically created training datasets are used, then training data availability is improved, but image quality deteriorates due to rotation operation
Solution Approach 1:
The patent uses real images captured by user devices as training data instead of synthetic copies. This preserves the original image quality while providing sufficient training data quantity through federated learning aggregation across multiple devices.
Solution Approach 2:
User devices automatically capture and contribute their own images to the training dataset without requiring external synthesis. The devices self-generate training data through normal image capture activities, eliminating the need for artificial image creation and rotation.
2Measurement precision
If user-labeled datasets are used, then training data accuracy is improved, but scene diversity deteriorates due to user camera handling limitations
Solution Approach 1:
The system uses sensor data (accelerometer, gyroscope) to provide accurate tilt angle feedback labels for captured images. This automated feedback mechanism ensures precise labeling while maintaining scene diversity, as users can capture images in various orientations without manual labeling constraints.
Solution Approach 2:
The patent replaces manual user labeling with automated sensor-based tilt angle detection. This substitution eliminates the mechanical limitation of user camera handling while preserving accurate tilt angle measurement through device sensors.
3Productivity
If centralized training is used, then model training efficiency is improved, but data privacy deteriorates due to data sharing requirements
Solution Approach 1:
The patent segments the training process into distributed components across multiple user devices. Each device trains local model copies independently using its own data, then shares only model updates (not raw data) with the central server for aggregation. This segmentation maintains training efficiency while preserving data privacy.
Solution Approach 2:
The system introduces model weight updates as an intermediary between raw training data and the central server. Instead of directly sharing sensitive image data, devices share processed model parameters that convey training information without exposing original data, thus maintaining privacy while enabling centralized model improvement.
4Adaptability or versatility
If training dataset is expanded to include more tilt angles, then model generalization is improved, but data bias deteriorates due to limited user capture patterns
Solution Approach 1:
The patent adds the dimension of device orientation by capturing images in both portrait and landscape modes across multiple user devices. This multi-dimensional approach ensures comprehensive coverage of tilt angles and orientations, eliminating data bias while improving model generalization to various capture scenarios.
Data Source
AI summary
Broadly speaking, the present techniques generally relates to methods, systems and apparatuses for performing image orientation correction, i.e. correcting or adjusting images that are tilted. In particular, the present application relates to a computer-implemented method for generating a training dataset for training a machine learning, ML, model using federated learning to perform image orientation correction, and methods for training the ML model using the generated training dataset. Advantageously, the method to generate a training dataset enables a diverse training dataset to be generated while maintaining user data privacy, where the diversity refers to the range of image tilt angles represented by the dataset. The present application also provides methods for training the ML model using the generated training dataset.