Image Rotation Correction Using Convolutional Neural Network
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image rotation correction methods struggle to accurately correct image rotation, especially in images lacking straight lines or horizon lines, and are not effective for a wide variety of image types, including portraits and landscapes, due to limitations in detecting small rotation angles and relying on human intuition or low-level vision features.
Innovation Solution
A deep neural network architecture utilizing a convolutional neural network with rectangle-shaped depthwise convolutions and an adapted loss function to detect rotation angles and orientations, capable of handling diverse image types by incorporating both classification and regression tasks, and suitable for real-time application on portable devices.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional image rotation correction methods are used, then the correction process is simple, but the accuracy is poor especially for images without straight lines or horizon lines
Solution Approach 1:
The patent replaces conventional mechanical/algorithmic line detection methods with a deep neural network system that uses learned features from training data to detect rotation angles, achieving superior accuracy especially for images without clear geometric structures
Solution Approach 2:
The patent transforms the rotation correction problem from detecting geometric features (lines, horizons) to predicting rotation parameters directly through a neural network that outputs rotation angle and orientation as continuous values, bypassing the need for explicit feature detection
2Measurement precision
If deep neural network with rectangle-shaped filters is used, then the accuracy for detecting rotation angles improves, but the computational complexity increases
Solution Approach 1:
The patent employs rectangle-shaped depthwise separable convolutions with specific aspect ratios that are locally optimized for detecting line structures and orientation patterns, providing targeted computational efficiency for the specific task of rotation detection rather than generic feature extraction
3Reliability
If an adapted loss function with multiple terms is used, then the training accuracy and orientation detection improve, but the training complexity increases
Solution Approach 1:
The patent segments the orientation detection task into two independent prediction branches: one for rotation angle and one for orientation direction, each optimized with specific loss function terms that address their respective prediction challenges independently
4Adaptability or versatility
If conventional methods relying on horizon lines are used, then the system is simple, but it cannot handle diverse image types including portraits and landscapes
Solution Approach 1:
The patent creates a universal rotation detection system trained on diverse image types including portraits, landscapes, and architectural photos, enabling the model to generalize across different image categories without requiring horizon lines or straight edges
Data Source
AI summary
Systems and methods are disclosed for automatically correcting a rotation of an image. One method comprises receiving an input image to be processed for rotation correction, and providing the input image to a machine learning model configured to predict a rotation angle and an orientation of an image. The machine learning model may comprise a convolutional neural network utilizing one or more rectangle-shaped filters to extract features from the input image. A rotation angle may be determined using a first branch of fully connected layer configured for a regression task of the machine learning model, and an orientation of the image may be determined using a second branch of fully connected layer configured for a classification task of the machine learning model. The input image may then be rotated based on the determined rotation angle and orientation.