Self-Supervised Neural Network Training for Robust Image Keypoints
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image feature extraction methods lack robustness against perspective and appearance changes, uneven keypoint distribution, and require manually constructed cost functions that are suboptimal and complex to set.
Innovation Solution
A method and apparatus for training a machine learning model using two deep neural networks to extract image features by matching keypoints with similar descriptors, reconstructing images based on these features, and optimizing networks through a self-supervised cost function that implicitly learns desired properties during training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manually constructed cost functions are used for training, then the training process can be implemented, but the cost functions are suboptimal and complex to set
Solution Approach 1:
The system uses self-supervised learning where the model generates its own training signals through image reconstruction. The second neural network reconstructs images from extracted features, and the reconstruction error automatically serves as the training cost function, eliminating the need for manual cost function design and ground truth labels.
Solution Approach 2:
Instead of directly optimizing feature extraction by designing complex cost functions, the patent inverts the approach by using image reconstruction as the optimization target. The feature extraction is improved indirectly through the reconstruction task, simplifying the training process.
2Reliability
If robust keypoint detection is achieved against perspective and appearance changes, then reliable feature matching is possible, but the training requires precise ground truth which is difficult to obtain
Solution Approach 1:
The system is self-supervised, using the input images themselves as training targets through the reconstruction task. No external ground truth data is needed - the model learns robust feature extraction by attempting to reconstruct images from its own extracted features, making training data preparation trivial.
Solution Approach 2:
The second neural network creates copies (reconstructions) of the original images from extracted features. These reconstructions serve as self-generated training targets, allowing the system to learn without requiring separate ground truth datasets.
3Measurement precision
If complex image processing operators are learned during training, then superior feature extraction is achieved, but the computational load increases
Solution Approach 1:
The system performs continuous useful action by using the same feature extraction pipeline for both the reconstruction task and the final application. The features learned during reconstruction training are directly applicable to feature matching, eliminating the need for separate optimization stages and improving training efficiency.
Data Source
AI summary
A method and an apparatus for training a machine learning model comprising a first and a second deep neural network.