Self-Learning Image Denoising via Similar Element Pairing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning techniques for image denoising require paired training data, which can be expensive or unavailable, limiting their effectiveness in denoising real-world images corrupted by various noises.
Innovation Solution
The method involves generating sets of similar elements for each noisy input element using a similarity module and creating training sample pairs to train an artificial neural network (ANN) unsupervisedly, allowing the network to learn noise suppression without labeled target data, leveraging the similarity and symmetry present in images across different dimensionalities.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If paired training data is used for training ANNs, then denoising performance is improved, but data acquisition cost and complexity increase
Solution Approach 1:
The system uses the noisy input data itself to generate training samples by finding similar elements within the same dataset, eliminating the need for external clean reference data. The noisy data serves dual purposes as both input and source of training targets through self-similarity comparisons.
Solution Approach 2:
Instead of requiring actual clean images, the system creates synthetic training targets by copying and adapting similar patches from the noisy dataset itself. The similar elements act as proxies for clean references, allowing the network to learn denoising patterns without accessing ground truth data.
2Measurement precision
If paired training data is used for training ANNs, then denoising performance is improved, but availability of training data decreases
Solution Approach 1:
The method enables the training system to be self-sufficient by extracting training signals from the noisy data itself. Any noisy dataset can serve as its own training source, making the approach universally applicable across different imaging modalities and noise conditions without requiring paired datasets.
Solution Approach 2:
The similar-element based training approach works across diverse image types (2D natural images, 3D CT, 4D spectral CT, MRI k-space, seismic data) and noise conditions, creating a universal training framework that adapts to any input data without requiring modality-specific paired training data.
3Measurement precision
If similar elements are generated for each noisy input element, then training sample quality is improved, but computational complexity increases
Solution Approach 1:
The system divides the training process into distinct modular stages: similarity computation, candidate selection, and pair generation. Each stage processes data independently and can be optimized separately, making the overall complex process manageable and scalable through systematic breakdown of operations.
Solution Approach 2:
Similar elements are pre-computed and stored before the actual training process begins. This preliminary similarity analysis creates an index or lookup structure that accelerates subsequent training iterations, avoiding redundant similarity computations during network training and reducing overall computational burden.
Data Source
AI summary
One embodiment provides a method of training an artificial neural network (ANN) for denoising. The method includes generating, by a similarity module, a respective set of similar elements for each noisy input element of a number of noisy input elements included in a single noisy input data set. Each noisy input element includes information and noise. The method further includes generating, by a sample pair module, a plurality of training sample pairs. Each training sample pair includes a pair of selected similar elements corresponding to a respective noisy input element. The method further includes training, by a training module, an ANN using the plurality of training sample pairs. Each set of similar elements is generated prior to training the ANN. The plurality of training sample pairs is generated during training the ANN. The training is unsupervised.