Self-Supervised Deblurring via Non-Trainable Convolution Layers
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image deblurring techniques are either iterative and time-consuming or require voluminous annotated training data, making them impractical for real-time deblurring of images, especially in medical imaging where high-resolution images are crucial for accurate diagnoses.
Innovation Solution
A self-supervised deep learning model is trained using a non-trainable convolution layer that mimics the point spread function of an imaging device, allowing it to deblur images without iterative processes and annotated ground truths, by learning the inverse of the point spread function from blurred images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If iterative deblurring techniques are used, then image deblurring performance is improved, but processing time increases significantly
Solution Approach 1:
The system performs preliminary action by training the machine learning model offline using self-supervised learning with the point spread function as ground truth. This pre-training eliminates the need for iterative processing during actual image deblurring, as the model can directly output deblurred images in a single forward pass, thus resolving the contradiction between deblurring performance and processing time.
Solution Approach 2:
The patent replaces the traditional iterative mathematical optimization mechanism with a machine learning-based direct inference mechanism. Instead of using iterative algorithms that repeatedly adjust parameters to converge on a solution, the system uses a pre-trained neural network that learns the inverse mapping from blurred to sharp images, achieving both speed and quality improvements.
2Loss of time
If non-iterative deblurring techniques are used, then processing time is reduced, but voluminous amounts of annotated training data are required
Solution Approach 1:
The system applies self-service by using the point spread function itself as the ground truth for training. Instead of requiring external annotated image pairs, the method leverages the known physical blur characteristics of the imaging system to create self-contained training data, eliminating the need for manual annotation and reducing data collection requirements while maintaining fast inference speeds.
Solution Approach 2:
The point spread function serves as an intermediary that bridges the gap between the blurred input images and the desired sharp output. By using the point spread function as a mediator in the training process, the system can generate synthetic training data without requiring actual annotated image pairs, thus reducing the quantity of training data needed while enabling non-iterative processing.
3Measurement precision
If annotated training data is used, then deblurring accuracy is improved, but data acquisition becomes impracticable
Solution Approach 1:
The patent applies copying by creating synthetic training data that replicates the characteristics of real blurred images. Instead of requiring actual annotated image pairs that are difficult to obtain, the system generates virtual training data by convolving sharp images with the point spread function, creating realistic blurred image pairs for training without needing to acquire difficult-to-obtain annotated data.
Solution Approach 2:
The point spread function acts as an intermediary that enables the creation of training data without requiring direct acquisition of annotated images. By using the point spread function as a mediator, the system can synthesize training data that accurately represents the imaging system's characteristics, making data acquisition practical while maintaining high deblurring accuracy.
Data Source
AI summary
Systems/techniques that facilitate self-supervised deblurring are provided. In various embodiments, a system can access an input image generated by an imaging device. In various aspects, the system can train, in a self-supervised manner based on a point spread function of the imaging device, a machine learning model to deblur the input image. More specifically, the system can append to the model one or more non-trainable convolution layers having a blur kernel that is based on the point spread function of the imaging device. In various aspects, the system can feed the input image to the model, the model can generate a first output image based on the input image, the one or more non-trainable convolution layers can generate a second output image by convolving the first output image with the blur kernel, and the system can update parameters of the model based on a difference between the input image and the second output image.