Generative Adversarial Network for Intensity Image Noise Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Intensity images, such as those from OCT, sonography, and lidar, suffer from speckle noise, which degrades image quality and makes it difficult to diagnose medical issues or measure structural features accurately, as current methods require multiple images and manual intervention, lacking efficiency and consistency.
Innovation Solution
A neural network-based generative adversarial network is employed to enhance intensity images by minimizing loss functions, including mean squared error and perceptual loss, to reduce speckle noise and improve image quality without the need for additional images or manual intervention, using residual blocks and discriminator networks to maintain realistic image enhancement.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional methods are used to reduce speckle noise, then image quality can be improved, but multiple images and manual intervention are required, reducing productivity and increasing time consumption
Solution Approach 1:
The patent replaces manual intervention and conventional image processing methods with an automated neural network system. The generative adversarial network automatically processes single intensity images to reduce speckle noise, eliminating the need for manual operations and multiple image acquisitions, thus improving both image quality and processing efficiency
Solution Approach 2:
The neural network is pre-trained on datasets of intensity images with and without speckle noise, enabling it to automatically recognize and reduce noise patterns. This preliminary training allows the system to perform noise reduction on new images without requiring manual guidance or multiple acquisitions, resolving the contradiction between quality improvement and productivity
2Object-affected harmful factors
If conventional noise reduction methods are applied, then speckle noise can be reduced, but edge details and structural features may be lost, worsening measurement precision
Solution Approach 1:
The generative adversarial network applies different processing characteristics to different regions of the image. The discriminator component specifically evaluates edge regions and structural features, ensuring that noise reduction is applied selectively without blurring important diagnostic details. This local differentiation resolves the contradiction between noise reduction and detail preservation
Solution Approach 2:
The adversarial training mechanism provides continuous feedback between the generator and discriminator networks. The discriminator evaluates whether processed images retain realistic structural features and edge details, providing feedback that guides the generator to preserve these features while reducing noise. This feedback loop ensures measurement precision is maintained while eliminating speckle noise
3Reliability
If multiple images are acquired to improve image quality, then diagnostic accuracy can be enhanced, but the duration of imaging and time consumption increase
Solution Approach 1:
The neural network is pre-trained to recognize noise patterns and structural features from datasets containing multiple reference images. This preliminary learning enables the system to achieve diagnostic accuracy equivalent to multiple images by processing a single image, eliminating the need for repeated acquisitions and reducing imaging time while maintaining reliability
Solution Approach 2:
The generative adversarial network creates a cleaned version (copy) of the input image by removing speckle noise while preserving structural features. This generated copy provides the diagnostic accuracy that would otherwise require multiple original images, thus reducing imaging duration while maintaining reliability
Data Source
AI summary
A generative adversarial network including a generator portion and a discriminator portion is constructed. The network is configured such that the network operates to enhance intensity images, wherein an intensity image is obtained by illuminating an object with an energy pulse and measuring the return strength of the energy pulse, and wherein a pixel of the intensity image corresponds to the return strength. As a part of the configuring, a loss function of the generative adversarial network is minimized, the loss function comprising a mean square error loss measurement of a noisy intensity image relative to a mean square error loss measurement of a corresponding clean intensity image. An enhanced intensity image is generated by applying the minimized loss function of the network to an original intensity image, the applying improving an image quality measurement of the enhanced intensity image relative to the original intensity image.