Modular Neural Networks for Diverse High-Resolution Image Conversion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing and conversion systems using deep neural networks have complex structures, are difficult to train, and produce output images with limited diversity.
Innovation Solution
An image processing method utilizing a generative neural network with down-scale, residual, and up-scale modules, combined with a super-resolution neural network, to enhance image resolution and diversity by incorporating noise images and employing loss functions for training, ensuring consistency and diversity in output images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a deep neural network is used for image processing and conversion, then the image processing capability is improved, but the system structure becomes complicated and difficult to train
Solution Approach 1:
The patent divides the deep neural network into multiple functional modules: encoder module, decoder module, and attention module. Each module has a specific function, making the overall complex system more manageable and easier to train by breaking it down into smaller, independent training units.
Solution Approach 2:
The patent introduces an attention module as an intermediary between the encoder and decoder. This attention mechanism serves as a mediator that selectively focuses on relevant features, improving the network's processing capability while maintaining a structured architecture that is easier to train compared to a fully connected deep network.
2Reliability
If a deep neural network is used for image conversion, then the processing capability is improved, but the output image lacks diversity
Solution Approach 1:
The patent employs an attention mechanism that dynamically adjusts the weight of different features during image processing. This dynamic adaptation allows the network to produce diverse output images by selectively emphasizing different features based on the input, thereby improving both processing capability and output diversity.
Solution Approach 2:
The patent changes the parameters of the attention mechanism during training to enable the network to produce diverse outputs. By adjusting attention weights and other parameters, the system can adapt to different image conversion tasks and generate varied results while maintaining high processing capability.
3Adaptability or versatility
If noise images are incorporated into the training process, then the diversity of output images is improved, but the training complexity increases
Solution Approach 1:
The patent applies noise images as input to the encoder module before the main processing occurs. This preliminary action of adding noise diversifies the training data early in the process, enabling the network to learn from varied inputs without significantly increasing overall training complexity, as the noise integration is built into the existing modular architecture.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
There are provided an image processing method, an image processing apparatus and an image processing device. The image processing method comprises: executing image conversion processing on an input image with a first noise image using a generative neural network, to output a first output image; and executing high resolution conversion processing on the first output image with a second noise image using a super-resolution neural network, to output a second output image.