Generative Neural Network Image Conversion for Diverse, Consistent Outputs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing systems using deep neural networks are complex to train and lack diversity in output images, leading to inconsistent and limited conversion results.
Innovation Solution
An image conversion method utilizing generative neural networks trained with content and result diversity loss functions to ensure consistency and diversity in output images, employing downsampling, residual, and upsampling modules with noise inputs to generate varied conversion results.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If deep neural network is used for image conversion processing, then image conversion capability is achieved, but the system structure becomes complicated and difficult to train
Solution Approach 1:
The neural network is divided into three independent modules: downsampling module, residual module, and upsampling module. Each module has a specific function and can be trained independently, reducing the overall training difficulty while maintaining image conversion capability.
Solution Approach 2:
Random noise is introduced as an intermediary element during the image conversion process. The noise is added in the residual module to generate diverse output images, enabling the system to achieve variety without complicating the overall structure.
2Adaptability or versatility
If deep neural network is used for image conversion processing, then image conversion is achieved, but the output image lacks diversity
Solution Approach 1:
The system dynamically adjusts the input by adding random noise to the original image before processing. This dynamic input variation enables diverse output images while keeping the network structure simple and easy to train.
Solution Approach 2:
Random noise with varying parameters is added to the input image, changing the input parameters dynamically. This allows the generation of diverse output images without requiring a complex network structure, as the diversity comes from parameter variation rather than structural complexity.
3Adaptability or versatility
If noise image channel is added to generate diverse output, then output diversity is achieved, but input complexity increases
Solution Approach 1:
The original image and random noise are merged in the residual module through addition. This simple merging operation generates diverse output images without requiring complex input processing, as the noise is directly added to the existing image data.
Data Source
Figure 1~2
Figure 3
Figure 4~5
AI summary
Provided are an image processing method, processing apparatus and processing device. Image conversion is realized by utilizing a generative neural network combined with image content features, so that a converted output image has diversity and is kept consistent with an input image. The image processing method comprises: acquiring an input image (110); utilizing a generative neural network to perform image conversion processing on the input image (120); and outputting an output image that has been subjected to image conversion processing (130). The input image has N channels, N being a positive integer greater than or equal to 1; an input of the generative neural network comprises a noise image channel and the N channels of the input image; and an output of the generative neural network is an output image comprising N channels.