Blended Neural Network Image Processing for Lower Hardware Cost
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing technologies require multiple neural network processors for different image functions, leading to high hardware costs due to the need for specialized circuits for each function.
Innovation Solution
A method involving training a blending neural network model by combining convolutional neural network (CNN) and generative adversarial network (GAN) models, generating blending parameters to execute multiple image processing tasks with a single neural network processor, using deep learning technology to reduce hardware costs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple neural network processors are used for different image functions, then image processing accuracy and specialization are improved, but hardware cost and device complexity increase
Solution Approach 1:
The patent implements a single neural network processor that can perform multiple image processing functions by dynamically loading different neural network models. The system includes a model management module that stores multiple models and a processing module that can switch between them based on the required image effect, allowing one processor to replace multiple specialized processors while maintaining processing accuracy through model-specific optimizations
Solution Approach 2:
The patent combines multiple neural network models into a single integrated processing system. Different image processing functions (such as denoising, super-resolution, style transfer) are merged into one processor through the model management module, which loads and executes appropriate models based on the input requirements, thereby reducing hardware complexity while preserving the specialized capabilities of each function
2Adaptability or versatility
If multiple neural network processors are used for different image functions, then image processing capability is improved, but device complexity increases
Solution Approach 1:
The neural network processor is designed as a universal platform capable of executing various image processing tasks by loading different pre-trained models. The model management module enables the single processor to adapt to different image effects (denoising, Sharpening, super-resolution, etc.) without requiring separate dedicated hardware for each function, thus improving versatility while controlling device complexity
Solution Approach 2:
The system employs dynamic model loading and switching mechanisms that allow the processor to change its functionality based on real-time requirements. The model management module dynamically selects and loads appropriate neural network models based on the desired image effect, enabling the device to adapt its capabilities without physical reconfiguration, thereby maintaining versatility with reduced complexity
3Device complexity
If a single neural network circuit is used for multiple image effects, then hardware cost is reduced, but processing efficiency may be affected
Solution Approach 1:
Multiple neural network models are pre-trained and stored in the model management module before runtime. When an image processing task is received, the system quickly loads the pre-prepared model corresponding to the required effect, avoiding the time-consuming process of training models on-demand. This preliminary preparation maintains processing efficiency while using a single hardware circuit
Solution Approach 2:
The system uses software-based model copying and loading instead of physical hardware replication. Multiple neural network models exist as software entities that can be loaded into the single processor as needed, replacing the need for multiple physical processors. This virtual copying approach reduces hardware costs while maintaining the ability to execute multiple functions efficiently through rapid model switching
Data Source
AI summary
An image processing method includes: training a first neural network model configured to execute a first image processing, according to multiple training data, to generate multiple first parameters associated with the first neural network model, in which the multiple first parameters includes multiple weights; training a second neural network model configured to execute a second image processing, which is different from the first image processing, according to the multiple training data and the multiple weights, to generate multiple second parameters associated with the second neural network model; and mixing the multiple first parameters with the multiple second parameters, to generate multiple blending parameters for a blending neural network model, in which the blending neural network model is configured to execute the first image processing and the second image processing on an input image, to output an optimized image.