Compressed Style Transfer Model for Mobile Terminals
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing technologies, specifically offline style transformation models, suffer from calculation redundancy and storage space waste due to fixed parameters, making them unsuitable for mobile terminals and unable to support multiple feedforward neural networks.
Innovation Solution
An image processing method and apparatus that uses a preset training model with a reduced structure, comprising three convolution layers, two residual layers, and three deconvolution layers, which deletes sample images not satisfying the image feature value extraction condition, resulting in a compressed model that can be applied to mobile terminals with improved utility.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If an offline style transformation model with fixed parameters is used, then style transformation can be performed quickly, but calculation redundancy and storage space waste occur
Solution Approach 1:
The patent extracts only the essential parameters from the complete style transformation model. By identifying and removing redundant parameters that do not significantly affect the output result, the model size is reduced from 13 MB to under 1 MB while maintaining acceptable transformation quality. This extraction principle directly addresses the storage space waste problem.
Solution Approach 2:
The patent changes the parameter set of the style transformation model by selecting a subset of parameters that are most critical for the transformation effect. Instead of using all original parameters, the optimized model uses only necessary parameters, thereby reducing storage requirements while preserving the core functionality of quick style transformation.
2Speed
If an offline style transformation model with fixed parameters is used, then style transformation can be performed quickly, but calculation redundancy occurs
Solution Approach 1:
The patent removes redundant calculation components from the model by extracting only the essential parameters. This eliminates unnecessary computational operations during style transformation, reducing calculation redundancy while maintaining the speed advantage of using a pre-trained model.
Solution Approach 2:
The patent applies partial action by using only the necessary portion of the original model parameters rather than the complete set. This partial parameter approach avoids excessive calculation associated with processing all original parameters, thereby reducing calculation redundancy while maintaining acceptable transformation performance.
3Quantity of substance
If a compressed model is created to reduce storage space, then the model can be applied to mobile terminals, but model complexity must be reduced
Solution Approach 1:
The patent extracts the core functional parameters from the complex original model, creating a simplified version that fits within mobile terminal storage constraints. By removing non-essential parameters and structural elements, the model complexity is reduced while retaining the fundamental style transformation capability, enabling deployment on mobile devices.
Solution Approach 2:
The patent changes the model parameters to a optimized subset that maintains transformation quality while reducing model size. This parameter optimization process simplifies the model structure by eliminating redundant parameters, thereby reducing complexity to a level suitable for mobile terminal deployment while preserving the essential functionality.
4Loss of substance
If redundant parameters are removed from the model, then storage space is saved, but some parameters may affect output quality
Solution Approach 1:
The patent carefully selects and changes the parameter set by identifying which parameters are critical for output quality and which can be removed. Through this selective parameter optimization, the model achieves significant storage space reduction while maintaining acceptable output quality by preserving the essential parameters that contribute most to transformation accuracy.
Solution Approach 2:
The patent applies different quality requirements to different parameters by identifying which parameters require high precision (those affecting output quality) and which can be simplified or removed. This local quality approach ensures that critical parameters are maintained with sufficient precision while non-critical parameters are optimized for storage efficiency, resolving the contradiction between space savings and quality maintenance.
Data Source
AI summary
An image processing method and apparatus, and a computer readable medium are provided. The method includes obtaining an image. The image is processed using a preset training model that is a function relationship model of a feature sample image and an activation function of the feature sample image. The feature sample image includes an image satisfying an image feature value extraction condition. A target image is obtained that corresponds to the image according to a processing result of the preset training model.