Dual CNN Stylized Image Processing Architecture
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image processing methods struggle to achieve a balance between stylized migration effects and processing speed, with existing deep learning technologies requiring significant computing resources and being too complex for real-time or commercial applications.
Innovation Solution
The proposed method employs a dual-CNN architecture with sequential convolutional and up-sampling layers to extract and combine content and style features from input images, allowing for efficient stylized migration while maintaining high processing speed.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If deep learning technology with CNN is applied to image processing, then stylized migration effect is improved, but computing resources and system complexity increase significantly
Solution Approach 1:
The patent divides the image processing system into three independent CNN modules: a first CNN for extracting content features, a second CNN for extracting style features, and a third CNN for reconstructing the stylized image. This segmentation allows each module to specialize in specific tasks, reducing overall system complexity while maintaining high stylized migration effect quality.
Solution Approach 2:
The patent extracts and separates content features and style features into different CNN modules. The first CNN extracts only content features, the second CNN extracts only style features, and the third CNN combines them for reconstruction. This extraction approach simplifies each individual module while achieving complex stylized migration effects through modular composition.
2Manufacturing precision
If deep learning technology with CNN is applied to image processing, then stylized migration effect is improved, but processing speed decreases
Solution Approach 1:
By segmenting the processing into three separate CNN modules that can operate independently and in parallel, the patent reduces the computational bottleneck. The content extraction, style extraction, and image reconstruction can be performed simultaneously on different data, significantly improving processing speed while maintaining high stylized migration effect quality.
Solution Approach 2:
The patent introduces a dimensional separation where content and style features are processed in separate computational dimensions (different CNN modules). This allows for parallel processing and reduces the sequential computational burden, thereby improving processing speed without sacrificing the quality of stylized migration effects.
3Adaptability or versatility
If existing deep learning technologies are used, then image processing capability is improved, but the system becomes too complex for real-time applications
Solution Approach 1:
The patent segments the image processing capability into three specialized modules that can be independently optimized and deployed. This modular architecture makes the system more adaptable to different real-time application scenarios while reducing the complexity barrier for deployment, as each module can be independently tuned for specific performance requirements.
Solution Approach 2:
The patent creates a universal image processing system where the three CNN modules can handle various image processing tasks (content extraction, style extraction, stylized reconstruction) through a consistent modular architecture. This universality enables real-time applications across different domains while maintaining manageable system complexity through standardized module interfaces.
Data Source
AI summary
An image processing method and an image processing device are provided. The image processing method includes steps of extracting a feature of an inputted first image by a first CNN, and reconstructing and outputting an image by a second CNN. The first CNN includes a plurality of first convolutional layers connected sequentially to each other and a plurality of first pooling layers each arranged between respective adjacent first convolutional layers, and each first convolutional layer is configured to generate and output a first convolutional feature. The second CNN includes a plurality of second convolutional layers connected sequentially to each other and a plurality of composite layers each arranged between respective adjacent second convolutional layers, and each composite layer is an up-sampling layer.