Stylized Image Generation via Local Neural Network Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing technologies face delays and inefficiencies in generating stylized images, particularly due to high computing resource requirements and mismatches between features in processed and original images, leading to poor processing effects.
Innovation Solution
A method and apparatus for stylized image generation that involves determining initial pairing data, training a style model, processing original images to obtain style images, performing deformation processing to create target style images, and training a stylization conversion model using these images to reduce processing delays and improve feature matching.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If image processing is performed on a server using related algorithms, then stylized images can be obtained, but processing delay increases due to network transmission and server processing time
Solution Approach 1:
The patent segments the image processing function from the server and deploys it locally on the client device. The neural network model is split into multiple computing units that can process images locally, eliminating the need for continuous server communication and reducing processing delay while maintaining processing quality.
Solution Approach 2:
The patent performs preliminary action by pre-processing and optimizing the neural network model before deployment. The model is converted into a lightweight format suitable for client-side execution, and necessary computations are prepared in advance to enable fast local processing without server round-trips.
2Manufacturing precision
If algorithms with greater computing resource requirements are deployed on a server, then more complex image processing can be achieved, but deployment on client devices becomes infeasible
Solution Approach 1:
The patent applies parameter changes by optimizing the neural network model's structure and parameters for client-side deployment. The model is converted into a lightweight format with adjusted computational parameters that reduce resource requirements while preserving processing capability. This enables complex algorithms to run on devices with limited computing resources.
Solution Approach 2:
The patent extracts the essential image processing functionality from the server environment and isolates it into a standalone neural network model that can execute locally. By taking out only the necessary computational core and removing dependencies on server infrastructure, the system achieves complex processing capability on client devices without requiring full server-grade resources.
3Productivity
If standard image processing algorithms are used, then processing can be performed, but feature correspondence between processed and original images deteriorates
Solution Approach 1:
The patent implements feedback mechanisms within the neural network training process. The model learns to maintain feature correspondence by receiving feedback during training on how well it preserves original image features while applying stylization. This feedback-driven learning enables the model to achieve both processing efficiency and feature correspondence.
Solution Approach 2:
The patent uses a composite approach by combining multiple processing functions into a single integrated neural network model. The model simultaneously performs stylization, feature preservation, and optimization tasks that would require separate algorithms otherwise. This composite structure improves both efficiency and feature correspondence by coordinating these functions within one unified system.
Data Source
AI summary
Embodiments of the present disclosure provide a stylized image generation method and apparatus, an electronic device, and a storage medium. The method comprises: determining a plurality of initial pairing data, and performing training on the basis of the initial pairing data to obtain a style model to be used; determining, on the basis of a preset screening condition, original images to be processed from original images, and processing, on the basis of said style model, each original image to be processed to obtain a style image to be used; performing deformation processing on the style image to be used to obtain a target style image, and taking each original image to be processed and the target style image corresponding thereto as stylized pairing data; and training, on the basis of the stylized pairing data, a stylized conversion model to be trained to obtain a target stylized conversion model.