Integrated Deep Learning Network for Face and Background Composition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for composing a background and a face using deep learning networks require multiple networks, leading to high computational costs and slow processing times due to excessive GPU computations.
Innovation Solution
Integration of face extraction, color correction, and background creation deep learning networks into a single end-to-end network, utilizing adaptive object normalization and feature vector masks to generate a composite image, reducing GPU computations by about ⅓ and enhancing processing speed.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If multiple deep learning networks (face extraction, color correction, background composition) are used to compose background and face, then the quality of composite image is improved, but computational cost and processing time increase significantly
Solution Approach 1:
The patent merges face extraction network, color correction network, and background composition network into a single integrated deep learning network. This consolidation allows the system to perform all necessary operations (face extraction, color correction, and background composition) within one unified model, eliminating the need to sequentially execute multiple separate networks and thereby significantly reducing computational cost and processing time while maintaining image quality
2Manufacturing precision
If multiple deep learning networks are used to compose background and face, then the naturalness of composition is improved, but server cost increases
Solution Approach 1:
The patent consolidates multiple specialized networks into one integrated network that performs face extraction, color correction, and background composition simultaneously. This merging reduces the total computational resources required, thereby lowering server costs and energy consumption while preserving the naturalness of the composed image through the coordinated operation of integrated modules
3Measurement precision
If multiple deep learning networks are used for face and background composition, then composition accuracy is improved, but time consumption increases
Solution Approach 1:
The patent integrates multiple processing stages into a single deep learning network that operates in one forward pass. This eliminates the sequential time consumption of executing multiple separate networks while maintaining composition accuracy through the coordinated transformation of features and masks within the unified model architecture
Data Source
AI summary
The present invention relates to a method and an apparatus for composing a background and a face by using a deep learning network, comprising: receiving an input of an original face image and a converted face image, and extracting data preprocessing and feature vectors for each image; generating a face feature vector mask from the extracted feature vectors; and generating a composite image by performing adaptive object normalization on the basis of the generated face feature vector mask.