Generative Network for Face Sharpening via Cascaded Super-Resolution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current deep learning technologies for face sharpening, such as those using three independent models, are inefficient and produce low-quality outputs with limited visual improvement compared to the original image, due to the tedious training process and poor model quality.
Innovation Solution
A method involving a generative network with a cascaded structure of super-resolution and down-sampling sub-networks, along with feature extraction sub-networks, that performs super-resolution and down-sampling processing on images to generate result images, and adjusts parameters based on total image loss conditions, including content, against, weighted L1, and eye region losses.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If three independent models are used for face sharpening, then the processing capability is enhanced, but the training process becomes tedious and complex
Solution Approach 1:
The patent merges three independent models (super-resolution model, face recognition model, and face alignment model) into a single integrated image processing model. This unified model performs all three functions through a single training process, eliminating the need for separate model training while maintaining the comprehensive face sharpening capability. The integration reduces training complexity and improves processing efficiency.
2Reliability
If three independent models are used for face sharpening, then the processing capability is enhanced, but the model quality and visual improvement are limited
Solution Approach 1:
The patent implements a feedback mechanism where the image processing model uses output from its own internal face recognition and alignment sub-models to guide the super-resolution processing. The face recognition model identifies key features and the face alignment model determines optimal processing parameters, which are then fed back to enhance the super-resolution output. This internal feedback loop ensures that the sharpening process is optimized for face structures, producing higher quality visual results.
3Productivity
If simple training process is used, then the training efficiency is improved, but the model quality and visual improvement are limited
Solution Approach 1:
The patent segments the image processing model into functional sub-models (super-resolution sub-model, face recognition sub-model, and face alignment sub-model) that work together within a unified training framework. Each sub-model has a specific function but they are trained jointly as a single integrated system. This segmentation allows for efficient training while maintaining high output quality, as each component can be optimized for its specific task while benefiting from the coordinated training of the entire system.
Data Source
AI summary
Provided are a training method and apparatus for an image processing image processing model, and an image processing method and apparatus. The training method comprises: acquiring a sample image and a first reference image, wherein the information quantity and resolution of the sample image are respectively lower than those of the first reference image; inputting the sample image into a generative network in an image processing model, and carrying out super-resolution processing and down-sampling processing on the sample image by means of the generative network, so as to generate and output at least one result image; determining the total image loss of the at least one result image according to the first reference image; and adjusting parameters of the generative network according to the total image loss, so that the total image loss of at least one result image output by the adjusted generative network meets an image loss condition.