Facial Image Object Removal Using Attention Maps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current facial image beautification techniques require large amounts of paired data for model training, making it difficult to collect and increasing training costs.
Innovation Solution
A method and apparatus for image processing that uses a model trained on unpaired data, specifically generating an attention map of the object to be removed, allowing for the removal of predetermined objects from facial images without the need for paired data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If paired data is used for model training, then the accuracy of object removal is improved, but the data collection difficulty and training cost increase
Solution Approach 1:
The patent introduces an attention map as an intermediary element that guides the object removal process. The attention map highlights the target object's location and characteristics, enabling the model to focus on relevant features without requiring paired training data. This intermediary mechanism bridges the gap between unpaired data and accurate object removal, resolving the contradiction by maintaining high accuracy while eliminating the need for difficult paired data collection
Solution Approach 2:
The patent changes the training paradigm from paired data to unpaired data, fundamentally altering the data parameter. By using unpaired data with attention maps, the model learns to identify and remove objects based on attention guidance rather than direct pixel-to-pixel mapping. This parameter change enables training with easily collected unpaired data while maintaining removal accuracy through the attention mechanism
2Manufacturing precision
If paired data is used for model training, then the object removal quality is improved, but the training cost increases
Solution Approach 1:
The attention map serves as a mediator that enables high-quality object removal without requiring large quantities of paired training data. By providing spatial and feature guidance through the attention map, the model can achieve precise object removal using unpaired data, thus reducing the training data quantity requirement while maintaining removal quality
Solution Approach 2:
The patent uses the attention map to copy or transfer the object's spatial and feature information to guide the removal process. Instead of requiring numerous paired examples, the attention map captures essential object characteristics that can be applied during inference, reducing the need for extensive paired training data while preserving removal quality
Data Source
AI summary
Embodiments of the disclosure provide a method, apparatus, electronic device (700), and storage medium for image processing. The method includes: inputting a to-be-processed facial image to a predetermined model (S110); and outputting, by the predetermined model, a target facial image (S120) with a predetermined object removed from the to-be-processed facial image; wherein the predetermined model is trained and generated based on an attention map (a) of the predetermined object. Since the predetermined model is trained based on the attention map (a) of the predetermined object, it is able to first generate the attention map (a) of the predetermined object based on unpaired data training, and then train to remove the predetermined object from the facial image with the attention map (a) of the predetermined object.