Image Parameter Adjustment Using Attention Loss for Face Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Vision transformers (ViT) based feature extractors face challenges in image recognition tasks due to focusing on irrelevant regions, leading to reduced attention on important information, thereby decreasing recognition performance, especially in face authentication tasks.
Innovation Solution
An image processing device that divides input images into partial images, converts them into tokens, and adjusts parameters to reduce attention loss by enhancing the relevance between tokens using an attention map and learning units to focus on important regions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a ViT-based feature extractor uses self-attention mechanisms to extract features from partial images, then feature extraction capability is improved, but attention is diverted to irrelevant regions containing noise, reducing recognition performance
Solution Approach 1:
The patent applies preliminary anti-action by introducing a loss function that penalizes attention to irrelevant regions before the main recognition task. The gradient of this loss function is used to adjust the attention map, proactively preventing the model from focusing on noisy regions during the training process, thereby preserving attention for important regions.
Solution Approach 2:
The patent converts the harmful effect of attention to irrelevant regions into a benefit by using the attention map of irrelevant regions as a signal for optimization. The loss function is designed to minimize attention to these regions, and the gradient of this loss is used to adjust the model parameters, turning the problematic attention distribution into a useful training signal.
2Reliability
If self-attention mechanisms collect information from all partial images, then comprehensive feature extraction is achieved, but information from irrelevant regions contaminates the feature vector, decreasing recognition performance
Solution Approach 1:
The patent converts the harmful noise from irrelevant regions into a useful training signal. By designing a loss function that measures attention to irrelevant regions and using its gradient to adjust the model, the previously harmful attention distribution becomes a beneficial optimization guide that reduces the impact of noise on recognition performance.
Solution Approach 2:
The patent implements feedback by using the attention map as an intermediate signal that is evaluated and fed back to adjust the model. The loss function computes the attention to irrelevant regions, and this feedback is used to update the attention mechanism through gradient descent, creating a closed-loop system that continuously improves attention distribution.
3Reliability
If a weighted map is used to enhance robustness against position-dependent noise, then robustness to positional noise is improved, but robustness against noise with low position dependence (such as concealment by objects) cannot be enhanced
Solution Approach 1:
The patent achieves universality by creating an attention adjustment mechanism that handles multiple types of noise through a unified approach. The loss function based on attention maps works for both position-dependent noise and concealment by objects, as it directly targets the attention distribution regardless of the noise type, making the solution versatile across different noise scenarios.
Solution Approach 2:
The patent applies parameter changes by dynamically adjusting the attention map parameters based on the computed loss. Instead of using a fixed weighted map, the attention weights are modified through gradient descent using the loss gradient, allowing the model to adapt its attention parameters to handle different noise types effectively.
Data Source
AI summary
An image processing device including a processor or circuit configured to: acquire an input image; divide the input image into a plurality of partial images to obtain a partial image sequence; convert the partial image sequence into a token sequence by respectively converting each of the partial images includes in the partial image sequence into a token having a fixed-dimensional vector; obtain an encoded representation sequence based on an attention map indicating a degree of association between tokens and the token sequence; obtain a feature vector from the encoded representation sequence; and adjust at least parameters of the feature vector to reduce an attention loss value. The attention loss value corresponds to an error between a target value of the degree of association between the partial images and a value indicated by the attention map.