Facial Image Recognition Using Cropped Regions and Temporal Fusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image recognition technologies face challenges in improving accuracy, particularly in dynamic environments and with partial facial information, such as in intensive care units where patients often wear masks.
Innovation Solution
A neural network-based image recognition method that trains on diverse user images, employing architectures with fusion and timing neural networks to extract features from both reference and input images, and uses cropping to focus on specific facial areas for improved recognition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional image recognition methods are used, then the system is simple to implement, but the recognition accuracy deteriorates in dynamic environments and with partial facial information
Solution Approach 1:
The patent applies segmentation by dividing the facial image into multiple regions (e.g., masked region and unmasked region) and processing each region through separate neural network branches. This allows the system to focus on relevant facial features while ignoring occluded areas, thereby improving recognition accuracy without requiring a complete overhaul of the entire image processing pipeline.
Solution Approach 2:
The patent introduces temporal dimension by incorporating video frames sequences and using timing neural networks to process temporal information. This transforms the problem from static 2D image recognition to dynamic 4D (3D spatial + time) recognition, enabling the system to leverage temporal consistency and motion cues to improve accuracy in dynamic environments.
2Measurement precision
If the recognition model processes complete facial information, then recognition accuracy is maximized, but the system fails when partial facial information is available (e.g., masked faces)
Solution Approach 1:
The patent extracts and processes only the relevant portions of facial information that are visible and useful for recognition. By using region-specific neural network branches and attention mechanisms, the system extracts discriminative features from unmasked regions while excluding or down-weighting masked regions, thereby maintaining high accuracy even when complete facial information is unavailable.
Solution Approach 2:
The patent dynamically adjusts processing parameters based on the quality and completeness of input information. When partial facial information is detected, the system changes parameters such as region weighting, feature fusion strategies, and confidence thresholds to optimize recognition performance under constrained conditions.
3Reliability
If the system uses multiple neural networks with complex architectures, then recognition accuracy in dynamic environments improves, but the computational complexity and training difficulty increase
Solution Approach 1:
The patent designs a unified multi-branch neural network architecture where different branches serve multiple functions. For example, the same backbone network is used across different branches with only minor modifications for region-specific processing, and the timing neural network handles both temporal modeling and feature fusion tasks. This modular universal design improves reliability in dynamic environments while controlling overall complexity through code reuse and shared components.
Data Source
AI summary
An image recognition method and an electronic apparatus configured for image recognition are provided. A training sample set is provided to train a recognition model including neural networks to recognize a classification label to which an image to be tested belongs through the trained recognition model. The training sample set includes image sets respectively belonging to users. During the training process, training images corresponding to classification labels are obtained from a first image set in the training sample set as reference images for training; a training image is obtained from a second image set different from the first image set as an input image for training; the reference images for training and the input image for training are obtained as inputs to the neural networks for training. The input to each neural network includes at least one of the reference images for training and the input image for training.