Face Recognition Neural Network Resisting Angle and Occlusion Interference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional face recognition methods are limited in their ability to handle angle and occlusion interference, leading to unsatisfactory recognition performance, especially when dealing with side faces and images with glasses.
Innovation Solution
A fast face recognition method that uses a large dataset to train a neural network with Inception-Resnet V1 structure, combining softmax and center loss functions, and employing GPU acceleration to extract features and calculate Euclidean distances, allowing for robust comparison of face images regardless of angle and occlusion.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional shallow structure methods are used to extract face image features, then the method is simple and fast, but the recognition accuracy is limited and unsatisfactory
Solution Approach 1:
The patent transforms the face recognition approach by changing the parameter of network depth, transitioning from shallow structures to deep neural networks with multiple convolutional layers. This parameter change enables the extraction of more sophisticated hierarchical features, significantly improving recognition accuracy while maintaining computational feasibility through optimized architecture design.
Solution Approach 2:
The patent introduces the dimension of network depth by stacking multiple convolutional layers, transforming the traditional single-layer or few-layer architecture into a deep hierarchical structure. This dimensional expansion allows the model to learn features at multiple abstraction levels, from edge detection to complex facial patterns, thereby resolving the accuracy-complexity contradiction.
2Reliability
If standard training datasets are used without sufficient side faces and occluded images, then the training process is fast, but the recognition performance under angle and occlusion interference is poor
Solution Approach 1:
The patent applies preliminary action by proactively augmenting the training dataset with side face images and occluded face images before training begins. By pre-including these challenging cases in the training data, the model learns to handle angle and occlusion interference from the outset, improving robustness without requiring excessive additional training data during deployment.
Solution Approach 2:
The patent implements preliminary anti-action by introducing adversarial examples (side faces and occluded faces) into the training set beforehand. This allows the model to develop resistance to angle and occlusion interference during training, effectively counteracting these harmful factors before they can degrade recognition performance in real applications.
3Measurement precision
If deep neural networks are trained with large datasets to improve accuracy, then recognition performance improves, but training time increases and convergence becomes slow
Solution Approach 1:
The patent applies segmentation by dividing the deep neural network into multiple convolutional layers, each responsible for extracting features at a specific level of abstraction. This segmentation allows for more efficient training compared to monolithic architectures, as each layer can be optimized independently and gradients can flow more effectively through the hierarchical structure, reducing overall training time while maintaining high accuracy.
4Reliability
If the training set contains more side faces and occluded images (10%-20% proportion), then angle and occlusion resistance improves, but the training data processing complexity increases
Solution Approach 1:
The patent applies universality by using a unified data processing pipeline that handles both standard face images and challenging cases (side faces, occluded faces) through the same convolutional neural network architecture. The multi-functional network processes diverse input types uniformly, extracting relevant features regardless of image quality or orientation, thereby improving angle and occlusion resistance without requiring separate processing pathways that would increase complexity.
Data Source
AI summary
The present invention provides an angle interference resistant and occlusion interference resistant fast face recognition method, comprising: first collecting a training set of images which have been detected and cropped, adjusting and expanding the training set, and conducting standardized pre-processing; inputting the same into a constructed neural network for training, and saving a parametric model; adjusting test data to a suitable size and number of channels, and also conducting standardized pre-processing; inputting the same into a prediction network to obtain feature vectors of face images; and determining whether two faces are from the same person by calculating a distance between the feature vectors of the face images.