Linear Activation CNN for Biometric Authentication Overfitting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing biometric recognition systems face challenges with overfitting, which affects the precision of biometric-based authentication due to the use of non-linear activation functions in convolutional neural networks (CNNs).
Innovation Solution
Implementing a convolutional neural network (CNN) with linear activation functions, such as ReLU, PReLU, or FReLU, at certain convolution layers to enhance biometric feature extraction and improve recognition performance by reducing overfitting.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If non-linear activation functions are used in CNN convolution layers, then the quality of fit and expressiveness of feature representation are improved, but overfitting occurs which reduces recognition precision
Solution Approach 1:
The patent changes the activation function parameter from non-linear (e.g., ReLU, Sigmoid, Tanh) to linear activation function. This parameter change in the CNN architecture resolves the overfitting issue while maintaining feature representation quality, directly addressing the contradiction between recognition precision and overfitting.
2Measurement precision
If non-linear activation functions are used to enhance feature expressiveness, then the quality of fit is improved, but the alignment of features with image positions deteriorates
Solution Approach 1:
The patent applies linear activation functions instead of non-linear ones in the CNN convolution layers. This parameter change preserves the spatial alignment information between extracted features and their corresponding positions in the original biometric image, preventing the loss of alignment information while maintaining feature expressiveness.
Data Source
AI summary
The present teaching relates to method, system, medium, and implementations for biometric-based authentication. When an input image representing biometric information of a person to be authenticated is received, a feature vector of the input image is generated based on the biometric information captured in the input image via a convolution neural network (CNN) that is realized using one or more linear activation functions. The person is authenticated based on the feature vector in accordance with a biometric-based recognition model.