Ear Recognition Using CNN and Handcrafted Feature Fusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing ear recognition technologies face challenges in unconstrained settings due to pose variation, different acquisition devices, low resolution, illumination variations, and occlusions, which affect recognition performance and make it difficult to identify individuals accurately in real-world scenarios.
Innovation Solution
A convolutional neural network (CNN) based framework that combines CNN-learned features and handcrafted features for ear normalization and description, using a two-stage landmark detector to normalize images and generate discriminant features, which are then fused to improve recognition accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional ear recognition methods are used on unconstrained ear images, then the system is simple to implement, but recognition accuracy deteriorates due to pose variation, illumination variations, and occlusions
Solution Approach 1:
The patent segments the ear recognition process into multiple specialized modules: landmark detection module, geometric normalization module, handcrafted feature extraction module, and deep learning feature extraction module. Each module handles a specific aspect of the recognition task, allowing the system to address pose variation, illumination variations, and occlusions through coordinated operation of these segmented components.
Solution Approach 2:
The patent combines multiple feature extraction approaches (handcrafted features like LBP, HOG, Gabor filters and deep learning features from CNNs) to create a composite feature representation. This composite approach leverages the strengths of different feature types to achieve robust recognition accuracy under varying conditions while maintaining a structured system architecture.
2Reliability
If geometric normalization is applied to address pose variation, then recognition accuracy improves, but processing time increases
Solution Approach 1:
The patent applies geometric normalization as a preliminary step before feature extraction. By detecting landmarks and normalizing the ear geometry first, the system prepares the input data in an optimal state for subsequent feature extraction, reducing the computational burden during recognition and improving overall processing efficiency while maintaining accuracy.
3Reliability
If multiple feature types are fused to improve recognition robustness, then recognition accuracy under occlusion improves, but computational complexity increases
Solution Approach 1:
The patent merges handcrafted features (LBP, HOG, Gabor) with deep learning features from CNNs into a unified feature representation. This combination leverages the complementary strengths of different feature types: handcrafted features provide robustness to local variations and occlusions, while deep learning features capture global structural information, together achieving superior recognition accuracy under challenging conditions.
Data Source
AI summary
A system and method for identifying a subject based upon ear recognition using a convolutional neural network (CNN) and handcrafted features, wherein an ear in an image is cropped using ground truth annotations and landmark detection is performed to obtain the information required to normalize pose and scale variations. The normalized images are then described by different feature extractors and matched through distance metrics. Finally, scores are fused and a subject identification decision is made.