Shared DNN Backbone for Joint Facial Feature Extraction and Quality Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing facial recognition systems face inefficiencies due to the complexity and resource-intensive nature of separate Deep Neural Network (DNN) backbones for facial feature extraction and quality prediction, leading to reduced accuracy and increased computational requirements.
Innovation Solution
A shared DNN backbone is trained for both facial feature extraction and quality prediction using a custom-labeled training dataset, allowing the model to jointly perform these tasks and reduce CPU and memory consumption by up to 50%.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If separate DNN backbones are used for facial feature extraction and quality prediction, then task specialization is improved, but device complexity and computational resource consumption increase
Solution Approach 1:
The patent merges two separate DNN backbones into a single shared backbone that performs both facial feature extraction and quality prediction tasks. This consolidation reduces system complexity and computational overhead while maintaining functional separation through distinct output heads, directly resolving the contradiction between specialization and complexity.
Solution Approach 2:
The shared DNN backbone is designed to perform multiple functions simultaneously - extracting facial features for recognition and predicting image quality scores. This multi-functional design eliminates the need for separate specialized models, reducing device complexity while maintaining task performance through a unified architecture.
2Reliability
If separate DNN backbones are used for facial feature extraction and quality prediction, then functional independence is improved, but use of energy and computational resources increase
Solution Approach 1:
By merging separate DNN backbones into a single shared model, the patent reduces redundant computational operations and resource consumption. The shared backbone processes inputs once and generates both feature embeddings and quality predictions simultaneously, eliminating duplicate computations while maintaining functional independence through separate output layers.
Solution Approach 2:
The patent recovers and reuses intermediate feature representations from the shared backbone for both facial recognition and quality assessment tasks. Instead of processing inputs separately through independent backbones, the system extracts features once and leverages these representations for multiple purposes, reducing overall computational energy consumption.
3Device complexity
If a shared DNN backbone is used for both facial feature extraction and quality prediction, then device complexity is reduced, but measurement precision may deteriorate
Solution Approach 1:
The shared DNN backbone is segmented into distinct functional components with separate output heads - one for facial feature extraction and another for quality prediction. This segmentation allows each task to have specialized processing pathways while sharing the common backbone, maintaining measurement precision for both tasks despite the unified architecture.
Solution Approach 2:
Different parts of the network are optimized for different tasks - the shared backbone learns general facial image representations, while task-specific output layers are fine-tuned for their respective functions. This local optimization ensures that each task maintains high precision despite sharing the underlying backbone architecture.
4Reliability
If separate DNN backbones are used for facial feature extraction and quality prediction, then task-specific optimization is improved, but productivity decreases
Solution Approach 1:
The patent combines separate processing pipelines into a single unified model that performs both facial feature extraction and quality prediction in one forward pass. This merging eliminates redundant processing steps and improves computational efficiency, directly enhancing productivity while maintaining task-specific performance through specialized output layers.
Solution Approach 2:
The shared backbone enables continuous extraction of useful features that serve both tasks simultaneously. Instead of sequentially processing inputs through separate models, the system performs both functions in a continuous, integrated manner, maximizing processing efficiency and productivity without sacrificing task-specific optimization.
Data Source
AI summary
Systems and methods for joint feature extraction and quality prediction using a shared machine learning model backbone and a customized training dataset are provided. According to an embodiment, a computer system receives a training dataset including example images each labeled with a particular category of a set of categories, and trains a deep neural network (DNN) based on the training dataset to jointly perform for an input image (i) facial feature extraction in accordance with the facial feature extraction algorithm and (ii) a quality scoring in accordance with a quality prediction algorithm. In the embodiment, the DNN, once trained with the training dataset labeled using a custom labeling scheme is used for the facial feature extraction and the quality prediction. The facial feature extraction algorithm and the quality prediction algorithm share a common DNN backbone of the DNN.