Parallel Face Recognition Pipelines for Low-Resolution Multi-Face Latency
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing facial recognition technologies in mobile devices struggle to efficiently execute multiple deep neural network (DNN) operations for recognizing a plurality of low-resolution faces in complex urban spaces, leading to long latencies and difficulties in rapid recognition of moving targets.
Innovation Solution
A facial recognition method and apparatus utilizing a content and resource-adaptive parallel processing pipeline, comprising a detection pipeline, a control pipeline, and multiple recognition pipelines, to dynamically schedule and execute DNN operations based on facial recognition difficulty levels, leveraging both resource-limited mobile devices and cloud computing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If multiple DNN operations are executed simultaneously for facial recognition in complex urban spaces, then recognition accuracy and coverage are improved, but processing latency increases and device resources are overwhelmed
Solution Approach 1:
The patent segments the facial recognition workload into multiple independent DNN operations that can be executed in parallel. Different face regions detected in the input image are distributed to separate recognition pipelines, each handling specific faces independently. This segmentation enables simultaneous processing of multiple faces without requiring sequential execution, thereby improving recognition coverage while managing processing latency through parallelization.
Solution Approach 2:
The patent introduces a pipeline dimension to organize and manage multiple DNN operations. By structuring the system as a pipeline with detection, control, and recognition stages, the patent adds a temporal and organizational dimension to the processing flow. This pipeline structure allows multiple DNN operations to progress through different stages simultaneously, improving throughput while maintaining manageable latency through staged processing.
2Productivity
If multiple DNN operations are executed simultaneously for facial recognition, then processing throughput is improved, but device resource consumption exceeds available capacity in mobile devices
Solution Approach 1:
The patent implements dynamic scheduling of DNN operations based on real-time device resource availability and image content characteristics. The control pipeline monitors resource usage and dynamically adjusts which face regions are assigned to recognition pipelines, prioritizing critical recognitions while deferring or skipping less important ones when resources are constrained. This dynamic adaptation enables the system to maintain high throughput when resources are abundant while preventing resource exhaustion when capacity is limited.
Solution Approach 2:
The patent applies different processing qualities to different face regions based on their importance and characteristics. Not all detected faces require the same level of recognition processing - the system can allocate full DNN operations to critical faces while using simplified processing for less important ones. This local quality differentiation optimizes resource consumption by concentrating computational power where most needed while reducing overhead for secondary recognitions.
3Use of energy by moving object
If a single DNN operation is executed for facial recognition, then device resource consumption is reduced, but the system cannot recognize multiple low-resolution faces efficiently in complex environments
Solution Approach 1:
The patent creates a universal recognition pipeline architecture that can handle multiple face recognition tasks using the same underlying DNN operations. Instead of requiring separate specialized processors for each face, the system uses a single flexible pipeline structure that can be dynamically configured to recognize multiple faces simultaneously. This multi-functionality enables the system to maintain low resource consumption per operation while achieving high multi-face recognition capability through repeated use of the same efficient pipeline.
Data Source
AI summary
A facial recognition method executed by a facial recognition apparatus includes detecting at least one face region in units of image blocks from an input image; determining an expected facial recognition difficulty level of the at least one face region; determining a recognition pipeline to perform facial recognition on each of the at least one face region among a plurality of recognition pipelines based on the expected facial recognition difficulty level; distributing the at least one face region to the determined recognition pipeline; and performing facial recognition from each of the at least one face region by a recognition pipeline to which the at least one face region is distributed.