Long-Distance Face Detection with Threshold-Guided Head Identification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing CNN-based face detection systems require significant computational resources, which are not available in lower-end devices, leading to the use of lower resolutions that complicate face detection, especially at longer distances.
Innovation Solution
Implement a two-step face detection process using lower resolution images with multiple confidence thresholds. Results above a high threshold are accepted, below a low threshold are discarded, and intermediate results undergo additional head identification using higher resolution images if necessary.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If lower resolution images are used for face detection, then computational resources are reduced, but face detection accuracy deteriorates especially at longer distances
Solution Approach 1:
The detection process is segmented into two stages: first, a lightweight face detector processes low-resolution images to identify candidate regions; second, a more accurate head identification model processes only these candidates at higher resolution. This segmentation allows the system to use computational resources efficiently while maintaining accuracy for critical detections.
Solution Approach 2:
The system applies different processing qualities to different regions of interest. Low-resolution processing is applied to the entire image for initial detection, while high-resolution processing is applied locally only to candidate regions that require further verification. This local quality enhancement maintains accuracy where needed without the global computational cost.
2Measurement precision
If higher resolution images are used for face detection, then face detection accuracy improves, but computational resource requirements increase significantly
Solution Approach 1:
The system performs preliminary filtering using a lightweight face detector on low-resolution images before applying the computationally intensive head identification model. This preliminary action eliminates most non-face regions, so the high-resolution processing is only applied to a small subset of candidate regions, significantly reducing overall computational requirements.
Solution Approach 2:
Instead of applying high-resolution processing to the entire image, the system applies it partially only to candidate regions identified by the initial detector. This partial action provides sufficient accuracy for detection while avoiding the excessive computational cost of processing the full high-resolution image.
3Use of energy by moving object
If simpler CNN architectures are used, then computational requirements are reduced, but detection confidence and reliability decrease
Solution Approach 1:
The system merges two different detection approaches: a simple, fast face detector and a more reliable head identification model. The simple detector provides speed and low computational requirements, while the head identification model provides reliability and accuracy. By combining these two models in a cascade architecture, the system achieves both efficiency and reliability.
Solution Approach 2:
The simple face detector acts as an intermediary that filters the input image to produce candidate regions. This intermediary step allows the more reliable head identification model to focus only on promising candidates, thereby achieving high reliability without requiring the intermediary model itself to be complex or computationally intensive.
Data Source
AI summary
Performing face detection with a lower resolution image and evaluating results against high and low thresholds. Results above the high threshold are accepted and results below the low threshold are discarded. Results between the two thresholds receive additional processing. Instead of face detection, head identification is used on the candidates. If the head size is above a limit, head identification is performed. If the head size is below the limit, a higher resolution version of the area is utilized for head identification. The results of the head identification are evaluated against a threshold and for results above the threshold, the original face detection results are accepted and below the threshold the original face detection results are discarded.