Cropped Image Object Detection for Fast and Accurate Video Masking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection methods in video streams suffer from reduced accuracy due to the need for neural networks to operate at lower resolutions, leading to potential visibility of masked objects through digital zooming and loss of data.
Innovation Solution
A method involving image cropping and recursive analysis using multiple neural networks to enhance object detection accuracy, where images are scaled to a lower resolution for initial analysis and further cropped for refined detection, with indications provided when probabilities meet certain thresholds.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If the neural network operates at lower resolution to ensure fast processing, then processing speed is improved, but object detection accuracy deteriorates
Solution Approach 1:
The image processing is divided into two stages: first, the full image is downsampled to low resolution for fast initial detection; second, only suspicious regions are cropped and processed at high resolution for accurate classification. This segmentation of processing resolution by region allows speed and accuracy to coexist.
Solution Approach 2:
Different parts of the image are processed at different resolutions. The majority of the image is processed at low resolution for speed, while only specific regions containing potential objects of interest are processed at high resolution for accuracy. This local quality differentiation resolves the contradiction between overall processing speed and local detection accuracy.
2Productivity
If the image is downscaled to lower resolution for neural network operation, then processing efficiency is improved, but data loss occurs
Solution Approach 1:
The method extracts only the necessary portions of the high-resolution image (suspicious regions) for detailed analysis. Instead of processing the entire high-resolution image which would be computationally expensive, only relevant regions are cropped and analyzed at high resolution, thus avoiding unnecessary data loss while maintaining processing efficiency.
Solution Approach 2:
A preliminary low-resolution analysis is performed first to identify suspicious regions before conducting detailed high-resolution analysis. This preliminary action filters out most of the image area that doesn't contain objects of interest, allowing the system to maintain high processing efficiency while preserving necessary detail information only where needed.
3Device complexity
If the neural network operates at reduced resolution, then computational load is reduced, but detection reliability deteriorates
Solution Approach 1:
The processing resolution is made dynamic rather than static. The system starts with low-resolution processing for the entire image, then dynamically increases resolution only for regions that require further investigation. This dynamic adaptation of resolution based on detected needs reduces computational load while maintaining detection reliability for critical regions.
Data Source
AI summary
A method and control unit for object detection in a video stream captured with an image acquisition device, especially for the purpose of anonymizing objects in the video stream.