Image Object Masking With Distance-Based Pixel Replacement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for anonymizing personal data in images, such as faces and license plates, distort object contours and introduce artificial edges, which negatively impact learning-based object detection systems, complicating the development and testing of automated driving systems.
Innovation Solution
A method that uses distance-dependent masking to replace original pixel values with averages from surrounding pixels, maintaining object contours and minimizing edge artifacts, ensuring accurate detection while anonymizing sensitive areas.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If automatic object detection with bounding boxes is used to identify faces and license plates, then detection speed and automation are improved, but the rectangular masking areas do not correspond to actual object contours, causing excessive image distortion and loss of information
Solution Approach 1:
The patent applies pixel-wise classification where each pixel is independently labeled as belonging to a face or license plate or not, allowing the masking to precisely follow object contours rather than using uniform rectangular bounding boxes. This local quality approach ensures that only the necessary pixels are masked, minimizing information loss while maintaining automation.
2Reliability
If more image areas are obscured to ensure reliable detection (reduce false negatives), then data protection is improved, but more unnecessary areas are distorted (increase false positives)
Solution Approach 1:
The patent replaces the mechanical bounding box approach with a pixel-level classification system using neural networks. This substitution allows for precise contour-based masking that accurately identifies only the pixels belonging to faces or license plates, achieving high detection reliability without unnecessary distortion of surrounding areas.
3Reliability
If traditional blurring methods are used to anonymize detected areas, then anonymization is achieved, but object contours are distorted and artificial edges are introduced, negatively impacting learning-based detection systems
Solution Approach 1:
The patent extracts and removes the pixels belonging to faces and license plates from the image entirely, replacing them with a uniform color or pattern. This extraction approach avoids the contour distortion and artificial edge introduction caused by traditional blurring methods, preserving the integrity of surrounding object shapes while achieving effective anonymization.
Data Source
Figure 1
Figure 2
Figure 3a~3c
AI summary
The method according to the invention for masking objects contained in an image involves pixel values being determined for the pixels of an image (1). An image mask is generated (2), which has at least one first region, in which pixels of a predetermined object class are arranged, and at least one second region, in which pixels of this predetermined object class are not arranged. For the pixels in the at least one first region, the distance from the particular closest pixel in the at least one second region is determined (3). The pixels in the at least one first region are each assigned a window (4) that has a window size dependent on the determined distance. From the pixel values of the pixels within the window assigned to a particular pixel, at least one new pixel value is determined for the respective pixel (5). The original pixel values of the pixels in the at least one first region are replaced with the respective new pixel values (6).