Image Masking With Distance-Based Windows for Contour-Preserving Anonymization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for anonymizing faces and motor vehicle license plate numbers in images for automatic driving systems are inefficient, leading to information loss and distortion, as they often result in unnecessary image regions being made unrecognizable and introduce new edges or artifacts, which can mislead learning-based object detection methods.
Innovation Solution
A method that determines pixel values for an image, generates a mask with predefined object classes, calculates distance to the nearest non-object region pixel, assigns a window size based on this distance, and replaces original pixel values with new values calculated from neighboring pixels within the window, minimizing edge creation and distortion.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If automatic object detection with artificial neural networks is used to detect faces and license plates, then detection speed and automation are improved, but the detected rectangular regions do not correspond to actual object contours, resulting in excessive image regions being made unrecognizable
Solution Approach 1:
The patent applies local quality by using distance-dependent window sizes for pixel value determination. Pixels closer to object boundaries use smaller windows to preserve edge details, while pixels farther from boundaries use larger windows for better anonymization. This localized adaptation of processing parameters resolves the contradiction by making the anonymization process spatially variable rather than uniform.
2Reliability
If uniform blurring or pixelation is applied to detected regions, then anonymization is achieved, but contours and edges are toned down and new edges or artifacts are introduced
Solution Approach 1:
The patent changes the parameter of window size dynamically based on distance to non-object regions. Instead of using a fixed window size for all pixels, the window size varies continuously with distance, creating a smooth gradient transition that avoids abrupt edge effects and artifacts while maintaining effective anonymization.
3Measurement precision
If manual detection of faces and license plates is performed, then precise object contours can be identified, but the process is too laborious for large numbers of images
Solution Approach 1:
The patent applies preliminary action by first using automatic neural network detection to identify candidate regions, then refining the masking boundaries based on distance calculations. This two-stage approach combines the speed of automatic detection with the precision of contour-aware processing, achieving both goals without requiring fully manual detection.
Data Source
AI summary
A method for masking objects contained in an image is provided, wherein pixel values are determined for pixels of an image. An image mask is generated, which comprises at least one first region in which pixels of a predefined object class are arranged and at least one second region in which pixels of said predefined object class are not arranged. The distance to the nearest pixel in the second region is determined for pixels in the at least one first region. Pixels in the at least one first region are assigned a window that has a window size that depends on the determined distance. At least one new pixel value is determined for a pixel from the pixel values of the pixels within the window assigned to the pixel. The original pixel values of pixels in the one first region are replaced with the respective new pixel values.