Saliency Map Compression for Machine Learning Feature Preservation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing media compression methods fail to consider the importance of specific elements in media files, leading to uniform compression that hinders user and machine perception and limits the effectiveness of systems like facial recognition and machine learning models.
Innovation Solution
Constructing a saliency map to identify and prioritize elements based on their importance, allowing for targeted compression and modification of media files to preserve relevant features while reducing file size.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If uniform compression is applied to all elements of media files, then file size is reduced, but perception quality and machine learning effectiveness deteriorate
Solution Approach 1:
The patent applies different compression quality levels to different regions of the media file based on a saliency map. Important regions (high saliency) are compressed with higher quality to preserve perception and machine learning effectiveness, while less important regions (low saliency) are compressed with lower quality to maximize file size reduction. This resolves the contradiction by making compression quality spatially variable rather than uniform.
Solution Approach 2:
The patent segments the media file into multiple regions with different compression characteristics based on the saliency map. Each region is independently compressed according to its importance level, allowing the system to optimize the trade-off between file size and quality for each segment separately, thereby achieving both size reduction and quality preservation where needed.
2Quantity of substance
If uniform compression is applied to all elements of media files, then file size is reduced, but machine learning model performance deteriorates
Solution Approach 1:
The patent identifies regions critical for machine learning tasks (such as facial features in images) using the saliency map and applies higher compression quality specifically to these regions. This ensures that the features most important for machine learning model performance are preserved with high fidelity, while other regions can be compressed more aggressively, thus maintaining model performance while reducing file size.
Solution Approach 2:
The patent segments the media file into regions of interest for machine learning and applies differential compression. By identifying and protecting critical feature regions through segmentation based on saliency, the system ensures that machine learning models receive sufficient information from important areas while achieving overall file size reduction.
3Quantity of substance
If compression is applied to reduce file size, then storage requirements are reduced, but quality is compromised
Solution Approach 1:
The patent applies local quality variation through the saliency map, where important regions maintain high quality (low compression) and less important regions accept lower quality (high compression). This resolves the contradiction by ensuring that quality is preserved where it matters most while achieving significant storage reduction in less critical areas, optimizing the overall quality-size trade-off.
Data Source
AI summary
A method may include obtaining a machine learning model configured to label media files of a given type. The method may include obtaining a saliency map including multiple elements corresponding to aspects of the given type of media files, each element corresponding to encoding parameters. A given encoding parameter of an associated element may be based on an identified salience of the associated element in the machine learning model determining a label. The method may include obtaining a first media sample of the given type. The method may include deriving a second media sample from the first media sample based on applying the given encoding parameter to the associated element of the first media sample according to the saliency map. The method may include labeling, by the machine learning model, the second media sample.