Multilayer Laplacian Resizer for Vision Systems
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional image resizing methods, such as nearest neighbor, bilinear, and bicubic resizing, are not optimized for machine perception in deep learning-based visual recognition tasks, leading to potential loss of visual features and performance degradation in image analysis, while learned resizers require high computational resources and struggle with transferability to new tasks.
Innovation Solution
A lightweight multilayer Laplacian resizer is jointly trainable with neural network models, utilizing few trainable parameters to boost details in specific frequency sub-bands, with trainable scaling and bias parameters and a nonlinearity function applied to generate a resized image that can be used for various computer vision tasks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional resizing methods (nearest neighbor, bilinear, bicubic) are used, then computational efficiency is maintained, but visual feature preservation and machine perception performance deteriorate
Solution Approach 1:
The patent segments the image into multiple frequency sub-bands using Laplacian pyramids, allowing different processing strategies for different frequency components. This segmentation enables selective enhancement of important visual features while maintaining computational efficiency through hierarchical processing.
Solution Approach 2:
The patent applies local quality enhancement by selectively boosting important frequency sub-bands based on their relevance to visual features. Different regions and frequency components receive tailored processing, with trainable parameters adapting the enhancement strength locally rather than applying uniform resizing across the entire image.
2Reliability
If learned resizers with deep neural networks are used, then image resizing performance for machine perception is improved, but computational overhead and parameter requirements increase significantly
Solution Approach 1:
The patent extracts only the essential frequency components needed for machine perception using Laplacian pyramids, discarding redundant information. This extraction approach retains the critical visual features required for downstream tasks while eliminating unnecessary computational complexity associated with processing all image details.
Solution Approach 2:
The patent changes the parameter representation by using a small set of trainable parameters that modulate frequency sub-band responses, rather than employing large neural network weight matrices. This parameter optimization reduces computational overhead while maintaining the ability to adapt to different machine perception tasks.
3Measurement precision
If learned resizers are used, then resizing accuracy is improved, but transferability to new tasks and datasets deteriorates due to excessive model capability
Solution Approach 1:
The patent achieves universality by designing a frequency-based resizer that can be applied across multiple computer vision tasks (classification, detection, segmentation) and different datasets. The Laplacian pyramid structure and frequency sub-band processing provide a task-agnostic framework that maintains high resizing accuracy while adapting easily to new applications.
Solution Approach 2:
The patent introduces dynamics through trainable parameters that can be adapted to different tasks and datasets. The model transitions from static conventional resizing to a dynamic system where frequency enhancement strengths are learned and adjusted based on specific application requirements, enabling both high accuracy and good transferability.
Data Source
AI summary
The technology provides an image resizer that is jointly trainable with neural network classification (recognition) models, and is designed to improve classification performance. Systems and method include applying an input image to a baseline resizer to obtain a default resized image, and applying the input image to a plurality of filters. Each respective filter in the plurality is configured to perform sub-band filtering on the input image to obtain a sub-band filtered result. This includes applying the sub-band filtered result to the baseline resizer to obtain a respective resized result, and also includes applying to the respective resized result a scaling parameter, a bias parameter, and a nonlinear function to obtain a respective filtered image. The process then combines the default resized image and the respective filtered images to generate a combined resized image.