Image Compression Retaining Feature Information at Original Coordinates
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current deep learning image processing models, such as neural networks and CNNs, require significant memory and computational resources to handle large 3D images or high-resolution 2D videos, leading to inefficient processing and reduced accuracy due to memory limitations in traditional GPUs.
Innovation Solution
The method involves compressing images while retaining feature information at original coordinates, creating subsets of regions of interest, and training tailored models using these subsets to reduce computational and storage requirements, allowing for efficient image classification without the need for large datasets.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If full-resolution 3D images are processed using traditional deep learning models, then image quality and feature information are preserved, but memory consumption and computational resources increase significantly
Solution Approach 1:
The patent divides the full-resolution image into multiple patches or blocks, processes each patch independently through the neural network, and then reconstructs the full image from processed patches. This segmentation approach reduces the memory burden of handling entire high-resolution images while preserving local feature information through the patch-wise processing approach.
Solution Approach 2:
The patent transforms the spatial domain image processing into frequency domain processing using Fourier transforms. By converting images to frequency representations, the system can perform compression and processing operations that reduce memory requirements while maintaining essential feature information through selective frequency component processing.
2Productivity
If traditional GPU memory limitations are imposed, then computational resources are constrained, but processing efficiency and model accuracy decrease
Solution Approach 1:
By segmenting images into patches and processing them independently, the system enables parallel processing across multiple GPU cores or even multiple devices, improving overall processing efficiency. The patch-based approach allows better utilization of limited GPU memory while maintaining model accuracy through consistent processing of all image regions.
Solution Approach 2:
The patent dynamically adjusts processing parameters such as patch size, overlap regions, and frequency threshold values to optimize the balance between processing efficiency and model accuracy. These parameter changes allow the system to adapt to different GPU memory capacities while maintaining reliable processing results.
3Quantity of substance
If image compression is applied, then storage and processing requirements are reduced, but feature information and pattern recognition capability are lost
Solution Approach 1:
The patent applies Fourier transforms to convert images from spatial domain to frequency domain, enabling compression by selectively retaining only the most significant frequency components. This frequency-based compression reduces data size while preserving essential pattern information that is critical for recognition tasks, as frequency representations capture global image structures and textures.
Solution Approach 2:
The system extracts and processes only the most relevant frequency components or salient features from the full image spectrum, discarding redundant information. This extraction approach reduces data size while maintaining the essential pattern recognition capability needed for accurate image analysis.
Data Source
AI summary
In an approach to image compression while retaining feature information at original coordinates, one or more computer processors compress an image while retaining pattern and feature information at original coordinates, wherein the image is a full resolution 3D image. The one or more computer processors create one or more image subsets based on the one or more determined regions of interest identified by utilizing a first pass selection of the regions of interest based on the compressed image. The one or more computer processors train one or more models based the created one or more image subsets that contain one or more regions of interest based on the compressed image.