Fast Dense Patch Search and Quantization for Image Processing

Solution Overview

Problem

The process of searching for nearest neighbor patches in digital images is time-consuming and processor-intensive, especially when comparing hundreds of thousands of image patches to millions of external database patches, even with approximate nearest neighbor search techniques.

Innovation Solution

The implementation of fast dense patch search and quantization, which groups image patches into partitions, determines a partition center patch for each, and forms patch groups with similar patches, using single-nearest neighbor (1-NN) or k-nearest neighbor (k-NN) distance determinations to allocate nearest neighbors, thereby accelerating the search process.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If exhaustive nearest neighbor search is performed for every image patch, then matching accuracy is improved, but processing time and computational intensity increase significantly

Engineering Contradiction:

Improvematching accuracyVSAvoidprocessing time

Core Design Contradiction:

Measurement precisionVSLoss of time

Solution Approach 1:

The patent segments the database patches into multiple clusters, where each cluster is represented by a cluster center. Instead of searching all database patches for every image patch, the system only searches within the relevant cluster, dramatically reducing the search space while maintaining matching accuracy through localized exhaustive search within clusters.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent performs preliminary clustering of database patches before the actual nearest neighbor search. By pre-organizing database patches into clusters based on their content similarity, the system prepares the data structure in advance, enabling efficient localized searches without requiring exhaustive comparison across the entire database for each query.

Inventive Principle:

Principle #10Preliminary action

2Measurement precision

If exhaustive nearest neighbor search is performed for every image patch, then matching accuracy is improved, but processor intensity increases significantly

Engineering Contradiction:

Improvematching accuracyVSAvoidprocessor intensity

Core Design Contradiction:

Measurement precisionVSPower

Solution Approach 1:

The patent segments the database patches into multiple clusters, where each cluster is represented by a cluster center. Instead of searching all database patches for every image patch, the system only searches within the relevant cluster, dramatically reducing the search space while maintaining matching accuracy through localized exhaustive search within clusters.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent performs preliminary clustering of database patches before the actual nearest neighbor search. By pre-organizing database patches into clusters based on their content similarity, the system prepares the data structure in advance, enabling efficient localized searches without requiring exhaustive comparison across the entire database for each query.

Inventive Principle:

Principle #10Preliminary action

3Productivity

If approximate nearest neighbor search is used, then processor efficiency is improved, but matching accuracy deteriorates

Engineering Contradiction:

Improveprocessor efficiencyVSAvoidmatching accuracy

Core Design Contradiction:

ProductivityVSMeasurement precision

Solution Approach 1:

The patent segments the database patches into multiple clusters, where each cluster is represented by a cluster center. Instead of searching all database patches for every image patch, the system only searches within the relevant cluster, dramatically reducing the search space while maintaining matching accuracy through localized exhaustive search within clusters.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent performs preliminary clustering of database patches before the actual nearest neighbor search. By pre-organizing database patches into clusters based on their content similarity, the system prepares the data structure in advance, enabling efficient localized searches without requiring exhaustive comparison across the entire database for each query.

Inventive Principle:

Principle #10Preliminary action

Data Source

PatentUS9286540B2Fast dense patch search and quantization

Publication Date: 2016.03.15 ADOBE INC

AI summary

In techniques for fast dense patch search and quantization, partition center patches are determined for partitions of example image patches. Patch groups of an image each include similar image patches and a reference image patch that represents a respective patch group. A partition center patch of the partitions is determined as a nearest neighbor to the reference image patch of a patch group. The partition center patch can be determined based on a single-nearest neighbor (1-NN) distance determination, and the determined partition center patch is allocated as the nearest neighbor to the similar image patches in the patch group. Alternatively, a group of nearby partition center patches are determined as the nearest neighbors to the reference image patch based on a k-nearest neighbor (k-NN) distance determination, and the nearest neighbor to each of the similar image patches in the patch group is determined from the nearby partition center patches.