Image Prediction Tiling With Overlap for Boundary Accuracy

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

High-capacity and high-cost computational resources are required for processing high-resolution pathology slide images, leading to inaccuracies in prediction results at the boundaries of divided images, which deteriorate the overall prediction performance.

Innovation Solution

The target image is divided into overlapping sub-images, with bilinear weights applied to prediction results for each pixel, and these results are merged to generate a final prediction result, ensuring smooth transitions at boundaries.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Use of energy by moving object

If the target image is divided into multiple sub-images for processing, then computational resource requirements are reduced, but prediction accuracy at boundary regions deteriorates

Engineering Contradiction:

Improvecomputational resource requirementsVSAvoidprediction accuracy at boundaries

Core Design Contradiction:

Use of energy by moving objectVSMeasurement precision

Solution Approach 1:

The target image is divided into multiple sub-images with overlapping regions, allowing parallel processing of smaller image segments while maintaining boundary accuracy through the overlap. Each sub-image is processed independently by the prediction model, reducing computational resource requirements compared to processing the entire high-resolution image at once.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

Different weighting strategies are applied to different regions of the sub-images. Boundary regions and overlapping areas receive special weight treatment (e.g., lower weights or averaging) to prevent discontinuities and maintain prediction accuracy at boundaries, while central regions use standard prediction weights.

Inventive Principle:

Principle #3Local quality

2Productivity

If sub-images are processed independently, then processing efficiency is improved, but continuity of prediction results at split lines deteriorates

Engineering Contradiction:

Improveprocessing efficiencyVSAvoidcontinuity of prediction results

Core Design Contradiction:

ProductivityVSStability of the object's composition

Solution Approach 1:

Overlapping regions are incorporated into the sub-image division strategy before processing begins. This preliminary design ensures that boundary regions are included in multiple sub-images, allowing the independent processing of each sub-image to produce results that can be seamlessly merged through the overlapping areas.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The prediction results from multiple sub-images are merged by combining the results from overlapping regions. This merging process restores continuity at split lines by using the redundant information from overlapping areas to smooth transitions and eliminate discontinuities that would otherwise occur at boundaries.

Inventive Principle:

Principle #5Merging (Combining)

3Measurement precision

If overlapping sub-images are used, then prediction accuracy at boundaries is improved, but computational redundancy increases

Engineering Contradiction:

Improveprediction accuracy at boundariesVSAvoidcomputational redundancy

Core Design Contradiction:

Measurement precisionVSQuantity of substance

Solution Approach 1:

Weighting parameters are applied to different regions of sub-images to optimize the balance between accuracy and redundancy. By adjusting the weights of overlapping and boundary regions, the system achieves improved boundary prediction accuracy while controlling the computational cost through selective emphasis on critical regions rather than uniform processing of all overlapping areas.

Inventive Principle:

Principle #35Parameter changes

Data Source

PatentUS20260023822A1Method and system for performing prediction work on target image

Publication Date: 2026.01.22 LUNIT

AI summary

Provided is a method for performing a prediction work on a target image, including dividing the target image into a plurality of sub-images, generating prediction results for a plurality of pixels included in each of the plurality of divided sub-images, applying weights to the prediction results for the plurality of pixels, and merging the prediction results for the plurality of pixels applied with the weights.