Instance Segmentation Feedback for Densely Packed Objects
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing segmentation networks, such as Mask Region-based Convolutional Neural Network (R-CNN), struggle to accurately predict segmentation maps for densely packed tiny objects affected by low-image quality, non-uniform illumination, blurring, and low-resolution, often missing instances due to varying textures and configurations, and require pre and post-processing steps for robust performance.
Innovation Solution
A method and system that iteratively enhances input images, uses a trained mask R-CNN model to generate initial segmentation maps, predicts missing instances based on contour maps, and applies the model iteratively until a predefined threshold is met, refining the segmentation maps to include all instances.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing segmentation networks (Mask R-CNN) are used for densely packed objects, then the model structure is simple and fast, but the segmentation accuracy deteriorates due to missed instances and poor performance on low-image-quality inputs
Solution Approach 1:
The patent applies preliminary action by performing image enhancement before segmentation to improve input image quality. It also performs preliminary contour detection and missing instance prediction before final segmentation, ensuring all instances are captured in subsequent iterations.
Solution Approach 2:
The patent implements feedback through iterative processing where the segmentation results from one iteration are fed back to generate updated contour maps, which are then used to predict missing instances in the next iteration. This continuous feedback loop refines segmentation accuracy until convergence.
2Reliability
If iterative prediction of missing instances is performed, then the completeness of instance detection improves, but the processing time increases
Solution Approach 1:
The iterative prediction process uses feedback from contour map comparisons to identify and predict missing instances. Each iteration refines the segmentation based on previous results, systematically improving completeness while controlling processing time through threshold-based termination.
Solution Approach 2:
The patent replaces manual instance counting and verification with automated computational methods. The system automatically compares contour maps, identifies missing instances, and predicts their locations using trained models, eliminating time-consuming manual processes.
3Reliability
If pre and post processing steps are added to handle low image quality, then the robustness of segmentation improves, but the overall processing complexity increases
Solution Approach 1:
The patent applies preliminary action by enhancing image quality before the main segmentation process. Image enhancement operations are performed in advance to improve illumination uniformity and reduce noise, ensuring the segmentation model receives optimized input and reducing the need for complex post-processing corrections.
Data Source
AI summary
A system and method for instance segmentation of densely packed objects is disclosed. The method includes receiving an input image, enhancing the input image to generate an enhanced input image, inputting the enhanced input image to a trained mask R-CNN model to generate a first segmentation map, iteratively predicting one or more missing instances of densely packed objects from the input image based on comparison of a contour map generated for the input image with a contour map generated for the first segmentation map, comparing an area of the one or more missing instances with a predefined threshold value, applying the trained masked R-CNN model on each of the one or more missing instances to generate a second segmentation map, and obtaining a refined segmentation map based on the first segmentation map and the second segmentation map.