Marker Position Display Map for Contiguous Object Identification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional image processing methods, particularly using convolutional neural networks (CNNs), face challenges in accurately identifying multiple contiguous target objects of the same type, such as vertebrae in medical images, often resulting in erroneous identification where adjacent vertebrae are misclassified as a single entity.
Innovation Solution
An image processing apparatus and method that generates a marker position display map indicating the positions of target objects within an analysis image, which is then input to a semantic segmentation model to enhance the accuracy of identifying contiguous target objects by using labels to differentiate between adjacent objects, thereby improving the model's output image identification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a conventional CNN is used to identify target objects in an analysis target image, then the identification process is simple, but the accuracy of identifying multiple contiguous target objects of the same type deteriorates
Solution Approach 1:
The patent divides the identification task into multiple processing stages: initial identification by CNN, extraction of identified position information, generation of marker position display maps, and secondary identification. This segmentation allows the system to handle multiple contiguous target objects by processing them in discrete steps rather than attempting to identify all objects simultaneously, thereby improving accuracy without overwhelming system complexity.
Solution Approach 2:
The patent performs preliminary identification of target objects using a CNN before generating marker position display maps. This preliminary action provides position information that guides subsequent processing steps, allowing the system to prepare necessary data structures and markers in advance, which improves the accuracy of identifying multiple contiguous objects while organizing complexity into manageable preliminary and subsequent phases.
2Measurement precision
If markers are used to indicate positions of target objects and a marker position display map is generated, then the accuracy of identifying contiguous target objects is improved, but the processing complexity increases
Solution Approach 1:
The patent introduces a marker position display map as an intermediary data structure between the CNN identification process and the final output. This intermediary contains position information and markers that facilitate the identification of multiple contiguous target objects by providing structured guidance data, thereby improving accuracy while containing complexity within a dedicated intermediate representation rather than dispersing it throughout the entire system.
3Measurement precision
If a semantic segmentation model is used with marker position display maps to identify target objects, then the identification accuracy of multiple contiguous objects is improved, but the processing time increases
Solution Approach 1:
The patent segments the processing into parallelizable stages: CNN-based initial identification, marker position display map generation, and semantic segmentation model processing. This segmentation enables certain operations to be performed independently and potentially in parallel, reducing the overall processing time while maintaining the accuracy benefits of the multi-stage approach with marker position display maps.
Data Source
AI summary
An image processing apparatus comprising: a processor and a memory connected to or incorporated in the processor, in which the processor acquires an analysis target image in which a plurality of contiguous target objects of the same type appear, receives an input of a marker indicating positions of the target objects in the analysis target image, generates a marker position display map indicating a position of the marker in the analysis target image, inputs the analysis target image and the marker position display map to a semantic segmentation model, and outputs, from the semantic segmentation model, an output image in which the target objects are identified.