Structural Preprocessing for Minuscule Target Detection in Noisy Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing algorithms struggle to distinguish between minuscule detection targets and noise in captured images, particularly in low-quality images, leading to reduced precision in particle analysis and classification of cell types and tissues.
Innovation Solution
A detection system utilizing a machine learning model, specifically convolutional neural networks, that focuses on the structures around minuscule detection targets to extract candidates and uses preprocessing techniques like Hough transform and Blob Filter to enhance detection accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Difficulty of detecting and measuring
If image processing algorithms are used to detect minuscule detection targets, then detection capability is provided, but the ability to distinguish between detection targets and noise deteriorates
Solution Approach 1:
The patent segments the detection task into multiple stages: first extracting candidate regions based on structural characteristics (using Hough transform and Blob Filter), then classifying these candidates using machine learning models. This segmentation allows the system to handle the complexity of distinguishing minuscule detection targets from noise by breaking down the detection process into manageable steps, each focusing on specific features.
Solution Approach 2:
The patent introduces an intermediary classification model that sits between the image processing stage and the final detection output. This intermediary model receives candidate regions extracted from images and classifies them as either detection targets or noise, serving as a mediator that resolves the ambiguity between similar-looking structures and actual detection targets.
2Productivity
If conventional image processing algorithms are used, then processing is performed, but precision of particle analysis deteriorates due to noise extraction
Solution Approach 1:
The patent implements feedback mechanisms where the classification model continuously learns from the distinction between true detection targets and noise. The system processes images, classifies candidates, and uses this information to refine its understanding, thereby maintaining high processing throughput while improving precision through iterative learning and adaptation.
Solution Approach 2:
The patent changes the parameters and approaches used in image processing by introducing machine learning models that can adapt to different image characteristics. Instead of using fixed thresholds and parameters, the system dynamically adjusts its detection criteria based on learned patterns from training data, allowing it to maintain precision across varying conditions while preserving processing efficiency.
3Measurement precision
If detection is performed at high magnification, then minuscule detection targets can be detected, but image quality deteriorates making distinction from noise difficult
Solution Approach 1:
The patent segments the detection process to first identify candidate regions based on structural features that are preserved even in low-quality images, then applies classification to distinguish true targets from noise. This segmentation allows the system to leverage the structural information that remains reliable even when overall image quality deteriorates at high magnifications.
Solution Approach 2:
The patent performs preliminary extraction of candidate regions based on structural characteristics before final classification. By preparing and pre-processing the images to highlight structural features early in the process, the system can compensate for quality deterioration and still identify potential detection targets for further analysis.
Data Source
Figure 1
Figure 2
Figure 3~4
AI summary
A detection system includes a learning apparatus and a detection apparatus. The learning apparatus has a first preprocessing unit that extracts, from a captured image used for training, which includes a noise image and an image of an object having a detection target, a training image including an image of the detection target, based on a structure around the detection target, and a training unit that trains a machine learning model based on the training image. The detection apparatus has an acquisition unit that acquires the machine learning model, a second preprocessing unit that extracts, from a captured image used for detection, which includes the noise image and the image of the object having the detection target, candidates for being the detection target image in which the detection target is captured, based on the structure around the detection target, and a detection unit that detects, in the captured image used for detection, a detection target image from among the detection target image candidates, by using the machine learning model.