Method and System for Feature Detection in Imaging Applications

Simple SummaryContent extracted from patent full text and abstract with AI.

This invention describes a method and system for detecting specific features (such as particles, nanoparticles, or other objects) in digital images using machine learning, particularly deep neural networks. A key innovation is the generation of large quantities of synthetic, automatically labelled training images from digital templates, eliminating the need for manual image annotation. This synthetic data can be used to train high-performing feature detection algorithms for various imaging modalities.

Use CasesContent extracted from patent full text and abstract with AI.

• Automated detection and characterization of particles or nanoparticles in scientific microscopy images (SEM, HIM, etc.)
• Biomedical image analysis, such as identifying and segmenting cells or other biological features
• Quality control in manufacturing, for example detecting defects in materials or products through imaging
• Environmental monitoring, such as identifying pollutants or contamination in samples using imaging
• Forensic or security applications, like automatic recognition of specific shapes, symbols, or license plates in surveillance footage
• Research in materials science for high-throughput analysis of sample properties through automated feature extraction
• Automated statistical analysis of large image datasets in science or industry

BenefitsContent extracted from patent full text and abstract with AI.

• Removes the need for time-consuming and error-prone manual annotation of training data
• Enables the generation of extensive, diverse, and customizable synthetic training datasets tailored to specific detection tasks
• Improves the accuracy and robustness of machine learning models for feature detection, even when real labelled data is scarce or unavailable
• Reduces costs and resource requirements associated with data preparation for machine learning applications in imaging
• Suitable for a wide range of imaging modalities and feature types, making it highly flexible and broadly applicable
• Supports high-throughput, automated analysis enabling faster research, diagnostics, and quality control
• Robust to image artefacts and variations since synthetic training data can include realistic noise and variability

Technical Classifications (CPCs)

Inventors & Applicants

Patent Abstract

The invention proposes a method and system for using a machine learning algorithm in the detection of at least one predetermined feature in digital imaging applications. In accordance with the proposed method, large quantities of reliable training data may be generated automatically. This allows to improve the detection performance of deep convolutional neural networks which are trained using the so-generated data, without requiring the actual imaging of manual annotation of large training datasets.