Segmentation Model Training Using Occlusion-Free Background Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image segmentation models struggle with accurately differentiating between background and foreground, particularly in uneven or irregular areas, due to insufficient training data and reliance on ground truth accuracy.
Innovation Solution
A computer-implemented training method for a segmentation model that utilizes first image data representing either background or foreground, with annotation data defining the presence of these elements, allowing for error calculation and model updating without requiring target segmentation data, and incorporating non-visible electromagnetic radiation imaging to enhance training accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional segmentation models are used with standard training data, then the model can be trained with available resources, but the model incorrectly classifies background features as foreground features
Solution Approach 1:
The patent applies preliminary action by capturing images at specific wavelengths (e.g., infrared) before the segmentation process to identify occlusion-free regions. These pre-identified regions are then used as reliable training data to train the segmentation model, ensuring that the model learns from accurate background representations without foreground interference.
Solution Approach 2:
The patent introduces an intermediary approach by using multi-wavelength imaging to capture images at different wavelengths (e.g., visible and infrared). These multi-wavelength images serve as intermediaries to identify occlusion-free background regions, which are then used to train the segmentation model for more accurate foreground-background differentiation.
2Measurement precision
If detailed target segmentation data is used for training, then the model can achieve high accuracy, but the complexity and resource requirements increase significantly
Solution Approach 1:
The patent applies the taking out principle by extracting only the essential training information from occlusion-free background regions identified in multi-wavelength images. Instead of using complete detailed segmentation data, the method extracts specific reliable background regions to create simplified training datasets that maintain accuracy while reducing complexity.
Solution Approach 2:
The patent uses a disposable approach by generating training data from easily captured multi-wavelength images rather than requiring complex, expensive, and time-consuming manual annotation processes. The multi-wavelength imaging provides a quick and efficient way to obtain reliable training data without significant resource investment.
3Adaptability or versatility
If the segmentation model processes images with occlusions, then it can handle real-world scenarios, but the classification accuracy decreases due to foreground-background confusion
Solution Approach 1:
The patent applies preliminary action by pre-identifying occlusion-free background regions using multi-wavelength imaging before the segmentation process. These pre-identified regions are used to train the model to distinguish background from foreground, enabling the model to maintain high classification accuracy when processing images with occlusions in real-world scenarios.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
The method improves the segmentation model's ability to accurately separate background and foreground, reducing the need for detailed segmentation and enhancing the precision of image processing, especially in environments with occlusions.
Implementation Method 1
The mask signal is generated by capturing an image comprising non-visible electromagnetic radiation emitted by emitters disposed in the billboard
Data Source
AI summary
Aspects of the present invention relate to a computer-implemented training method for training a segmentation model to segment an image. The method includes receiving a plurality of first training data sets. The first training data sets each include first image data representing a first image consisting of a background or of a foreground and first annotation data identifying the presence solely of the background or the foreground in the first image. The first image data is captured by at least one visible electromagnetic radiation imaging device. The method includes, for each first training data set, processing the first image data using the segmentation model to generate a first candidate segmentation; and supplying the first annotation data to an error calculating algorithm to determine a first error for the first candidate segmentation. The segmentation model is updated in dependence on the determined first error. According to a further aspect of the present invention there is provided a system for training a segmentation model to segment an image. Aspects of the present invention also relate to an image processing system and method.