Neural Network Spectral Image Characterization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for characterizing samples using infrared thermography face limitations due to sensitivity to external reflections, emissivity variations, and nonuniform energy deposition, which affect the accuracy of thermal signature analysis, particularly in non-destructive testing and coating thickness estimation.
Innovation Solution
A method utilizing a hybrid neural network architecture, comprising a pre-trained convolutional neural network and a classifier like SVM, that processes spectral images to extract features and classify samples, even under nonuniform excitation conditions, with the network trained on both natural and virtual images to enhance robustness and reduce computational requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional infrared thermography methods are used to characterize samples, then thermal images can be acquired and processed, but the results are affected by external reflections, emissivity variations, and nonuniform energy deposition
Solution Approach 1:
The neural network is trained in advance on a large dataset of spectral images with known characteristics. This preliminary training enables the network to learn and compensate for various harmful factors such as external reflections, emissivity variations, and nonuniform energy deposition patterns, so that when actual sample characterization is performed, these factors have already been accounted for in the network's processing logic
Solution Approach 2:
A deep neural network is introduced as an intermediary between the raw spectral image data and the final characterization results. The network acts as a mediator that processes the thermal images, automatically filtering out harmful factors and extracting meaningful features, thereby improving measurement precision without requiring manual correction of reflections or emissivity variations
2Measurement precision
If deep learning algorithms with multiple layers are used to process spectral images, then feature extraction and classification accuracy improve, but computational power requirements and data processing time increase
Solution Approach 1:
The deep neural network is pre-trained offline on a comprehensive dataset, performing the computationally intensive learning process in advance. Once trained, the network's weights and biases are fixed, allowing the actual sample processing to be performed much faster with significantly reduced computational power requirements, as only forward propagation through the already-trained network is needed
Solution Approach 2:
The patent employs a deep neural network with multiple layers that may be more computationally intensive than strictly necessary for some applications. However, this excessive computational capacity during training ensures that the network learns robust features and can handle a wide variety of conditions, providing high measurement precision across diverse scenarios while allowing flexibility to truncate or simplify the network for applications with stricter computational constraints
3Ease of manufacture
If uniform energy deposition is used during thermal excitation, then thermal signature analysis becomes simpler, but high-power flashes or complex heating systems are required to achieve uniformity
Solution Approach 1:
The neural network serves as an intermediary that compensates for nonuniform energy deposition. Instead of requiring complex hardware to achieve uniform heating, the network learns the patterns of nonuniformity during training and automatically corrects for them during processing, thereby maintaining manufacturing simplicity while achieving the effect of uniform energy deposition through software-based correction
Solution Approach 2:
Rather than attempting to eliminate nonuniform energy deposition through complex heating systems, the patent accepts the nonuniformity as an inherent characteristic and uses it as training data for the neural network. The network learns to recognize and compensate for these nonuniform patterns, converting what would be a harmful factor into a learnable feature that improves the robustness of the characterization system
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
This approach enables accurate and efficient characterization of samples with improved noise resistance and reduced computational demands, allowing for real-time processing on low-cost on-board systems, effectively handling variations in energy deposition and measurement conditions.
Implementation Method 1
Specific detectors allow this radiation to be sensed at certain wavelengths and these wavelengths to be converted into luminance values related to the surface temperature of the object, creating thermal images
Implementation Method 2
far-infrared electromagnetic radiation, also called radiative heat, which is continuously emitted by any body having a temperature above absolute zero
Data Source
AI summary
A method for characterizing a sample using spectral images of the sample. At least one volume of values of an observed parameter is generated from the images for a plurality of coordinates of the pixels of the images and a plurality of acquisitions. At least one set of input data from the volume is extracted, with the input data corresponding to the values of the parameter, for a pixel of given coordinates in various acquisitions, to which values at least one conversion function has been applied. The at least one neural network is trained using the input data in order to extract therefrom at least one feature of the sample to be characterized. The at least one feature extracted by the neural network is used to perform a classification of the input data into a plurality of classes, each class being representative of at least one feature of the sample.