Shearography Object Classification Using CNN and Derivative Peaks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Shearographic imagery often suffers from low signal-to-noise ratio (SNR) and high noise due to speckles, making it difficult to detect and classify buried objects effectively.
Innovation Solution
A method combining a derivative peaks process as a first stage cueing algorithm with a convolutional neural network (CNN) for classifying detected objects in shearograms, allowing for the detection of both high and low amplitude shearograms and accommodating new clutter types without manual fine-tuning, using extracted image clips to distinguish between targets and clutter.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If shearographic imagery is used to detect buried objects, then object detection capability is provided, but signal-to-noise ratio is low and noise from speckles is high
Solution Approach 1:
The patent segments the shearographic image processing into distinct stages: first applying a derivative peaks process to detect potential objects, then using a convolutional neural network to classify detected objects. This segmentation allows each stage to focus on specific aspects (detection vs. classification) and reduces the impact of noise by processing information in controlled steps rather than attempting holistic analysis of the noisy full image.
Solution Approach 2:
The patent introduces an intermediary classification stage between raw image acquisition and final object identification. The convolutional neural network acts as an intermediary that processes the derivative peaks data, learning to distinguish true objects from noise patterns. This intermediary layer filters out speckle noise while preserving genuine object signals through trained pattern recognition.
2Productivity
If traditional image processing is used on shearograms, then processing can be performed, but classification accuracy is poor due to low SNR
Solution Approach 1:
The patent replaces traditional mechanical/image processing methods with a data-driven convolutional neural network approach. Instead of using conventional image processing algorithms that struggle with low SNR, the system uses a trained neural network that has learned optimal features for object classification from training data, achieving superior accuracy despite noisy input conditions.
Solution Approach 2:
The patent changes the processing parameters by transitioning from fixed threshold-based derivative analysis to adaptive neural network-based classification. The neural network dynamically adjusts its decision boundaries based on learned patterns from training data, allowing accurate classification even when input SNR varies, whereas traditional methods use fixed parameters that cannot adapt to changing noise conditions.
3Measurement precision
If manual fine-tuning is used for classification, then detection can be adapted to specific targets, but new clutter types cannot be accommodated without re-tuning
Solution Approach 1:
The patent creates a universal classification system using a convolutional neural network that can handle multiple target types and clutter categories through a single trained model. The neural network is trained on diverse data including various targets and clutter types, enabling it to generalize to new clutter types without requiring manual re-tuning, whereas traditional methods would need separate optimization for each specific target-clutter combination.
Solution Approach 2:
The patent implements self-service through automated machine learning training where the system learns optimal classification parameters from training data without human intervention. The neural network automatically adapts to different target and clutter types during training, eliminating the need for manual fine-tuning when encountering new clutter types, as the system self-adjusts its classification boundaries based on learned patterns.
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 enhances the ability to classify objects in shearograms with improved SNR, enabling the detection of weaker or partially imaged targets and reducing noise interference, allowing for effective classification of buried objects as threats or non-threats.
Implementation Method 1
Speckles are statistical interference patterns which occur after reflection of a coherent wave off a rough surface, giving the image a grainy structure.
Implementation Method 2
Speckles are statistical interference patterns which occur after reflection of a coherent wave off a rough surface
Implementation Method 3
The two images are coherently superposed. The lateral displacement is called the shear of the images. The superposition of the two images is called a shearogram, which is an interferogram of an object wave with the sheared surface wave as a reference wave.
Implementation Method 4
The absolute difference of two shearograms recorded at different physical loading conditions of the target surface, part or area is an interference fringe pattern which is directly correlated to the difference in the deformation state of the target area between taking the two images thereof.
Data Source
AI summary
A system for classifying objects detected in shearographic images identifying via a derivative peaks method, process, or algorithm is accomplished with a CNN (CNN). The system utilized training techniques of the CNN to classify objects as threats or non-threats. The system utilizes extracted clips from the shearographic image that contain the presence of an object so that the extracted clip containing the object is/are evaluated by the CNN to classify the object on or below a surface that has been insonified or otherwise stimulated to generate the shearographic image.