Image based novelty detection of material samples
Systems and methods for detecting/identifying novel material samples are provided. A test sample image is processed with a trained transformation function to obtain a transformed matrix. A measure of similarity of the test image based on the transformed matrix is compared to a threshold to determine whether the test sample is novel to a batch of material samples that are provided to train the transformation function.
1. A computer-implemented method of anomalous product identification, the method comprising:
providing, by a processor, a test image of a test sample as an input to a trained transformation function that is obtained by training a transformation function with training images of a batch of material samples;
executing, by the processor, the trained transformation function using the test image as the input, to obtain a transformed matrix;
comparing, by the processor, the transformed matrix to a reference matrix to compute a measure of similarity associated with the test image;
computing, by the processor, a threshold value from reconstruction errors of validation images by processing the validation images with the trained transformation function;
comparing the measure of similarity to the threshold value; and
based on the comparison of the measure of similarity to the threshold value, determining, by the processor, that the test sample is novel to the batch of material samples.
2. The method of claim 1 , wherein the trained transformation function includes a one class support vector machine (SVM) algorithm.
3. The method of claim 2 , further comprising:
extracting, by the processor, computer-vision features from the test image to obtain a feature vector; and
providing, by the processor, the feature vector to the one class SVM algorithm.
4. The method of claim 2 , wherein the measure of similarity is a scalar number.
5. The method of claim 1 , wherein the trained transformation function includes a projection-based reconstruction algorithm.
6. The method of claim 5 , wherein the projection-based reconstruction algorithm includes a principal component analysis (PCA) technique.
7. The method of claim 5 , further comprising extracting, by the processor, computer-vision features from the test image to obtain a feature vector, wherein executing the trained transformation function comprises encoding and decoding the feature vector to obtain the transformed matrix as a reconstructed vector having a same dimension as the feature vector.
8. The method of claim 7 , wherein the measure of similarity represents a reconstruction error between the transformed vector and the feature vector.
9. The method of claim 1 , wherein the trained transformation function includes an autoencoder-based reconstruction algorithm that includes one or more neural networks.
10. The method of claim 1 , wherein the batch of material samples contain substantially same material configurations and/or are produced by substantially same processes.
11. The method of claim 1 , further comprising identifying the test sample to be counterfeit based on the determination that test sample is novel to the batch of material samples.
12. The method of claim 1 , wherein the validation images include one or more of the training images.
13. The method of claim 1 , wherein the validation images and the training images are different and are obtained from material samples of a same batch of material samples.
14. The method of claim 1 , further comprising identifying the test sample as a new sample to be identified based on the determination that test sample is novel to the batch of material samples.
15. A system for anomalous product identification, the system comprising:
a memory configured to store a test image of a test sample; and
a processor in communication with the memory, the processor being configured to:
provide test image stored to the memory as an input to a trained transformation function that is obtained by training a transformation function with training images of a batch of material samples;
execute the trained transformation function using the test image as the input, to obtain a transformed matrix;
compare the transformed matrix to a reference matrix to compute a measure of similarity associated with the test image;
compute a threshold value from reconstruction errors of validation images by processing the validation images with the trained transformation function;
compare the measure of similarity to the threshold value; and
based on the comparison of the measure of similarity to the threshold value, determine whether the test sample is novel to the batch of material samples.
16. The system of claim 15 , wherein the processor is further configured to identify the test sample to be counterfeit based on the determination that test sample is novel to the batch of material samples.
17. The system of claim 15 , wherein the processor is further configured to identify the test sample as a new sample to be identified based on the determination that test sample is novel to the batch of material samples.
18. The system of claim 15 , wherein the validation images include one or more of the training images.
19. The system of claim 15 , wherein the validation images and the training images are different and are obtained from material samples of a same batch of material samples.