Deep group disentangled embedding and network weight generation for visual inspection
A method is provided for visual inspection. The method includes learning, by a processor, group disentangled visual feature embedding vectors of input images. The input images include defective objects and defect-free objects. The method further includes generating, by the processor using a weight generation network, classification weights from visual features and semantic descriptions. Both the visual features and the semantic descriptions are for predicting defective and defect-free labels. The method also includes calculating, by the processor, a cosine similarity score between the classification weights and the group disentangled visual feature embedding vectors. The method additionally includes episodically training, by the processor, the weight generation network on the input images to update parameters of the weight generation network. The method further includes generating, by the processor using the trained weight generation network, a prediction of a test image as including any of defective objects and defect-free objects.
1. A computer-implemented method for visual inspection, comprising:
learning, by a hardware processor, group disentangled visual feature embedding vectors of input images, the input images including defective objects and defect-free objects;
generating, by the hardware processor using a weight generation network, classification weights from visual features and semantic descriptions, both the visual features and the semantic descriptions for predicting defective and defect-free labels;
calculating, by the hardware processor, a cosine similarity score between the classification weights and the group disentangled visual feature embedding vectors;
episodically training, by the hardware processor, the weight generation network on the input images to update parameters of the weight generation network; and
generating, by the hardware processor using the trained weight generation network, a prediction of a test image as including any of defective objects and defect-free objects.
2. The computer-implemented method of claim 1 , wherein the classification weights are generated given limited exemplar class information below a threshold amount.
3. The computer-implemented method of claim 2 , wherein the limited exemplar class information comprises a number of semantic attributes below the threshold amount.
4. The computer-implemented method of claim 2 , wherein the limited exemplar class information comprises a number of labeled examples below the threshold amount.
5. The computer-implemented method of claim 1 , wherein said training step comprising training the classification weight generator using a cosine similarity based softmax function.
6. The computer-implemented method of claim 1 , wherein the method is performed relative to a training set of images and a training set of exemplar class information corresponding to the training set of images.
7. The computer-implemented method of claim 1 , wherein the group disentangled visual feature embedding vectors are learned using a feature embedding network implemented by a deep neural network configured to perform multiclass classification, when labeled data is available.
8. The computer-implemented method of claim 1 , wherein the group disentangled visual feature embedding vectors are learned using a feature embedding network implemented by a deep autoencoder configured to perform input image reconstruction, when labeled data is unavailable.
9. The computer-implemented method of claim 1 , wherein the weight generation network maps the visual features and the semantic descriptions to the classification weights.
10. The computer-implemented method of claim 1 , wherein said learning step comprises:
splitting the group disentangled visual feature embedding vectors into k groups that are independent with each other; and
constructing new feature embedding vectors based on the k groups using a self-attention function.
11. The computer-implemented method of claim 1 , wherein said learning step comprises:
training a discriminator using the group disentangled visual feature embeddings vectors generated by a mini-batch of the input images as positive data;
shuffling the group disentangled visual feature embeddings vectors of the mini-batch in a group-wise manner to obtain negative data; and
training the discriminator with the negative data.
12. A computer program product for visual inspection, the computer program product comprising a non-transitory computer readable storage medium having program instructions embodied therewith, the program instructions executable by a computer to cause the computer to perform a method comprising:
learning, by a hardware processor, group disentangled visual feature embedding vectors of input images, the input images including defective objects and defect-free objects;
generating, by the hardware processor using a weight generation network, classification weights from visual features and semantic descriptions, both the visual features and the semantic descriptions for predicting defective and defect-free labels;
calculating, by the hardware processor, a cosine similarity score between the classification weights and the group disentangled visual feature embedding vectors;
episodically training, by the hardware processor, the weight generation network on the input images to update parameters of the weight generation network; and
generating, by the hardware processor using the trained weight generation network, a prediction of a test image as including any of defective objects and defect-free objects.
13. The computer program product of claim 12 , wherein said training step comprising training the classification weight generator using a cosine similarity based softmax function.
14. The computer program product of claim 12 , wherein the method is performed relative to a training set of images and a training set of exemplar class information corresponding to the training set of images.
15. The computer program product of claim 12 , wherein the group disentangled visual feature embedding vectors are learned using a feature embedding network implemented by a deep neural network configured to perform multiclass classification, when labeled data is available.
16. The computer program product of claim 12 , wherein the group disentangled visual feature embedding vectors are learned using a feature embedding network implemented by a deep autoencoder configured to perform input image reconstruction, when labeled data is unavailable.
17. The computer program product of claim 12 , wherein the weight generation network maps the visual features and the semantic descriptions to the classification weights.
18. The computer program product of claim 12 , wherein said learning step comprises:
splitting the group disentangled visual feature embedding vectors into k groups that are independent with each other; and
constructing new feature embedding vectors based on the k groups using a self-attention function.
19. The computer program product of claim 12 , wherein said learning step comprises:
training a discriminator using the group disentangled visual feature embeddings vectors generated by a mini-batch of the input images as positive data;
shuffling the group disentangled visual feature embeddings vectors of the mini-batch in a group-wise manner to obtain negative data; and
training the discriminator with the negative data.
20. A computer processing system for visual inspection, comprising:
a memory device including program code stored thereon;
a hardware processor, operatively coupled to the memory device, and configured to run the program code stored on the memory device to
learn group disentangled visual feature embedding vectors of input images, the input images including defective objects and defect-free objects;
generate, using a weight generation network, classification weights from visual features and semantic descriptions, both the visual features and the semantic descriptions for predicting defective and defect-free labels;
calculate a cosine similarity score between the classification weights and the group disentangled visual feature embedding vectors;
episodically train the weight generation network on the input images to update parameters of the weight generation network; and
generate, using the trained weight generation network, a prediction of a test image as including any of defective objects and defect-free objects.