Learning device, learning system, learning method, and program

Abstract

To perform learning so as to allow for robust detection on the shape distribution of a detection target.SOLUTION: A server 20 (a learning device) includes: a distribution acquisition unit (a rectangular shape distribution unit 32) for acquiring the shape distribution of teacher data 43 including annotation data related to predetermined shape information (rectangular information 56); and a selection unit 33 for selecting a parameter candidate used for learning from teacher data of the predetermined shape information on the basis of the shape distribution. The selection unit selects a parameter candidate used for learning from the teacher data so that the size of the predetermined shape related to the annotation data is impartial.SELECTED DRAWING: Figure 2

Claims

1. A distribution acquisition unit acquires the shape distribution of multiple training data sets, which include annotation data associated with images of good products, images of defective products, and predetermined shape information surrounding the defective parts of the products to be inspected. A selection unit selects parameter candidates that define predetermined shape information to be used for generating a trained model from the training data including the annotation data, based on the shape distribution. An inference unit that infers the target to be detected using a trained model generated using the parameter candidates selected by the selection unit, The system comprises a network unit having a convolutional network that performs convolution processing on an input image to be inspected, and the network unit being configured such that the layer outputs from the convolutional network are each branched into multiple layers with different resolutions, The selection unit selects, corresponding to the output of each layer, parameter candidates to be used to generate the trained model from multiple training data sets, including the annotation data, in a manner that avoids duplication. The network unit generates the trained model using the parameter candidates selected by the selection unit. Learning device.

2. The selection unit selects candidate parameters to be used to generate a trained model from the training data so that the size of the shapes represented by predetermined shape information associated with the annotation data is not biased. The learning device according to claim 1.

3. The system includes a generation unit that generates a list of parameter candidates selected by the selection unit so that the size of the shape represented by the predetermined shape information is not biased. The learning device according to claim 2.

4. The generation unit uses the size of the input image showing the object to be inspected to generate a list of candidate parameters to be used in generating the trained model from the training data, such that the size of the shape represented by the predetermined shape information is not biased. The learning device according to claim 3.

5. The selection unit selects shapes such that the shapes represented by predetermined shape information associated with the annotation data are not biased. The learning device according to claim 1.

6. The system includes a generation unit that generates a list of candidate parameters to be used in generating a trained model from the training data, so as to ensure that the external shape represented by predetermined shape information associated with the annotation data is not biased. The learning device according to claim 1.

7. The generation unit uses the size of the input image showing the object to be inspected to generate a list of candidate parameters to be used in generating the trained model from the training data, so as to ensure that the external shape represented by the predetermined shape information is not biased. The learning device according to claim 6.

8. The distribution acquisition unit acquires the shape distribution of a plurality of training data used for inspection of the appearance image of the object to be inspected. The learning device according to claim 1.

9. The predetermined shape information is rectangular information. The learning device according to claim 1.

10. A learning device according to any one of claims 1 to 9, The system includes an appearance inspection system that inspects the appearance image of the object to be inspected, The distribution acquisition unit of the learning device acquires the shape distribution of a plurality of training data used for inspection of the appearance image of the object to be inspected, which is performed by the appearance inspection system. Learning system.

11. The distribution acquisition unit performs a distribution acquisition step in which it acquires a predetermined shape distribution of multiple training data sets, which include annotation data associated with a good product image, a defective product image, and a predetermined shape information surrounding the defective part of the product being inspected, for a plurality of training data sets. The network unit performs convolutional processing on the input image to be inspected, and the layer output from the convolutional network is branched into multiple outputs, each with a different resolution. The selection unit includes a selection step in which it selects, based on the predetermined shape distribution, parameter candidates that define predetermined shape information to be used for generating a trained model from the training data including the annotation data, corresponding to each layer output from the convolutional network, without duplication. The network unit performs a generation step of generating the trained model using the parameter candidates selected by the selection unit, The inference unit includes an inference step of inferring the target to be detected using the generated trained model, Learning methods.

12. A procedure for obtaining a predetermined shape distribution of a plurality of training data sets, which include an image of a good product to be inspected, an image of a defective product to be inspected, and annotation data associated with predetermined shape information surrounding the defective portion of the product to be inspected, A procedure for branching and outputting multiple layer outputs from a convolutional network that performs convolutional processing on the input image to be inspected, each with a different resolution, and A procedure for selecting, based on the predetermined shape distribution, parameter candidates that define predetermined shape information to be used in generating a trained model from the training data including the annotation data, corresponding to each layer output from the convolutional network, without duplication; A step to generate the trained model using the parameter candidates selected in the selection step, The procedure for inferring the target to be detected using the generated pre-trained model, A program that causes a computer to execute something.

Images