Training data generation device, inspection device and program
A training data generation device includes: cut-out unit that receives an input of an image of an inspection target, specifies an inspection area of the inspection target in the image by a predetermined method, and cuts out the specified inspection area from the image; sorting unit that, on the basis of a sorting operation of sorting, as a learning-target image, an image of the inspection area cut out in the cut-out unit into at least either normal or non-normal, associates the learning-target image and a result of the sorting with each other; and training data memory unit that stores training data in which the learning-target image and the result of the sorting are associated with each other.
1. An inspection device, comprising:
a conveyor belt for feeding an inspection target to cause the inspection target to pass through an inspection region;
an electromagnetic-wave radiation unit for applying, as electromagnetic waves, X-rays to the inspection region;
an electromagnetic-wave detection sensor arranged at a position facing the electromagnetic-wave radiator across the conveyor belt and configured to detect the electromagnetic waves that have transmitted through the inspection target passing through the inspection region; and
a processor configured to
configure, as a two-dimensional image, an intensity distribution of the electromagnetic waves detected by the electromagnetic-wave detection sensor,
specify an inspection area of the inspection target in the two-dimensional image by a predetermined method,
cut out the specified inspection area from the two-dimensional image,
extract a sub-image from the cut-out two-dimensional image of the specified inspection area, the sub-image being smaller than the cut-out two-dimensional image of the specified inspection area, and
perform an inspection for determining whether the inspection target is normal based on a determination result that is output by inputting, as a determination-target image, the sub-image into a learned model, wherein
the processor is configured to extract, from the cut-out two-dimensional image of the specified inspection area, a plurality of sub-images including the sub-image such that neighboring sub-images among the plurality of sub-images partially overlap each other.
2. The inspection device according to claim 1 , wherein
the predetermined method for specifying the inspection area is a method for specification based on
positional information of one or more input reference points, and
information indicating a distance from the one or more input reference points.
3. The inspection device according to claim 1 , wherein
the predetermined method for specifying the inspection area is a method for specification based on a blob detected by binarizing the two-dimensional image based on a predetermined threshold of brightness/darkness of the two-dimensional image.
4. The inspection device according to claim 1 , wherein
the determination result output by the learned model is a score that indicates normality of the input determination-target image, and
the processor is configured to perform the inspection for determining whether the inspection target is normal based on the score.
5. The inspection device according to claim 4 , wherein
a plurality of determination-target images including the determination-target image is provided, wherein
the processor is configured to perform the inspection for determining whether the inspection target is normal based on the score output by the learned model in response to each of the plurality of determination-target images input into the learned model.