Adaptive system and method for inspection of imaged items
A visual inspection system and method are provided in which a reference image of an item is processed and based on the processing a pre-trained model is adjusted to obtain an adapted model. The adapted model, rather than the pre-trained model, is then used to process an inspection image of a same-type item, to obtain inspection information which may then be output to a user.
1 . A visual inspection method comprising:
(i) processing at least one reference image of an item;
(ii) based on the processing, adjusting a pre-trained model, without re-training on specific items, to obtain an adapted model by updating parameters of the pre-trained model without an iterative back-propagating process;
(iii) processing an inspection image of a same-type item, using the adapted model, to obtain inspection information; and
(iv) outputting the inspection information.
2 . The method of claim 1 , wherein the processing of the at least one reference image comprises encoding the reference image.
3 . The method of claim 1 , further comprising:
using an adaptation component to process the at least one reference image, the adaptation component being configured to cause a change to at least one parameter of the pre-trained model, and the change comprising one or a combination of: adding, removing, changing a value of or replacing a parameter of the pre-trained model.
4 . The method of claim 3 , wherein the adaptation component causes a change to a latent space of the pre-trained model.
5 . The method of claim 3 , wherein the adaptation component causes a change to an embedding of the pre-trained model.
6 . The method of claim 3 , wherein the adaptation component causes a change to a metric function of the pre-trained model.
7 . The method of claim 3 , wherein the adaptation component causes a change to a network architecture of the pre-trained model.
8 . The method of claim 3 , wherein the adaptation component causes a change to an input inspection image.
9 . The method of claim 3 , wherein the at least one parameter of the pre-trained model comprises a parameter of a classifier used to detect a defect in an image.
10 . The method of claim 3 , wherein the at least one parameter of the pre-trained model comprises a threshold of determining that an inspection image represents a defect.
11 . A system for visual inspection, the system comprising:
an adaptation component which receives a reference image of an item and causes a change to a parameter of a pre-trained model based on the reference image, without re-training on specific items, the change updating parameters of the pre-trained model without an iterative back-propagating process to produce an adapted model, and the adapted model comprising a classification component configured to receive an inspection image of a same-type item; and
an output module which receives input from the classification component and outputs inspection information to a user based on the input from the classification component.
12 . The system of claim 11 , further comprising:
an auto-encoder configured to process the reference image prior to being received at the adaptation component.
13 . The system of claim 11 , wherein the adaptation component causes a change to a latent space of the pre-trained model.
14 . The system of claim 11 , wherein the adaptation component causes a change to an embedding of the pre-trained model.
15 . The system of claim 11 , wherein the adaptation component causes a change to a metric function of the pre-trained model.
16 . The system of claim 11 , wherein the adaptation component causes a change to a network architecture of the pre-trained model.
17 . The system of claim 11 , wherein the adaptation component causes a change to a combination of a latent space of the pre-trained model, an embedding of the pre-trained model, a metric function of the pre-trained model and a network architecture of the pre-trained model.
18 . The system of claim 11 , wherein the adaptation component infers a design of a network architecture of the adapted model.
19 . The system of claim 11 , further comprising:
an input module which includes an image processing component, to process the inspection image prior to being received at the classification component.
20 . The system of claim 11 , wherein the classification component comprises:
an embedding module which embeds the reference image and the inspection image; and
a defect detector module which receives input from the embedding module and which determines existence and location of a defect in the inspection image, from the input.