Method, device and computer program for training an artificial neural network for object recognition in images
The invention relates to a method for training an artificial neural network (ANN) for object detection in images. The method includes a training phase of training the ANN with a training database (TDB) that includes training images, and measuring a performance of the trained ANN with a validation database that includes validation images. If the measured performance is not satisfactory, the method further includes an enhancing phase of detecting a mismatch between the validation images and a response of the trained ANN in response to the validation images, correlating the mismatch to a characteristic of the training images that impacts the training of the ANN, modifying the training database according to the characteristic, and training the ANN with the modified training database. The invention further relates to a non-transitory computer program and a device configured to carry out the method, and to an artificial neural network trained with the method.
1 . A method for training an artificial neural network (ANN) for object detection in images, said method configured to be implemented by a computer, and said method comprising:
a training phase that comprises
training the ANN with a training database (TDB) comprising training images, and
measuring a performance of the ANN that is trained with a validation database (VDB) comprising validation images;
wherein if the performance that is measured is not satisfactory, implementing an enhancing phase comprising
detecting a mismatch between the validation images and a response of the ANN that is trained in response to said validation images,
correlating the mismatch that is detected to at least one characteristic of the training images that impacts the training of the ANN,
modifying the training database (TDB) according to said at least one characteristic of the training images that impacts the training of the ANN, and
training the ANN with said training database that is modified;
wherein said correlating the mismatch comprises correlating said mismatch to at least one metadata, associated to said training images, and describing an environment, or an imaging condition, of an object to be detected in said training images.
2 . The method according to claim 1 , wherein, for a validation image of the validation images, the mismatch is detected by comparing a label of said validation image to a feature detected in said validation image by a predetermined algorithm.
3 . The method according to claim 2 , wherein the predetermined algorithm is a heatmap algorithm, or a gradient based algorithm.
4 . The method according to claim 1 , wherein said modifying the training database (TDB) comprises adding at least one new training image to said training database.
5 . The method according to claim 4 , further comprising generating the at least one new training image by 3D-simulation.
6 . The method according to claim 4 , further comprising automatically adding, to the at least one new training image, a label indicating a location of an object to be detected in said at least one new training image.
7 . The method according to claim 4 , further comprising automatically adding, to the at least one new training image, a tag indicating a type of an object to be detected in said at least one new training image.
8 . The method according to claim 1 , wherein said modifying the training database comprises selecting a subset of said training database (TDB).
9 . The method according to claim 1 , wherein the training database comprises
at least one training image, comprising a synthetic training image, obtained by simulation; and
optionally, at least one real image captured by imaging means.
10 . A non-transitory computer program comprising instructions which, when executed by a computer, cause the computer to carry out a method for training an artificial neural network (ANN) for object detection in images, said method configured to be implemented by a computer, and said method comprising:
a training phase that comprises
training the ANN with a training database (TDB) comprising training images, and
measuring a performance of the ANN that is trained with a validation database (VDB) comprising validation images;
wherein if the performance that is measured is not satisfactory, implementing an enhancing phase comprising
detecting a mismatch between the validation images and a response of the ANN that is trained in response to said validation images,
correlating the mismatch that is detected to at least one characteristic of the training images that impacts the training of the ANN,
modifying the training database (TDB) according to said at least one characteristic of the training images that impacts the training of the ANN, and
training the ANN with said training database that is modified;
wherein said correlating the mismatch comprises correlating said mismatch to at least one metadata, associated to said training images, and describing an environment, or an imaging condition, of an object to be detected in said training images.
11 . A device comprising:
one or more of
at least one software component, and
at least one hardware component
configured to implement a method for training an artificial neural network (ANN) for object
detection in images, said method comprising
a training phase that comprises
training the ANN with a training database (TDB) comprising training images, and
measuring a performance of the ANN that is trained with a validation database (VDB) comprising validation images;
wherein if the performance that is measured is not satisfactory, implementing an enhancing phase comprising
detecting a mismatch between the validation images and a response of the ANN that is trained in response to said validation images,
correlating the mismatch that is detected to at least one characteristic of the training images that impacts the training of the ANN,
modifying the training database (TDB) according to said at least one characteristic of the training images that impacts the training of the ANN, and
training the ANN with said training database that is modified;
wherein said correlating the mismatch comprises correlating said mismatch to at least one metadata, associated to said training images, and describing an environment, or an imaging condition, of an object to be detected in said training images.