Method for training a convolutional neural network
A method for training a convolutional neural network. For each of a multiplicity of training input images, the method includes processing of the training input image by the convolutional network; processing of a scaled version of the training input image by the convolutional network; determining a pair of convolutional layers of the convolutional network so that a convolutional layer of the pair generates a first feature map for the training input image which has the same size as a second feature map which is generated by the other convolutional layer of the pair for the scaled version of the training input image; and calculating a loss between the first feature map and the second feature map; and training the convolutional neural network to reduce an overall loss which includes the calculated losses.
1 . A method for training a convolutional neural network, the method comprising:
for each of a multitude of training input images:
processing the training input image by the convolutional network,
processing a scaled version of the training input image by the convolutional network,
determining a pair of convolutional layers of the convolutional network so that a convolutional layer of the pair generates a first feature map for the training input image which has the same size as a second feature map which is generated by the other convolutional layer of the pair for the scaled version of the training input image, and
calculating a loss between the first feature map and the second feature map; and
training the convolutional neural network to reduce an overall loss which includes the calculated losses.
2 . The method as recited in claim 1 , wherein the scaled version of the training input image is generated by scaling the training input image using a downsampling factor between successive convolutional layers of the convolutional network, or a power of the downsampling factor, or a reciprocal value of the downsampling factor, or a power of the reciprocal value of the downsampling factor.
3 . The method as recited in claim 1 , wherein the overall loss furthermore has a training loss for training the convolutional network for a predefined task, and the method further comprises weighting the calculated losses in the overall loss with regard to the training loss.
4 . The method as recited in claim 1 , further comprising:
determining multiple pairs of convolutional layers of the convolutional network for each of the multiplicity of training input images so that, for each pair of the multiple pairs, a convolutional layer of the pair generates a first feature map for the training input image, which has the same size as a second feature map which is generated by the other convolutional layer of the pair for a respective scaled version of the training input image, and a loss between the first feature map and the second feature map is calculated, and the overall loss includes the losses calculated for the pairs.
5 . The method as recited in claim 1 , wherein the neural network is trained using a training dataset of training input images, the multiplicity of training input images is selected from the training dataset, and scaled versions of the training input images are generated for the training input images so that the convolutional network has a pair of convolutional layers for each training input image of the multiplicity of training input images and each scaled version of the training input image, so that a convolutional layer of the pair generates a feature map for the training input image which has the same size as a feature map which is generated by the other convolutional layer of the pair for the scaled version of the training input image.
6 . A training device for a convolutional neural network, the training device configured to:
for each of a multitude of training input images:
process the training input image by the convolutional network,
process a scaled version of the training input image by the convolutional network,
determine a pair of convolutional layers of the convolutional network so that a convolutional layer of the pair generates a first feature map for the training input image which has the same size as a second feature map which is generated by the other convolutional layer of the pair for the scaled version of the training input image, and
calculate a loss between the first feature map and the second feature map; and
train the convolutional neural network to reduce an overall loss which includes the calculated losses.
7 . A non-transitory computer-readable medium on which is stored commands which, when executed by a processor, cause the processor to perform the following steps:
for each of a multitude of training input images:
processing the training input image by a convolutional network,
processing a scaled version of the training input image by the convolutional network,
determining a pair of convolutional layers of the convolutional network so that a convolutional layer of the pair generates a first feature map for the training input image which has the same size as a second feature map which is generated by the other convolutional layer of the pair for the scaled version of the training input image, and
calculating a loss between the first feature map and the second feature map; and
training the convolutional neural network to reduce an overall loss which includes the calculated losses.