Using image analysis algorithms for providing training data to neural networks

1. A method for training a neural network that receives a plurality of input digital images and for each specific input digital image outputs data for determining a relevance level for groups of pixels in the specific input digital image, the method comprising:

determining a training data set by:

for a first plurality of digital training images, determining a relevance level for groups of pixels in each specific digital training image of the first plurality of digital training images by using a first image analysis algorithm configured to automatically label the group of pixels with a relevance level, wherein the first image analysis algorithm is configured to automatically label the group of pixels with a relevance level by:

calculating a spatial statistical measure for the group of pixels;

calculating a temporal statistical measure for the group of pixels by applying a temporal filtering on a sequence of images of the first plurality of digital training images, the sequence of images including the specific digital training image;

calculating a weighted statistical measure by weighting the temporal statistical measure and the spatial statistical measure for the group of pixels; and

labeling the group of pixels with a relevance level based on the weighted statistical measure; and

including the first plurality of digital training images and the labels in the training data set,

wherein each label relates to a relevance level for a group of pixels in an image among the first plurality of digital training images and each label is used as ground truth for the group of pixels when training the neural network; and

using the training data set as an input for training of the neural network.

2. The method of claim 1 , wherein determining the training data set comprises:

for a second plurality of digital training images, determining a relevance level for groups of pixels in each specific digital training image of the second plurality of digital training images by using a second image analysis algorithm configured to automatically label the group of pixels with a relevance level, wherein the second image analysis algorithm differs from the first image analysis algorithm.

3. The method of claim 2 , wherein the second plurality of digital training images comprises at least one digital training image being part of the first plurality of digital training images.

4. The method of claim 2 , wherein the second plurality of digital training images comprises at least one digital training image not being part of the first plurality of digital training images.

5. The method of claim 1 , wherein the training data set is further determined by manually label groups of pixels in at least one digital training image among the first plurality of digital training images with a relevance level.

6. The method of claim 2 , wherein the training data set is further determined by manually label groups of pixels in at least one digital training image among the second plurality of digital training images with a relevance level.

7. A non-transitory computer-readable storage medium including instructions, which when executed by a device having a processing capability, cause the device to execute instructions for training a neural network that receives a plurality of input digital images and for each specific input digital image outputs data for determining a relevance level for groups of pixels in the specific input digital image, the instructions causing the device to perform operations comprising:

determining a training data set by:

for a first plurality of digital training images, determining a relevance level for groups of pixels in each specific digital training image of the first plurality of digital training images by using a first image analysis algorithm configured to automatically label the group of pixels with a relevance level, wherein the first image analysis algorithm is configured to automatically label the group of pixels with a relevance level by:

calculating a spatial statistical measure for the group of pixels;

calculating a temporal statistical measure for the group of pixels by applying a temporal filtering on a sequence of images of the first plurality of digital training images, the sequence of images including the specific digital training image;

calculating a weighted statistical measure by weighting the temporal statistical measure and the spatial statistical measure for the group of pixels; and

labeling the group of pixels with a relevance level based on the weighted statistical measure; and

including the first plurality of digital training images and the labels in the training data set,

wherein each label relates to a relevance level for a group of pixels in an image among the first plurality of digital training images and each label is used as ground truth for the group of pixels when training the neural network; and

using the training data set as an input for training of the neural network.

8. A computer comprising a processor arranged for training a neural network that receives a plurality of input digital images and for each specific input digital image outputs data for determining a relevance level of groups of pixels in the specific input digital image, wherein the training of the neural network comprises:

determining a training data set by:

for a first plurality of digital training images, determining a relevance level for groups of pixels in each specific digital training image of the first plurality of digital training images by using a first image analysis algorithm configured to automatically label the group of pixels with a relevance level, wherein the first image analysis algorithm is configured to automatically label the group of pixels with a relevance level by:

calculating a spatial statistical measure for the group of pixels;

calculating a temporal statistical measure for the group of pixels by applying a temporal filtering on a sequence of images of the first plurality of digital training images, the sequence of images including the specific digital training image;

calculating a weighted statistical measure by weighting the temporal statistical measure and the spatial statistical measure for the group of pixels; and

labeling the group of pixels with a relevance level based on the weighted statistical measure; and

including the first plurality of digital training images and the labels in the training data set,

wherein each label relates to a relevance level for a group of pixels in an image among the first plurality of digital training images and each label is used as ground truth for the group of pixels when training the neural network; and

using the training data set as an input for training of the neural network.