Abstract: A method for training a convolutional neural network comprises sliding a first, a second and a third window in a predefined path simultaneously on each training image of a training data set, a binary image containing a highlighted region of interest of the each training image, and a labeled image containing a plurality of tags representing one or more features of interest of the each training image. A target matrix is obtained for each sample of the labeled image that lies within the highlighted region of interest. The target matrix is a probabilistic distribution of the plurality of tags that is based on a proportion of existence of a feature of interest represented by a tag in the each sample. The CNN is trained to recognize each sample to contain the proportion of the at least one feature of interest based on the target matrix of the each sample.

Abstract: A method for training a convolutional neural network comprises sliding a first, a second and a third window in a predefined path simultaneously on each training image of a training data set, a binary image containing a highlighted region of interest of the each training image, and a labeled image containing a plurality of tags representing one or more features of interest of the each training image. A target matrix is obtained for each sample of the labeled image that lies within the highlighted region of interest. The target matrix is a probabilistic distribution of the plurality of tags that is based on a proportion of existence of a feature of interest represented by a tag in the each sample. The CNN is trained to recognize each sample to contain the proportion of the at least one feature of interest based on the target matrix of the each sample.