System and method for expanding and training convolutional neural networks for large size input images

A method, comprising:

• creating a trained convolutional neural network (CNN), training of the CNN comprising;
• evaluating training input images using first convolutional layers, the number of said first convolutional layers being based on a first size of input images, each layer of said first convolutional layers receiving layer input signals comprising features of said input images and generating layer output signals,
• responsive to a training input image being a second size, said second size being larger than said first size, adding additional convolutional layers to said CNN, the number of said additional convolutional layers being based on said second size in relation to said first size,
• initializing said additional convolutional layers using weights from said first convolutional layers, and
• creating feature maps comprising said layer output signals;
• selecting a plurality of feature maps of an output of at least one selected layer of said trained CNN according to values attributed to said feature maps by said trained CNN;
• for each feature map of said plurality of feature maps, determining a location corresponding thereto in an image space of a test input image;
• defining interest points of said test input image, based on said locations corresponding to said plurality of feature maps;
• and
• classifying said test input image based on said interest points.

A system, comprising:

• a processing unit;
• a memory operationally connected to said processing unit, said memory comprising feature maps comprising layer output signals from a trained convolutional neural network (CNN);
• and
• an imaging device connected to said processing unit, said imaging device receiving a test image comprising a set of pixels,
• said processing unit selecting a plurality of feature maps from said memory,
• for each of said plurality of feature maps, said processing unit determining a location corresponding thereto in an image space of said test image,
• said processing unit defining interest points of said test image, based on said locations corresponding to said plurality of feature maps, and
• said processing unit classifying said test image based on said interest points.