Fully Convolutional Neural Network
In the previous section, we performed pixel-wise classification using a deep convolutional neural network. We have performed the inference in patch-based mode, meaning that for each output pixel, we have run the model on one small patch of the input image, centered on the output pixel position. While this kind of network architecture is easy to implement, it is not efficient in terms of processing. Due to patch overlap, the data is copied multiple times with different memory alignment (for each different patch) and passed to the model. The costliest operations implemented in the deep net are the convolutions, which are massively parallel and can be implemented so that the intermediary results are reused ...
Become an O’Reilly member and get unlimited access to this title plus top books and audiobooks from O’Reilly and nearly 200 top publishers, thousands of courses curated by job role, 150+ live events each month,
and much more.
Read now
Unlock full access