preview

What Is CNN Architecture : U-NET

Decent Essays
Open Document

CNN Architecture: U-NET
Built upon the ‘fully convolutional network’ paradigm, it supplements the contracting network with successive layers of upsampling operators instead of pooling operators; this provides the network with the power to localize the learning. The architecture consists of contracting and expanding paths as seen in fig. The contraction path reduces the original image through a series of convolutions, ReLU and maxpooling to extract relevant features. This is followed by an expansion path which through a sequence of upconvolutions, ReLU and concatenation with high-resolution features from contraction path reproduces original image as a segmented output. Thus the expansive path is almost symmetric to contracting path yielding …show more content…

The final layer used 1 x 1 convolution to map each 64-component feature vector to the required number of classes. The network was quite heavy with a total of 23 convolution layers and necessitates the selection of an appropriate input tile size so that the 2 x 2 max pooling led to pixel sizes that were even numbers. It may be noted that the unpadded convolutions led to output images that were smaller than the input by a fixed width.
Ever since the original paper was published, it was highly popular for semantic segmentation. The absence of U-net related work in earth observation and the success of
U-net for similar semantic segmentation tasks motivated its choice for this project.
III. REINFORCEMENT LEARNING
Reinforcement learning which has its origins in behavioural psychology is a computational approach to learn from interaction with an environment. It adopts the approach of learning about stimulus or environment through reward and punishment. Unlike supervised or unsupervised learning, data isn’t independently identically distributed but dynamic and sequential. It is a kind of semi-supervised learning that lies midway between supervised and unsupervised learning. Adaptive learning through observed data and predicting future outcome are at the soul of reinforcement learning.
While supervised learning uses class labels for training data and unsupervised learning uses

Get Access
Related
Get Access