U-Net is undoubtedly a medical image segmentation on the basis of a network structure model is very important. Many models are modified based on the magic to U-Net.
The first feature of the U-Net is completely symmetrical, the left side of the encoding, decoding the right. Understanding of the main U-Net to make a comparison with the FCN. If you are not familiar with the venue to FCN here.
The decoder FCN relatively simple, only a deconvolution operation, after convolution structure not kept pace. U-Net using a deconvolution operation on the sampling and the operation performed crop, wherein FIG crop and do comparison with the original. The second difference is skip connection, FCN with the add operation (summation), U-Net using a stack operation (concatenation).
Add operation: FIG consistent feature dimensions, numerical values of the features in FIG make additions.
Stack Operation: FIG consistent feature size, the dimensions of the overlay (3 + 3 * 3 * 3 * 3 * 3 * 3 * 64 = 67)
