Upsampling and resizing

Similarly, an average-unpooling operation was developed to mirror average-pooling. The latter operation takes a pooling region of k × k elements and averages them into a single value. Therefore, an average-unpooling layer takes each value of a tensor and duplicates it into a k × k region, as illustrated in Figure 6-6:

Figure 6-6: Example of an average-unpooling operation (also known as upsampling)

This operation is nowadays used more often than max-unpooling, and is more commonly known as upsampling. For instance, this operation can be instantiated through tf.keras.layers.UpSampling2D() (refer to the documentation at https://www.tensorflow.org/api_docs/python/tf/keras/layers/UpSampling2D). This method is itself nothing more than a wrapper for tf.image.resize() (refer to the documentation at https://www.tensorflow.org/api_docs/python/tf/image/resize) when called with the method=tf.image.ResizeMethod.NEAREST_NEIGHBOR ...