Samsung patents faster image resizing for neural networks

1. A neural network operation apparatus that generates a target tensor comprising first pixels having respective first pixel values, the apparatus comprising: one or more processors; and memory storing instructions configured to cause the one or more processors to approximate a bilinear upsampling of the target tensor by a scale factor by: generating an upsampled tensor comprising blocks of second pixels having respective second pixel values by performing nearest neighbor (NN) upsampling on the target tensor, the blocks respectively corresponding to the first pixels and having a same dimension, wherein the NN upsampling comprises setting each blocks' second pixel values to the first pixel value of the first pixel corresponding thereto, wherein the dimension of the blocks is proportional to the scale factor; and generating a downsampled tensor comprising third pixels having respective third pixel values by performing a pooling operation of striding a window with a dimension proportional to the scale factor over the upsampled tensor, wherein the strides of the window respectively correspond to the third pixels, wherein for each stride of the window the second pixel values of the second pixels in the window are pooled to generate the third value of the third pixel corresponding to the stride of the window, and wherein the downsampled tensor approximates the bilinear upsampling of the target tensor by the scale factor.