LEARNING METHOD AND LEARNING DEVICE FOR IMPROVING PERFORMANCE OF CNN BY USING FEATURE UPSAMPLING NETWORKS, AND TESTING METHOD AND TESTING DEVICE USING THE SAME

摘要

In the present invention, a learning method for improving CNN performance using FUN is disclosed. This method includes (a) when the input image is obtained, causing the down-sampling block to apply a predetermined operation to the input image to obtain a down-sampling image, and (b) when the down-sampling image is obtained, the 1_1 To 1_K filter blocks by applying one or more convolution operations to the down-sampling image, respectively, 1_1, 1_2, ... , 1_K feature maps are obtained, respectively, (c) cause a specific up-sampling block to: (i) receive a specific feature map from a corresponding filter block, and (ii) obtain another predetermined feature map from the previous up-sampling block. After receiving and rescaling the specific feature map to have the same size as the predetermined feature map, (iii) performing a certain operation on the specific feature map and the rescaled predetermined specific map to generate a feature map of a predetermined up-sampling block. , (d) (i) obtaining an application-specific output from the application block and (ii) performing a first backpropagation process.