1xK Kx1 CNN LEARNING METHOD AND LEARNING DEVICE FOR CNN USING 1xK OR Kx1 CONVOLUTION TO BE USED FOR HARDWARE OPTIMIZATION AND TESTING METHOD AND TESTING DEVICE USING THE SAME

摘要

The present invention provides a method for learning a CNN parameter using a 1xK convolution operation or a Kx1 convolution operation provided to be used for hardware optimization that meets a KPI (Key Performance Index, key performance indicator), a learning apparatus comprising: (a ) Reshaping the feature map (Reshaped) by allowing the reshaping layer to two-dimensionally concatenate the features in each group consisting of K channels of the training image or the feature map processed from it Feature Map) and causing a subsequent convolutional layer to apply a 1xK convolution operation or a Kx1 convolution operation to the reshaped feature map to generate an Adjusted Feature Map; and (b) causing the output layer to refer to the adjustment feature map or features on the processed feature map, and causing the loss layer to refer to the output from the output layer and at least one GT (Ground Truth) corresponding thereto to determine the loss. It is characterized in that it comprises;