METHOD FOR GENERATING FEATURE MAP FOR A CONVOLUTIONAL NEURAL NETWORK ROBUST TO SCALE VARIATION AND COMPUTING DEVICE USING THE SAME

摘要

The present invention provides a method for generating a feature map for a convolutional neural network robust to scale variations, wherein (a) when a first feature map corresponding to an image or a convolution operation of the image is obtained at least once, the computing device , by inputting the first feature map as a first (1*1) convolutional layer to cause the first (1*1) convolutional layer to perform a (1*1) convolution operation on the first feature map to obtain the number of channels. generate an adjusted second feature map, and define the second feature map as (1*r) - where r is an integer greater than or equal to 1 - with an extension ratio (k*k) - where k is an integer greater than or equal to 2 - including a kernel 1st (k*k) convolutional layer to (n*r) where n is an integer greater than or equal to 2 - nth (k*k) convolutional layer including a (k*k) kernel having an extension ratio, respectively By inputting, each of the first (k*k) convolutional layer to the nth (k*k) convolutional layer divides the channels of the second feature map into at least two groups, and divides the channels of the second feature map into at least two groups. 3_1 feature map by performing a (k*k) convolution operation using an (1*r) expansion ratio or a (k*k) convolution operation using an (n*r) expansion ratio on each of the second feature maps corresponding to to generate a to 3_n feature map; and (b) the computing device inputs each of the 3_1 feature map to the 3_n feature map as a 2_1 (1*1) convolution layer to a 2_n (1*1) convolution layer, and the second_1 ( 1*1) the convolutional layer to the 2nd_n (1*1) convolutional layer, each of the 3_1 to the 3_n feature map is (1*1) convolutional operation, and the number of channels is adjusted The third (1*1) by generating 4_1 feature maps to 4_n feature maps, concatenating the 4_1 feature maps to the 4_n feature maps, and inputting them as a third (1*1) convolutional layer. A convolution layer performs a (1*1) convolution operation on the concatenated 4_1 to 4_n feature maps to generate a fifth feature map with an adjusted number of channels, the first feature map and the generating a sixth feature map by concatenating the fifth feature map; It relates to a method comprising