1x1 CNN LEARNING METHOD AND LEARNING DEVICE FOR CONVOLUTIONAL NEURAL NETWORK USING 1x1 CONVOLUTION FOR IMAGE RECOGNITION TO BE USED FOR HARDWARE OPTIMIZATION AND TESTING METHOD AND TESTING DEVICE USING THE SAME

摘要

The present invention provides a method of learning parameters of a CNN for image recognition used for hardware optimization to satisfy KPIs, (1) causing a first transposing layer or a pooling layer to perform corresponding identical each on a pooled ROI feature map. Concatenating each pixel of the location for each ROI to generate an integrated feature map; (2) (i) cause the second transposing layer to separate the volume-adjusted feature map from the integrated feature map for each pixel, and cause the classification layer to generate object information for each ROI, or (ii) reduce the object loss Including the step of backpropagation, in the present invention, since the same processor performs the convolution operation and the FC operation, the size of the chip can be reduced, and there is an advantage in that there is no need to install an additional line during the semiconductor manufacturing process.