Method and apparatus for learning CNN-based object detector using 1×1 convolution used for hardware optimization, test method and apparatus using the same {LEARNING METHOD AND LEARNING DEVICE FOR OBJECT DETECTOR BASED ON CNN USING 1x1 CONVOLUTION TO BE USED FOR HARDWARD OPTIMIZATION, AND TESTING METHOD AND TESTING DEVICE USING THE SAME}

摘要

PROBLEM TO BE SOLVED: To provide a learning method, a learning device, a test method and a test device for a CNN-based object detector capable of reducing the amount of calculation. A learning method includes a step of generating an initial feature map using a convolution layer 121 and an integrated feature map using a first transpose layer 124, and a first 1x1 convolution layer 125 and a second 1x1 convolution layer. The detection layer 129 is generated based on the step of generating the second adjustment feature map whose volume is adjusted by the volume layer 126 and the object class information generated by the pixel-by-pixel feature map obtained by separating the volume-adjusted feature map for each pixel. By calculating the object detection loss by referring to the object detection information generated by the above and the original correct answer, at least a part of the convolution layer 121, the first 1x1 convolution layer 125, and the second 1x1 convolution layer 126 Learning parameters. [Selection diagram] Figure 2