Cnn-based learning method, learning device for selecting useful training data and test method, test device using the same

Abstract

A convolutional neural network (CNN)-based learning method for selecting useful training data is provided. The CNN-based learning method includes steps of: a learning device (a) enables a first CNN module (i) to generate a first feature map, and enables a second CNN module to generate a second feature map; (ii) generates a first output indicating identification information or location informationof an object by using the first feature map, and calculates a first loss by referring to the first output and its corresponding GT; (b) enables the second CNN module (i) to change a size of the firstfeature map and integrate the first feature map with the second feature map, to generate a third feature map; (ii) generates a fourth feature map and to calculate a second loss; and (c) backpropagatesthe auto-screener loss generated by referring to the first loss and the second loss.

Description

technical field [0001] The present invention relates to a CNN-based learning method and learning device for selecting useful learning data, as well as a testing method and testing device using the same, more specifically, for selecting useful learning data In the learning method based on the CNN selected, the following steps are included: (a) when the learning device acquires at least one input image, perform the following processing: (i) make the first CNN module perform at least one convolution on the input image Operation, generate a first feature map, the first CNN module is used to obtain the identification information or position information of the specific object in the input image; and (ii) make the second CNN module perform convolution on the input image at least once product operation to generate a second feature map, and the second CNN module can automatically screen the useful learning data used in the learning process of the first CNN module; (b) the learning devi...

AI Technical Summary

Problems solved by technology

[0011] However, there is a problem that the number of training images contained in the training image database is not proportional to the performance of the learning device (e.g., the above-mentioned detector)

In contrast, if the number of training images in the learning database is increased by 10 times, the cost of building the database is very high, because it is necessary to manually create GT images for each of the 90,000 images. However, the useful data for improving the performance of the detector is only 4,500 more

And, if using the useful 4,500 images, the performance of the detector is further improved to 98% through the learning process, the cost of building the database required to improve the performance of the detector is even higher

Images