METHOD FOR AUTOMATICALLY EVALUATING LABELING RELIABILITY OF TRAINING IMAGES FOR USE IN DEEP LEARNING NETWORK TO ANALYZE IMAGES AND RELIABILITY-EVALUATING DEVICE USING THE SAME

摘要

The present invention relates to a method for evaluating the labeling reliability of a training image for use in learning of a deep learning network, wherein the reliability evaluation device causes the similar image selection network to have a similar shooting environment to the original image, which is an unlabeled image, selecting a verification image candidate group having a unique true label, and then causing an auto-labeling network to auto-label the original image candidate group and the verification image candidate group having the unique true label; (i) evaluating the reliability of the auto-labeling network with reference to the true label and auto-label of the easy validation image, and (ii) manual labeling with reference to the true label and manual label of the difficult validation image Evaluating the reliability of the labeling device; a method comprising the is provided. The above method may be utilized to recognize the surrounding by applying a bag-of-words (BoW) model, optimize a sampling process for selecting a valid image among similar images, and reduce annotation cost.