Method and device for acquiring sample images for inspecting label among auto-labeled images

Abstract

The invention provides a method for acquiring a sample image for label-inspecting among auto-labeled images for learning a deep learning network, optimizing sampling processes for manual labeling, andreducing annotation costs. The method is characterized by comprising steps that a sample image acquiring device generates first and a second images, enables convolutional layers to generate first andsecond feature maps, enables a pooling layer to generate first and second pooled feature maps, and generates concatenated feature maps; enables a deep learning classifier to acquire the concatenatedfeature maps and generate class information; and calculates probabilities of abnormal class elements in an abnormal class group, determines whether the auto-labeled image is a difficult image, and selects the auto-labeled image as the sample image for label-inspecting. Further, the method can be performed by using a robust algorithm with multiple transform pairs. By the method, hazardous situations are detected more accurately.

Description

technical field [0001] The present invention relates to a method and a device for obtaining at least one sample image for checking labels in at least one automatically labeled image for deep learning network learning. Background technique [0002] Recently, research has been conducted on methods of recognizing objects using machine learning, and the like. Through this machine learning process, deep learning using a neural network having a plurality of hidden layers between the input layer and the output layer has high recognition performance. [0003] Also, the above-mentioned neural network using the above-mentioned deep learning is generally learned by backpropagation using a loss. [0004] In order to learn such a deep learning network, it is necessary for a labeler to add a label (tag) to individual data points, that is, label training data. Preparing such training data (ie, accurately classifying data), especially where large amounts of training data are utilized and ...

AI Technical Summary

Problems solved by technology

[0006] However, in this conventional method, the throughput of inspectors is lower than that of the above-mentioned automatic labeling device, and it takes a long time to generate real labels for all training images

In order to improve the overall processing rate, it is necessary to increase the number of inspectors, but in this case, the problem of cost increase occurs

[0007] Moreover, there is also the problem that it is difficult to ensure a plurality of skilled inspectors who can withstand the processing rate of the above-mentioned automatic labeling device.

Images