A Befitting Image Data Crawling and Annotating System with CNN based Transfer Learning

With the Growth of Big Data technology, Artificial Intelligence technology, especially Deep Learning, is growing rapidly. In Deep Learning technology that like Convolutional Neural Network (CNN), collecting and training a sizeable befitting training data has a significant impact on technology maturity and completeness. But it is difficult for a person to collect a sizeable training data directly, and there are some problems in collecting it automatically with an ordinary web image crawler. For example, the name of a car is often obtained from a specific place name. When these keywords are crawled, the crawler retrieves only the text metadata of the image and gets the image. As a result, the crawler collects all images of different but has the same text. Also, even if the image of the car has been collected, only some of the car's interior or only a portion of the data that is not suitable for learning is also collected. In this paper, we present a method to classify and collect only data that is suitable for learning among crawled images using deep learning networks. Also, we offer a method to automatically generate training data by detecting objects and making object coordinates necessary for CNN learning at the same time.