Automatic Bird-Species Recognition using the Deep Learning and Web Data Mining

摘要

In this paper, only the bird-species recognition keywords are input, and the Web image is refined as the learning data. We propose a method to generate a bird-species recognition model through learning a model based on refined training data.First, if you enter the name of the targeted bird breed, the image will be collected from the Web using the image crawl. To refine the collected images into the training dataset, the corrupted image is corrected and deleted, the outlier is removed, and finally the image is expanded to obtain the refined training data. In the process of modifying and deleting the images, the white-background image is deleted and the header is initialized. In the outlier removal, the features are extracted using the deep learning of the image data collected for each keyword. Then, the cluster distance of each label is measured using the K-means clustering that forms the training data according to the measured cluster distance. Lastly, the image expansion is performed to improve the training of the training data and for the learning accuracy of the refined database.To evaluate the performance of the proposed method, CUB-200 [1], an existing Bird Breed Dataset, is divided into training data and test data. In addition, the recognition-rate change according to various parameter changes is confirmed. As a result, the proposed method shows different recognition rates than the data collected from the existing institutions. We can use this proposal to provide varieties of training data and automated processes compared with the existing methods.