[A-OS10] 陸域海洋相互作用ー惑星スケールの物質輸送

[AOS10-P02] A Proposal of a Preprocessing Using Deep Reinforcement Learning for Improving the Accuracy of Marine Animal Identification

*岡村 森¹、山敷 庸亮¹ (1.京都大学総合生存学館)

キーワード：海洋生物学

Since Krizhevsky et al. proposed a convolutional neural network method as a highly accurate model for image recognition in 2012, the field of computer vision has achieved rapid development. As a result, the algorithms are being used in every research field. In the field of marine ecosystems, techniques such as deep learning are beginning to be used for species identification. However, unlike ordinary images, the accuracy of image identification of species living in water tends to be relatively low. Images were taken underwater often have a blue-black filtered look due to the effect of light attenuation, which makes it impossible to extract enough features from the image, resulting in a phenomenon that makes it difficult to increase accuracy. In order to prevent this phenomenon, it is common practice to pre-process the images when performing identification using underwater images. Preprocessing of underwater images has been widely studied, and several methods have been proposed, such as those that take into account environmental factors at the time the image was acquired and those that use deep learning. In this study, we propose a preprocessing method for underwater images using deep reinforcement learning based on the above background. This method is implemented using Deep Q-Network Learning, a deep reinforcement learning algorithm commonly used in reinforcement learning, and a trained discriminative model, which is a preprocessing model specialized for improving the accuracy of image identification of specific species. Unlike many previous studies on preprocessing of underwater images, this preprocessing model does not need to take into account environmental factors such as the lighting conditions at the time the image was acquired, suspended particles in the water, or the presence of artificial light sources, The system spontaneously searches for an appropriate combination of pre-processing and learns the filter to increase the accuracy of the species discriminator. Therefore, it is expected to reduce all the costs related to the preprocessing of underwater images and to increase the accuracy of species identification. In this study, the automatic preprocessing model described above was used to verify the identification accuracy of Dendronephthya gigantea, a species of coral reef. a species of soft coral reef. As a result, the generated preprocessed identification model improved the prediction accuracy by 11% compared to the original model. In addition, by combining the generated preprocessed classifier model with the original model, we confirmed a significant improvement in the prediction error, from 2.05 of the original model to 1.27. This result confirms that the proposed preprocessing model for the classifier for underwater images can contribute to the improvement of the accuracy and robustness of the classifier. In future research, we plan to use the automatic preprocessing method developed in this study for the efficient evaluation of the ecosystem from underwater images.