A Classification Method of Image Feature Using Neural Metric Learning for Natural Environment Video

This paper proposes an image feature classification method that applies a distance learning neural network to image feature vectors extracted from an autoencoder. There is active research on similar image retrieval methods using image feature vectors extracted from neural networks. If the image classification performance is not sufficient, it is possible to further improve it by applying a distance learning neural network to convert it into an image feature vector for obtaining appropriate ranking results. In the proposed method, by constructing a model that connects an autoencoder and a distance learning neural network, the reusability of image features extracted from the autoencoder is maintained. In addition, it allows the model to flexibly be combine the autoencoder and distance learning neural network for the model construction. In the experiment, we evaluate the image classification accuracy using an aerial photo dataset provided by the Geospatial Information Authority of Japan and confirm the feasibility of the proposed method.