Shape retrieval using deep autoencoder learning representation

Deep learning has shown to be very effective in variety of applications including image classification and object recognition. In this paper we use deep autoencoder for compact shape representation learning and image retrieval. In this method the autoencoder is a 4-layer coding network, and the original shape images after scale normalization are used to pre-train the autoencoder in an unsupervised way. Then fine-tuning procedure is executed on the 4-layer coding network to refine the learned weight matrixes. Finally a learned 40-dimension vector for each shape image is used as its features, and the similarity between any two shapes is measured using standard cosine similarity. The image retrieval performance of the proposed method is evaluated on the Swedish leaf database using precision and recall measurement, and compared with the classical Fourier shape descriptor. The experimental results indicate that the proposed method reaches the higher precision at the same recall value among compared methods.