Deep Learning for Retrieval of Natural Flower Videos

Abstract

In this paper we have proposed a flower video retrieval system using Deep learning approach for training and learning features. Network is trained in three ways; with keyframes, with segmented flowers and with gradient of the flowers as input. For a given query video, the system retrieves similar videos from the database using Multiclass Support Vector Machine (MSVM). An extensive experimentation has been conducted on a relatively huge flower video data set of our own consisting of more than 2600 flower videos belonging to 30 different classes of flowers. It is observed that the complexity of the system can be reduced to a larger extend by training the DCNN using gradient of the flowers without compromising the performance. However, among the three ways, the system trained with segmented keyframes has shown best retrieval accuracy. It has also been observed that the proposed DCNN based approach out performs traditional retrieval approach.