Plankton Classification with Deep Convolutional Neural Networks

Traditional methods for measuring and monitoring plankton populations are time consuming and can not scale to the granularity or scope necessary for large-scale studies. Improved approaches are needed. Manual analysis of the imagery captured by underwater camera system is infeasible. Automated image classification using machine learning tools is an alternative to the manual approach. In this paper, we present a deep neural network model for plankton classification which exploits translational and rotational symmetry. In this work, we propose two constrains in the design of deep convolutional neural network structure to guarantee the performance gain when going deep. Firstly, for each convolutional layer, its capacity of learning more complex patterns should be guaranteed; Secondly, the receptive field of the topmost layer should be no larger than the image region. We also developed a "inception layer" like structure to deal with multi-size imagery input with convolutional neural network. The experimental result on Plankton Set 1.0 imagery data set show the feasibility and effectiveness of the proposed method.