Marine animal classification using combined CNN and hand-designed image features

Digital imagery and video have been widely used in many undersea applications. Online automated labeling of marine animals in such video clips comprises of three major steps: detection and tracking, feature extraction and classification. The latter two aspects are the focus of this paper. Feature extracted from convolutional neural network (CNN) is tested on two real-world marine animal datasets (Taiwan sea fish and Monterey Bay Aquarium Research Institute (MBARI) benthic animal), and yields better classification results than existing approaches. Appropriate combination of CNN and hand-designed features can achieve even higher accuracy than applying CNN alone. The group feature selection scheme, which is a modified version of the minimal-redundancy-maximal-relevance (mRMR) algorithm, serves as the criterion for selecting an optimal set of hand-designed features. Performance of CNN and hand-designed features are further examined for images with lowered quality that emulates bad lighting condition in water.