Real-world underwater fish recognition and identification, using sparse representation

摘要

In this paper we describe how a distributed real-time underwater video observational system, developed and operated in southern Taiwan, can be used for visual environmental monitoring of a coral reef ecosystem. The method makes use of an innovative fish recognition and identification technique for real-world automatic underwater observation. Our research demonstrated that advanced fish recognition and identification techniques can be used to study fish populations and to identify species of fish that appear for the first time in particular areas of interest. The observational system subsequently accumulates massive tera-scale video data that can be used for long-term studies on coral reef fish. The system has the capacity for efficient and accurate recognition of fishes from the video dataset, which is recorded in a setting of biological abundance in a coral reef ecosystem. A simple and effective preprocess for fish detection from the video data has been developed, inwhich multiple bounding-surrounding boxes are introduced to discriminate between swimming fish and other moving objects, such as moving sea anemones and drifting water plants. Additional data, including images of various features from a number of fish species, taken at various angles and illumination conditions, can form the basis for a fishcategory database. A maximum probability, partial ranking method, based on sparse representation-based classification (SRC-MP), is proposed for real-world fish recognition and identification. Eigenfaces and Fisherfaces are used to extract feature data, by means of the fish-category database. Two parameters-feature space dimension and partial ranking value-are used to optimize the solutions, in which the recognition and identification rates can respectively achieve 81.8% and 96%. Experimental results show that the proposed approach is robust and highly accurate for the use of fish recognition and identification of real-world underwater observational video data.