Fish Species Identification in Real-Life Underwater Images

Abstract

Kernel descriptors consist in finite-dimensional vectors extracted from image patches and designed in such a way that the dot product approximates a nonlinear kernel, whose projection feature space would be high-dimensional. Recently, they have been successfully used for fine-gradined object recogntion, and in this work we study the application of two such descriptors, called EMK and KDES (respectively designed as a kernelized generalization of the common bag-of-words and histogram-of-gradient approaches) to the MAED 2014 Fish Classification task, consisting of about 50,000 underwater images from 10 fish species.