| Plankton including phytoplankton and zooplankton form the base of the food chain in the ocean and are a fundamental component of marine ecosystem dynamics. The rapid mapping of plankton abundance together with taxonomic and size composition can help the oceanographic researchers understand how climate change and human activities affect marine ecosystems.; Recently the University of South Florida developed the Shadowed Image Particle Profiling and Evaluation Recorder (SIPPER), an underwater video system which can continuously capture the magnified plankton images in the ocean. The SIPPER images differ from those used for most previous research in four aspects: (i) the images are much noisier, (ii) the objects are deformable and often partially occluded, (iii) the images are projection variant, i.e., the images are video records of three-dimensional objects in arbitrary positions and orientations, and (iv) the images are binary thus are lack of texture information. To deal with these difficulties, we implement three most valuable general features (i.e., moment invariants, Fourier descriptors, and granulometries) and propose a set of specific features such as circular projections, boundary smoothness, and object density to form a more complete description of the binary plankton patterns. These features are translation, scale, and rotation invariant. Moreover, they are less sensitive to noise. High-quality features will surely benefit the overall performance of the plankton recognition system.; Since all the features are extracted from the same plankton pattern, they may contain much redundant information and noise as well. Different types of features are incompatible in length and scale and the combined feature vector has a higher dimensionality. To make the best of these features for the binary SIPPER plankton image classification, we propose a two-stage PCA based scheme for feature selection, combination, and normalization. The first-stage PCA is used to compact every long feature vector by removing the redundant information and reduce noise as well, and the second-stage PCA is employed to compact the combined feature vector by eliminating the correlative information among different types of features. In addition, we normalize every component in the combined feature vector to the same scale according to its mean value and variance. In doing so, we reduce the computation time for the later recognition stage, and improve the classification accuracy.; Due to the complexity of plankton recognition problem, it is difficult to pursue a single optimal classifier to meet all the requirements. In this work, instead of developing a single sophisticated classifier, we propose an ensemble learning framework based on the random sampling techniques including random subspace and bagging. In the random subspace method, a set of low-dimensional subspaces are generated by randomly sampling on the feature space, and multiple classifiers constructed from these random subspaces are combined to yield a powerful classifier. In the bagging approach, a number of independent bootstrap replicates are generated by randomly sampling with replacement on the training set. A classifier is trained on each replicate, and the final result is produced by integrating all the classifiers using majority voting. Using random sampling, the constructed classifiers are stable and multiple classifiers cover the entire feature space or the whole training set without losing discriminative information. Thus, good performance can be achieved. Experimental results demonstrate the effectiveness of the random sampling techniques for improving the system performance.; On the other hand, in previous approaches, normally the samples of all the plankton classes are used for a single classifier training. It may be difficult to select one feature space to optimally represent and classify all the patterns. Therefore, the overall accuracy rate may be low. In this work, we propose a pai... |
