Classification Of Plankton In-Situ Images

This paper studies the classification of plankton digital images based on handcrafted feature and traditional machine learning methods.In feature extraction,this paper not only uses common grayscale features,but also introduces a variety of features with color information,including color histograms,color coherence vectors,color correlograms and two kinds of gradient with color information.In order to alleviate the problem that the dimensionality of features with color information is too high,and to solve the problem that unstructured data cannot be trained in the classifier,this paper proposes the K-means color quantization method and introduces the statistical descriptor of single channel histogram.It is demonstrated by experiment that using the features with color information can obtains higher classification accuracy comparing with only using grayscale features.In the aspect of feature dimension reduction and classifier,this paper applies many combination procedures to improve the training efficiency and classification effect,and gives the applicable scenarios of these combinations.If the training efficiency and recognition rate of the model are considered,the combination of Linear Discriminant Analysis(LDA)and Support Vector Machine(SVM)is recommended.For a test data with 46 classes of 7647 plankton images,with combination of LDA and SVM,the accuracy rate of 96.50%is obtained.If interpretation is considered more,methods combining mutual information or Fisher score and linear SVM or logistic regression,is better choice.In conclusion,color features are effective features to distinguish plankton images,while the traditional classifier can bring more interpretation than CNN without losing classification accuracy.