Computer vision for virus image classification


In this paper we present a new ensemble of descriptors for the classification of transmission electron microscopy images of viruses that is based on texture analysis. A set of six well-assessed texture descriptors, namely Local Binary Pattern (LBP), Local Ternary Pattern (LTP), Dense LBP (DLBP), Rotation Invariant co-occurrence among LBP (RI), Local Phase Quantization (LPQ), and LBP Histogram Fourier (LHF), are combined with innovative approaches to improve their performances in virus classification. To generate new variants of the aforementioned descriptors, different approaches are applied: (i) the Edge approach (ED), which extracts the textural information from specific regions of the image instead of from the original image; (ii) the Bag of Features (BOF) scheme, which is used to build vocabularies of the most representative patterns; and (iii) Multi-Quinary coding (MQ). Moreover, to demonstrate the generalizability and applicability of ED and BOF, they are tested on additional datasets containing subcellular parts and tissues. We observe that ED improves the performance of the single standard descriptors, while BOF is particularly effective in fusion by sum rule with the standard application of texture descriptors (i.e., when they are extracted from the entire image). In contrast, the combination of MQ with LHF, DLBP, and RI does not improve virus classification. Using these results, we suggest a new ensemble of descriptors called NewF, which is based on the best methods investigated in this paper, as well as on some other state-of-the-art descriptors. NewF accuracy in virus classification is 85.7%, outperforming previous methods proposed in the literature for the same task (i.e., virus classification using the object scale dataset). The MATLAB code for our methods and NewF descriptor are available at

Keywords texture descriptors; feature selection; local binary patterns; ensemble of descriptors.

[full paper]