Recognizing white blood cells with local image descriptors

摘要: Abstract Automatic and reliable classification of images white blood cells is desirable for inexpensive, quick accurate health diagnosis worldwide. In contrast to previous approaches which tend rely on image segmentation a careful choice ad hoc (geometric) features, we explore the possibilities local descriptors, since they are simple approach, require no explicit segmentation, yet have been shown be quite robust against background distraction in number visual tasks. Despite its potential, this methodology remains unexplored problem. work, therefore characterized with well-known bag-of-words approach. Three keypoint detectors five regular sampling strategies studied compared. The results indicate that approach encouraging, both sparse dense can perform reasonably well (mean accuracies about 80% obtained), competitive segmentation-based approaches. Two main findings as follows. First, points, detector localizes keypoints cell contour (oFAST) performs somehow better than other two (SIFT CenSurE). Second, interestingly, partly contrary our expectations, including hierarchical spatial information, multi-resolution encoding, or foveal-like sampling, clearly outperform simpler uniform-sampling considered. From broader perspective expert intelligent systems, relevance proposed that, it very general problem-agnostic, makes unnecessary human expertise elicited form cues; only labels type required from domain experts.