Feasibility of computer-aided identification of foraminiferal tests

Over the past two decades, several investigators have worked on computerized systems to accelerate the identification of foraminiferal tests (forams). Leading examples have focused on fully-automatic identification using neural networks and supervised learning. This paper introduces an alternative semi-automatic or computer-aided approach. Such an approach reduces the workload associated with foram identification without the challenges of training set collection and fully-automatic recognition. The proposed method begins by photographing a collection of specimens sprinkled on a microscopy slide. Segmented images are then mapped into a canonical space where position, rotation, and scale are normalized. Specimens are clustered based on image similarity in the canonical space. A specialist then identifies the clusters by inspecting representative templates. Experimental results show that the identification effort can be reduced by 35%, yet the accuracy remains comparable to when every specimen is individually identified. Further reduction of effort was prevented by the significant variability of illumination direction in the canonical images. These results encourage further work on a computer-aided approach to foram identification.