On Confidences and Their Use in (Semi-)Automatic Multi-Image Taxa Identification

We analyzed classification confidences in biological multi-image taxa identification problems, where each specimen is represented by multiple images. We observed that confidences can be exploited to progress toward semi-automated identification process, where images are initially classified using a convolutional neural network and taxonomic experts manually inspect only the samples with a low confidence. We studied different ways to evaluate confidences and concluded that the difference of the largest and second largest values in unnormalized network outputs leads to best results. Furthermore, we compared different ways to use image-wise confidences when deciding on the final identification using all the input images of a specimen. The best results were obtained using a confidence-weighted sum rule over the unnormalized outputs. This approach also outperformed the evaluated supervised decision method.