On the impact of pre-training datasets for matching dendritic identifiers using residual nets

Baaz Jhaj, Ankita Shukla, Pavan K. Turaga, Michael N. Kozicki

Abstract: Dendrites are easy to synthesize branching structures that exhibit randomness; yet they are unique, non-repeatable, and identifiable with the right algorithmic innovations. This has created a novel application area where manufactured dendritic structures are being used as product identifiers - essentially "fingerprints for things". Unlike barcodes, which are linear structures, dendrites exhibit spatial randomness. This, coupled with a unique optical signal generated by light scattering from material inhomogeneities, ensures that each dendrite is unique and unclonable. While there have not yet been any established methods on reading dendritic patterns for verification using image data, identifying dendrites using computer vision techniques could have high potential. Due to limited data and low variance, dendrite identification can be considered to be a fine-grained classification task. In this paper, we examine how the selection of pre-trained models influences dendrite classification. The dendrites we work with share similarity to human fingerprints, thus we begin with a model trained for matching fingerprint data to extract features relevant to dendrites. Additionally, we explore broader pre-training approaches, using ImageNet-1K for our second model and ImageNet-21K for our third model. Surprisingly, our results indicate that even with the visual similarity with human fingerprints, more general pre-training with common image datasets achieves better performance on dendrite classification.