No abstract

Microplastic pollution poses severe environmental problems. Developing effective imaging tools for the identification and analysis of microplastics is a critical step to curtail their proliferation. Digital holographic imaging can record the morphological and refractive index information of such small plastic fragments, yet due to the heterogeneous sampling environments and variations in the microplastic shapes, traditional supervised learning methods are of limited use. In this work, we pioneer a zero-shot learning method that combines the holographic images with their semantic attributes to identify the microplastics in heterogeneous samples, even if they have not appeared in the training dataset. It makes use of the attention mechanism for image feature extraction, and the Kullback-Leibler divergence both to alleviate the domain shift problem and to guide the training of the mapping function. Experimental results demonstrate the effectiveness of our approach, and the potential use in a wide variety of environmental pollution assessment.

No abstract