Dipartimento di Ingegneria informatica, automatica e gestionale

Microplastics derived from fiber shredding are recognized by the scientific community as one of the main sources of microplastic water pollution, thus actualizing the need for techniques able to identify them with high accuracy. The recently released Holography Micro-Plastic Dataset offers the opportunity to test deep neural networks in their ability to distinguish between microplastics and other debris on a standard benchmark. The promising results obtained from the initial batch of experiments can be further improved through a combined approach which involves different image mapping techniques and recent state-of-the-art deep models. Within this framework, we analyze various image fusion schemas to merge the paired dataset images (amplitude and phase) into a single three-channel picture. We demonstrate that our proposed approach yields increased accuracy compared to both single-image data processing and other fusion techniques. Finally, the performance of our method is further enhanced by employing the Vision Transformer model as backbone, highlighting the effectiveness of the proposed approach in microplastics classification.