U-Net skip-connection architectures for the automated counting of microplastics

Water pollution is a widespread problem, with lakes, rivers, and oceans contaminated by an increasing amount of microplastics and other pollutants. Microplastic counting from microscope images is a laborious, time-consuming, and error-prone task. The ability of researchers to automate the detection and counting of microplastics would accelerate research and monitoring activities. This paper applies machine learning techniques to automatically segment and count microplastics in a given image, in challenging cluttered conditions. A U-Net neural network was trained to segment microplastics and image post-processing techniques were then applied to count the number of microplastics as well as highlight their position in an image. Different forms of skip connections from the U-Net encoder layers to decoder layers were tested to assess the impact of skip connections on the performance of the U-Net architecture. Our work shows that U-Net can achieve human-level performance in enumerating microplastics in cluttered images and that the standard skip-connection architecture is not necessarily optimal.