Microplastic Binary Segmentation with Resolution Fusion and Large Convolution Kernels

The term “microplastic” refers to plastic particles with a length or diameter of less than 5 mm that do not easily decompose in the natural environment and persist for a long time. These microplastics have adverse effects on the marine ecosystem when they enter the ocean. Therefore, it is necessary to estimate the amount of microplastics in rivers and sewers and to block the outflow of microplastics in areas where they are found to be present at high levels. However, estimating the amount of microplastics first requires detecting these particles, which is not an easy task to complete efficiently and accurately due to their small size and the difficulty involved in distinguishing them from organic materials. The current study therefore proposes a new model structure for microplastic segmentation. This model uses the multi-resolution fusion module (MRFM), which is known to significantly contribute to the segmentation performance in HRNet, and this model employs the EfficientNetV2B3 model as a backbone. We also utilize large convolution kernels to achieve better feature extraction from the inputs of three resolution stages that are closer to the input image resolution. The experimental results showed that the model using two MRFMs outperformed the model using feature pyramid network in the head network, with improvements of 3.28% in IoU and 2.67% in F1-score.