Multi-scale feature flow alignment fusion with Transformer for the microscopic images segmentation of activated sludge

Abstract

Accurate microscopic images segmentation of activated sludge is essential for monitoring wastewater treatment processes. However, it is a challenging task due to poor contrast, artifacts, morphological similarities, and distribution imbalance. A novel image segmentation model (FafFormer) was developed in the work based on Transformer that incorporated pyramid pooling and flow alignment fusion. Pyramid Pooling Module was used to extract multi-scale features of flocs and filamentous bacteria with different morphology in the encoder. Multi-scale features were fused by flow alignment fusion module in the decoder. The module used generated semantic flow as auxiliary information to restore boundary details and facilitate fine-grained upsampling. The Focal–Lovász Loss was designed to handle class imbalance for filamentous bacteria and flocs. Image-segmentation experiments were conducted on an activated sludge dataset from a municipal wastewater treatment plant. FafFormer showed relative superiority in accuracy and reliability, especially for filamentous bacteria compared to existing models.