Sewer-MoE: A tuned Mixture of Experts Model for Sewer Defect Classification

Sijie Xu, Yifan Gong, Xiangrong Zhu, Shuo Chang, Yuhe Wang, Liqi Yan

Abstract: Abstract: Inspection of pipelines is particularly important for the drainage industry, and automation of this process has received a lot of attention. We propose the Mixture of Experts for Sewer Defect Classification (Sewer-MoE), an innovative model for identifying pipe defects, in which we train multiple expert models and then merge them into a single multiclassification model. During the model training process, we produced an attention mechanism structure that allows each expert model to refer to the other expert models, while weighting each classification to emphasise the defect types with fewer occurrences, effectively improving the prediction accuracy. We evaluate our Mixture of Experts for Sewer Defect Classification (Sewer-MoE) on the Sewer-ML dataset, where we use our model to compare the model proposed by Xie et al. and the model proposed by Chen et al. with the original model after modifying them, and our model significantly outperforms the original model on the same size dataset.