Flash SewerNet: A lightweight expert network with Gaussian-multimodal fusion for sewer defect recognition

Sewer defect recognition from video inspections is a critical but challenging task due to three persistent issues in such harsh environments: severe image contrast degradation from sensor noise, which obscures defect features; high cross-similarity between different defect and material classes; and the difficulty in distinguishing infrequent defect types from the background, given scarce samples and weak features. These challenges significantly limit the accuracy, robustness, and practical deployment of existing models. To overcome these limitations, we propose Flash SewerNet, a lightweight expert network that effectively incorporates multimodal prior knowledge for robust sewer defect recognition. This network synergizes Gaussian modeling with multimodal priors, paired with an expert-assisted multi-label recognition head. This architecture is specifically co-designed for high stability and computational efficiency, making it suitable for real-time video analysis on resource-limited edge devices. Extensive experiments show that Flash SewerNet achieves state-of-the-art performance on several sewer inspection benchmarks. More importantly, we demonstrate its practical viability by deploying the model on an embedded platform, where it achieves a high inference speed sufficient for continuous, in-situ defect diagnosis in real-world sewer inspection videos. This work not only presents a robust detection solution but also highlights the essential role of multimodal domain knowledge in advancing vision systems for challenging real-world applications. The code and models are available at: https://github.com/zkkzxk/Flash-SewerNet.