PipeMamba: State Space Model for Efficient Video-based Sewer Defect Classification

In the community of sewer pipe defect analysis, CNN-based methods have suffered from modeling the long-range dependencies in the long-sequence videos, such as closed-circuit television (CCTV-) and quick-view (QV)-based videos, resulting in the suboptimal performance. Besides, Transformer-based methods are limited in the real-time applications due to their high computational complexity during the inference. Recently, Mamba-based methods have performed the significant capability with a trade-off between accuracy and efficiency, which present a promising direction for modeling an efficient defect inspection method. Nevertheless, directly adapting Mamba to sewer pipe defect inspection achieves a poor performance, since the multi-focus complementarity of representing the sewer defects is ignored, and the bidirectional scan method disrupts the spatial and semantic continuity of the sewer pipes. To address these challenges, we propose a method for the video-based sewer pipe defect classification using a state space model, namely PipeMamba. Specifically, a flow-based selection module (FSM) is designed to split the sewer video into the long-focus and shortfocus segments, where the multi-focus segments can be modeled to represent the multi-scale information of sewer defects. Then, a novel circle scan strategy (CSS) is proposed to specify the appearances of the sewer pipes, which aims to capture the sewer defects efficiently without disrupting the semantic consistency and spatial continuity of the sewer pipes. Extensive experiments conducted on the public benchmark demonstrate the superiority of PipeMamba, which achieves the inference speed nearly 3 to 7 times faster than the Transformer-based methods, while maintaining high classification performance. Code is available at: https://anonymous.4open.science/r/PipeMamba-3637.