CAFEN: A Correlation-Aware Feature Enhancement Network for Sewer Defect Identification

Sewer defect identification aims to find out defects from the sewer inspection video frames, such as cracks, breaks, collapses and so on, to prevent pipeline accidents. Currently, most of the identification methods are based on single-label classification models, semantic segmentation models, or object detection models, which have achieved considerable results. However, these methods have their own limitations respectively. For example, the single-label classification-based methods can not recognize multiple defects at a time, but there may exist different types of defects in one pipeline section. The semantic segmentation-based methods are easily influenced by illumination conditions and video resolutions. And the object detection-based methods always need to regress bounding boxes, which is unnecessary as we can use the accompanied geographical coordinates to fast localize the place where the defects occur. Moreover, it also brings additional difficulties to the training process when defects are spatially overlapped. To tackle these problems, in this paper, we propose a correlation-aware feature enhancement network (CAFEN) for sewer defect identification, significantly boosting the performance of defect classifiers. Specifically, we present a graph-based learning module, label correlation learning module (LCL), to capture the correlation information among defect labels. Then we introduce the convolutional block attention module (CBAM) to capture the context information from the feature maps. Extracted features from the above two modules are fused to further enhance the feature representation. Furthermore, we design a class importance weights (CIW) loss function to prioritize different types of defects, enabling our model better recognize the important defects. Experiment results demonstrate that our proposed CAFEN outperforms previous models and achieves state-of-the-art performance on the Sewer-ML bench-mark.