Pipeline defect detection based on feature remapping

Traditional underground drainage pipeline detection requires a lot of manpower to complete, which faces not only the challenge of time but also the challenge of efficiency. In recent years, the development of deep learning has played an amazing role in many fields, especially in pattern classification. Usually, there are some problems in real pipeline detection tasks, such as multiple sample classes which usually includes 1 normal class and N defect classes, the problem of normal class's imbalanced sample proportions, and the problem of similar features, that is, the features between partial normal class samples and defective class samples are similar. The above problems will lead to a lower precision rate of normal class. Consequently, first, for the multi-class problem, we propose to integrate 2-class classifier (normal class and defect class) and multi-class classifier (normal and specific defect classes). Second, for the problem of imbalance of normal samples, specifically, because the pipeline samples usually include a normal class and multiple defect classes, the proportion of normal samples in the multi-class classifier is small which will hurt the precision rate of normal class. Therefore, we propose to use feature remapping in the integration of 2-class classifier and multi-class classifier, that is, the trained 2-class classifier supervises the training of the multi-class classifier. At the same time, through feature remapping, the normal class features output of the multi-class classifier are approximated to the normal class features output by the 2-class classifier. These can not only alleviate the problem of normal sample imbalance, but also make full use of the gap between classes to ease the problem of similar features and improve the precision of normal classes. Third, in order to prevent information loss in the process of feature transfer along the network, we adopt the feature transfer enhancement module to improve the feature transfer capability of the network. Besides, the feature transfer enhancement module can also be beneficial to the problem of similar features. This method allows us to achieve a 9% gain in the normal category sample classification, which will be shown in the experiment.