Sewer pipelines, as a critical urban infrastructure, are becoming a growing concern for managers as they approach the end of their service life. On the other hand, accessing and inspecting sewer pipelines using methods that involve human entry are problematic and high-risk. Current methods for assessing sewer pipelines utilize various types of equipment for condition inspection. One of the most widely used technologies for sewer pipeline inspection is video (CCTV) cameras. However, employing camera-based methods in extensive sewer networks requires licensed operators to review hours of video footage, which is time-consuming, labor-intensive, and prone to errors. In this study, the performance of two well-known convolutional neural network architectures, VGG16 and ResNet50, was compared for the automatic multi-label detection and classification of defects in sewer inspection video imagery. The proposed models were implemented on a subset of the public Sewer-ML dataset. In both models, a transfer learning approach was employed, with partial fine-tuning applied to the last deep block of the network, using ImageNet pre-trained weights. To address the severe class imbalance, class importance weighting was applied in the loss function, along with independent optimal thresholding for each class. The results show that ResNet50, with an F2CIW score of 61%, outperformed VGG16, which achieved 58%, and demonstrated superior performance in 12 out of the 17 defect classes. Given that ResNet50 achieved this performance with approximately one-fifth the number of parameters of VGG16, this architecture is introduced as a more suitable option for practical deployment in real-time sewer inspection monitoring systems.