A deep-learning-based approach for cable duct recognition

As the scale of cable installation continues to expand, the arrangement of cables in manholes has become increasingly complex, making manual surveys more challenging. In response to this issue, this study adopts the YOLOX algorithm as the framework, using the lightweight Swin Transformer network as its backbone. In the recognition model, a convolution attention module is integrated, and the original localization loss function is replaced with the Efficient Intersection over Union (EIoU) loss function. Based on the principles mentioned above, the target recognition algorithm was tested. The test results indicate that the improved YOLOX tracking algorithm achieved an accuracy of 86.64%, with a 2.04% increase in precision, a 2.3% increase in recall, and a 2.14% increase in average detection accuracy. This effectively enables the recognition of cable duct openings.