A Railway Communication Cable Fixtures Detection Algorithm

The goal of this paper is to accomplishing the automatic detection for railway cable fixtures. Since Yolov4-tiny achieves relatively lower recall and precision on the fixture dataset, we modify its Cross Stage Partial (CSP) architecture to reduce the repeated gradient information, so that the modified architecture achieves a better performance on the communication cable fixtures dataset. Meanwhile, our algorithm modifies the classification loss function by introducing a heuristic penalty factor to pay more attention on the minor classes in the imbalanced dataset. Experimental results show that the proposed algorithm has increased recall and precision by 3.3% and 2.4% compared with the baseline algorithm, respectively.