Research on Optimization of Railway Obstacle Detection Model Based on Neural Architecture Search
The automatic perception of train operating environments leveraging neural networks has emerged as a new approach critical for ensuring the safe operation of trains. However, traditional neural network models primarily rely on a trial-and-error process conducted by human experts, along with accumula...
Saved in:
| Main Authors: | , , , , , |
|---|---|
| Format: | Article |
| Language: | zho |
| Published: |
Editorial Office of Control and Information Technology
2024-08-01
|
| Series: | Kongzhi Yu Xinxi Jishu |
| Subjects: | |
| Online Access: | http://ctet.csrzic.com/thesisDetails#10.13889/j.issn.2096-5427.2024.04.012 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | The automatic perception of train operating environments leveraging neural networks has emerged as a new approach critical for ensuring the safe operation of trains. However, traditional neural network models primarily rely on a trial-and-error process conducted by human experts, along with accumulated experience, which is not only time-consuming and tedious, but also hardly guarantees optimum performance for models. To address this issue, this paper proposes a method for optimizing railway obstacle detection models based on zero-cost neural architecture search. This method begins with the construction of an comprehensive space of potential model architectures. On this basis, the search scope is effectively constrained according to computational workloads required in real-world applications, ensuring that the selected models meet requirements both in accuracy and operational efficiency. The subsequent utilization of a zero-cost neural architecture search algorithm allows for quickly predicting the practical performance of various architectures, without the need for tedious and time-consuming actual training. Furthermore, a comparison of expected performance scores across different models leads to the selection of the optimal option as the ultimate solution. Experimental results demonstrated that this method achieved an average accuracy of 0.711 and an average inference time of 6.12 ms per frame, significantly outperforming baseline models with equivalent computational loads. |
|---|---|
| ISSN: | 2096-5427 |