Image change combined with CNN power subway vent valve state monitoring
Mengmeng, Lu,Gang, Li,Huasheng, Sun
DOI: https://doi.org/10.1007/s11760-023-02874-8
IF: 1.583
2024-01-25
Signal Image and Video Processing
Abstract:Currently, maintenance of wind valves in rail transit occurs frequently. The implementation of machine vision for remote operation and maintenance would significantly decrease labour requirements. As the number of wind valves on-site is high, it necessitates rich GPU resources. Although large deep learning algorithm models provide strong performance, they require expensive computing power. Therefore, the challenge lies in designing smaller models that reduce GPU computing power to increase efficiency and reduce costs. This study explores the use of a small model in combination with a "virtual coding" process, employing the itti encoder type device. The proposed approach achieves model size reduction while maintaining performance, resulting in a test set accuracy of 98.5%. The F1 index also reflects high accuracy at 99.6%, representing a 1% increase compared to no encoder operation. The approach meets the needs of remote operation and maintenance and can be applied in subway environments. Moreover, successful transition from manual detection to machine detection is realized. This paper suggests a "virtual coding" technique for flexible image transformation. The technique focuses on different data formats including image edge detection, segmentation, feature extraction, dimension reduction, fusion, saliency transformation and feature change. The proposed process offers significant technical summary value for deep learning to incorporate other image change technologies. To ascertain the efficacy of the method, this paper conducted verifications on additional datasets, and the findings confirmed a significant enhancement in its effectiveness.
engineering, electrical & electronic,imaging science & photographic technology