FishTwoMask R-CNN: Two-stage Mask R-CNN approach for detection of fishplates in high-altitude railroad track drone images
Authors: Saini, A., Singh, D. and Alvarez, M.
Journal: Multimedia Tools and Applications
Publisher: Springer Nature
ISSN: 1380-7501
DOI: 10.1007/s11042-023-15924-7
https://eprints.bournemouth.ac.uk/39962/
Source: Manual
FishTwoMask R-CNN: Two-stage Mask R-CNN approach for detection of fishplates in high-altitude railroad track drone images
Authors: Saini, A., Singh, D. and Alvarez, M.
Journal: Multimedia Tools and Applications
Volume: 83
Pages: 10367-10392
Publisher: Springer Nature
ISSN: 1380-7501
Abstract:Maintenance of railroad track safety is of utmost importance as derailment accidents cause significant loss to life and property. Inspection of railroad tracks and their components is necessary in order to ensure security and well-being of goods as well as humans. Fishplate is an essential component in the railroad track environment hence, periodic maintenance of fishplates is an imperative goal. In this paper, we propose a method for detection and segmentation of fishplate instances in high-altitude drone images (DI) for a closer-view and consequent inspection of fishplate instances. For this purpose, a novel two-stage Mask R-CNN-based framework termed as FishTwoMask R-CNN is proposed. A new fine-tuning strategy has been developed for the purpose of improving the detections in the second stage (Stage 2) which includes a training trick of modifying the loss weights for Stage 2 training. In the first stage (Stage 1), we detect fishplate instances, which are then cropped and fed as input to Stage 2, along with Stage 1 dataset. The Stage 2 network is then trained through a modified weighted loss and produces final detections for segmentation and further inspection. The”layers” hyper-parameter is assigned as “heads” for Stage 1 and updated to “4 + ” for Stage 2. Also, the critical analysis of Mask R-CNN hyper-parameters has been carried out during both the stages which has lead to an improved detection precision rate of 97% in Stage 2 as opposed to 47% in Stage 1. We evaluate our proposed approach on five different test image scenarios in order to view fishplate instance detection results. There has been statistical evaluation on out-of-distribution test images also in order to compute the metrics values. The comparative results have been evaluated using metrics of precision, recall, and F1-score on Mask R-CNN Stage 1 and Stage 2 along with Faster R-CNN and YOLOv5 methods. It is inferred that the proposed approach achieves appreciable metrics values and thus can be gathered suitable for fishplate instance segmentation in drone images.
https://eprints.bournemouth.ac.uk/39962/
Source: BURO EPrints