Paper #1 - Vision-Based Railway Track Monitoring Using Deep Learning
Key Notes
- Vision models to detect sunkinks, loose ballast and railway assets like switches and signals
Models Analysis
- Inception - 2015 - Instead of stacking convolutional layers in a sequential manner, if convolutions with different filter sizes can be processed in parallel and then concatenated and fed to next layer
- ResNet - 2015 - Inputs of a lower layer are added to the inputs of a higher layer due to which higher layers make inference based on both the extracted feature maps and the original inputs
- FRCNN - Region proposal method to identify bounding boxes with highest probability
Defects
Loose Ballast - A track is identified as having loose ballast when it doesn’t have enough gravel distributed in between tracks making depth of cross ties visible
Sunkink - model just needs to identify if the tracks are almost straight and parallel or not. Rail malformations known as a "sun kink" or "track buckle“
Steps
- (A) ROI IN THE FRAME,
- (B) IMAGE AFTER PERSPECTIVE TRANSFORMED WITH GAUGE MEASUREMENTS FROM PROCESSED FRAME
- (C) PROCESSED ROI
- (D) AND (E) TRAINING IMAGES SIMULATED IN PAINT,
- (F) SUNKINK BEING DETECTED IN A VIDEOFIGURE 4
Signal Color - Once the signal is detected, red and green color masks are applied on the image part inside the predicted bounding box to identify signal color.
Track Health Index - All the track defects that can be identified will be given a weightage based on their severity
Paper #2 - Railway Track Specific Traffic Signal Selection Using Deep Learning
Key Notes
- Challenges - Urban areas, multiple lines run in parallel
- Approach - position of signal is highly dependent on the position of the current track relative to the other tracks present.
- FRCNN model was used for signal detection and color detection was added on top of it.
Paper #3 - Holistically-Nested Edge Detection
Key Notes
- End-to-end edge detection system
- Canny Edge Draw back - exhibit spatial shift and inconsistency.
- Recent Architectures - [10], DeepContour [34], DeepEdge [2], and CSCNN [19].
HED
- CNN for patch-based edge prediction
- Structured Edges (SE) - incorporating structural information, fusing multi-scale responses
- End-to-end edge detection system, a strategy inspired by fully convolutional neural networks
- Multi-scale deep learning into four categories, namely, multi-stream learning, skip-net learning, a single model running on multiple inputs, and training of independent networks.
Techniques to Improve Signal - Noise in Image
- Gray conversion
- Histogram equalization
- Binary Image
- Classification Custom Model
Issues
- Crack in a track
- Rivets in a track
Ref - Link
Paper #4 - Automatic Detection of Objects of Interest from Rail Track Images
Keynotes
- Powerful, smaller and cheaper, automatic visual inspection systems
Detects the locations of the fasteners in an input rail track image
This system should be able to detect various objects such as fastening elements (bolts, insulated block joints, clamps, clips, etc.) by inspecting the images acquired by a digital camera installed under a diagnostic train.
- Traditional Techniques
- Ultrasound inspection
Magnetic methods, such as eddy current inspection, magnetic particle inspection (MPI), magnetic induction, magnetic flux leakage (MFL), electromagnetic acoustic transducer (EMAT)
- Ground penetrating radar (GPR)
- Laser light inspection
- Infrared inspection
- X-ray inspection
- Spectral analysis of surface waves (SASW)
- Impact-echo techniques
- Impulse-response techniques
Approach
- Feature extraction algorithm, and then provided to a classifier.
Happy Learning!!!
