Deep Learning based Surface Defect Detection of Rail Tracks
Benefits of the Project (less than 2500 characters)
2025-06-28 16:31:06 - Adil Khan
Project Title
Project Objectives
Deep Learning based Surface Defect Detection of Rail Tracks
Project Area of Specialization Artificial IntelligenceProject Summary- Railways are quick and cost effective medium for transportation / goods delivery. Due to wear and tear from train and natural environment, rail tracks develop variety of defects on its surface such as holes, cracks, wheel burns, and corrosion etc. These defects jeopardize the safety of railway thus railway inspection should be done periodically, preferably at night to avoid rush hour.
- At present, defect detection relies mainly on manual inspection, but this method is labor-intensive, has lower efficiency and yields less creditable results as workers tire easily. Thus, a need for reliable and automated defect inspection system is needed.
- This project aims to develop such a system that makes use of Computer Vision and Artificial Intelligence techniques to provide reliable and automated defect detection system for rail tracks, it will operate in real-time to detect defects. This system will mainly aim at common rail surface defects such as crack, holes and other surface deformations.
Project Objectives - Image Acquisition system – for capturing rail track’s surface image.
- Deep learning model – for processing input images to identify defects.
- GPS system – for extracting geographical location of defect.
- Data: Dataset is comprised of defective and non defective images of rail track’s surface and other surfaces such as steel. Some data is collected online but more data will be collected from local railway tracks.
- Neural Network: A encoder-decoder based network will be trained to reconstruct non-defective images. Once the network can reconstruct non-defective images with sufficient accuracy, it will be then fed defective image, since the reconstruction will be poor on defective images, network will classify that image anamolous. This way we would be able to detect defective railway track via visual information.
- Image capture mechanism: A remote controlled rig will be developed with rail track compatible tires. That rig will house image acquisition system and other hardware. After a certain distance covered, image acquisition system and GPS will be simultaneously triggered to get track’s image and corresponding gps coordinates. That image will read by embedded system on board and fed into neural network that we previously trained. Once image is classified as defective, maintenance engineer on site can then use GPS coordinates with the defective image to locate and repair defect.
- Safer railway travel.
- Quicker & reliable inspection of railtrack.
- Low operational costs as compared to manual labour.
- Maximum up-time of railway.
- Faster maintenance of defects.
- Image Acquisition System: Image acquisition system consists of 2 other subsystems: Illumination system and Camera System.
- Illumination System: It consists of LED light sources in all 4 directions (front, back, left, right). The light sources are placed at a low angle with respect to the surface, so that light reflected from defective area (i.e. crack or hole) appears as high contrast pixels in image. This is known as dark field illumination.
- Camera System: Camera is facing directly down at the track to capture images of rail track’s surface (shown below). Camera Specs: 60 fps, monochrome, global shutter mechanism.
- GPS System: A centimeter level accurate GPS system to be employed to accurately determine geo-location of defects.
- Data collecting rig: Data collecting rig will be designed and fabricated, with rail track compatible tires, to mount image acquisition system and house other hardware for data collection and track’s inspection.
- Deep learning model: Since dataset consists of unlabeled data, anomaly in images will be detected using unsupervised learning methods. Neural network architectures such as Auto encoders, Variational Auto encoders, Generative Adversarial Networks, Convolutional Auto encoders and their other variants are trained and tested. Best model will be employed in final system. Neural network’s task is divided into 2 chunks: Defect Classification and Defect Segmentation. Neural Network’s development is divided into following parts: Training, Fine-tuning, Testing.
- Training: In training, model is first trained on defect free images. Once the model can accurately reconstruct defect free images, it is used to fine tune parameters that will classify and segment defects in defective images.
- Fine-tuning: In fine-tuning, using trained model, minimum defective area and best threshold value pair is estimated to classify defective image.
- Testing: In testing phase, defective images are processed by model. Model’s accuracy is determined by true-positive ratio and true-negative ratio.
- Loss function: Structural similarity index measurement (SSIM) is used as loss function. A comparison between L2 loss and SSIM loss in segmenting defects is shown below. SSIM loss is better since it captures image’s structural, contrast and luminance information to derive results.
| Item Name | Type | No. of Units | Per Unit Cost (in Rs) | Total (in Rs) |
|---|---|---|---|---|
| Total in (Rs) | 80000 | |||
| Inspection Camera | Equipment | 1 | 15000 | 15000 |
| Illumination system | Equipment | 1 | 8000 | 8000 |
| 3D printed fixture for camera & illumination system. | Equipment | 1 | 10000 | 10000 |
| Data Collection rig | Equipment | 1 | 15000 | 15000 |
| GPS Hardware | Equipment | 1 | 7000 | 7000 |
| Jetson nano embedded system | Equipment | 1 | 10000 | 10000 |
| Equipment shipping | Miscellaneous | 1 | 5000 | 5000 |
| batteries | Miscellaneous | 1 | 5000 | 5000 |
| Motors and RC control for Data collectioon rig. | Equipment | 1 | 5000 | 5000 |