scumechanics/GPR-Object-Detection

This repository contains code to train object detection models like FRCNN for identifying objects in Ground Penetrating Radar scans. It also contains code to generate fake data using GANs.

Detecting Objects in Ground Penetrating Radars Scans

We used mmdetection library for fine-tuining Faster RCNN, Mask RCNN, and YOLO to identify objects from Ground penetrating radar scans.

Data:

• Used publicaly available dataset containing 171 annotated GPR scans from https://github.com/irenexychen/gpr-data-classifier.
• Converted the annotations from .xml format to .coco format using https://github.com/yukkyo/voc2coco.

Creating fake data using GANS:

• Code: GPR_GAN_071521_v.0.4.ipynb

GANs architecture(created from scratch):

• Generator:
  • Input: A tensor of size (100, 1, 1) filled with random numbers from a normal distribution(mean = 0, variance = 1)
  • Output: A tensor of size (3, 128 , 128) representing Image
  • NN Architecture: 6 transpose convolutions with batch normalization and ReLU activations.
• Discriminator:
  • Input: (3, 128, 128) size tensor either real or fake
  • Output: Classification of input as either real(1) or fake(0) 7 Conv layers with batch normalization and ReLU
• Loss: Binary Cross Entropy Loss

GANs results:

Hyperbola detection:

• Code: GPR_MMDetection_v0.022.ipynb
• We finetuned multiple object detection models with various checkpoints and due to the scarsity of GPR scans, the FRCNN model pretrained on MS-COCO dataset with resnet-101 as backbone gave the best results.

Results:

• We finetuned FRCNN with both real images and fake images from GANs
• Achieved mAP 0.90 in detecting hyperbolas