Mars Crater Detection and Segmentation with MaskRCNN

Dependencies

Tested On Windows 10 pro PC with GPU

conda install tensorflow-gpu
conda install keras

Detection of craters on planetary surface is very crucial for the better understanding of planet topography. It is also important for the selection of landing sites of various lander mission, path planning for rover missions. This project is about application of deep learning method for detection and semantic segmentation of craters in an image. Model trained using transfer learning on pretrained MaskRCNN model. Overall project can be devided into four parts as follow:-

1. Data Preparation

2. Data Inspection

3. Training

4. Testing

With splash.py the trained model was tested on video, although it was only trained with Mars data, learned model able to detect craters on the lunar surface as well.

Mars	Moon

Data Collection and Training-Validation data preparation

Image data collected from the sources and annotated with VIA tool to get json file. Which looks like:-

The Labeled dataset can be downloaded from dropbox. Dataset directory looks like:-

├── train
│   ├── img1.jpg
|   ├── :
|   ├── imgn.jpg
│   └── via_region_data.json
└── val
    ├── img.jpg
    ├── :
    ├── imgq.jpg
    └── via_region_data.json

train and val folder should be inside datasets directory(which is place at root directory)

Trained model mask_rcnn_crater_new.h5 dropbox

Dataset Sources:

1. High Resolution Imaging Science Experiment
2. Mars Science
3. ASU
4. USGS