Model_Information.txt

Key Terms:
1) RDB -> Residual Dense Blocks -> To extract local features from the convoluted matrix 
2) Convolution Layer Output -> Transfer the features (distinguishing feature) to every other position of the image to see how the features match that area. 
  			    -> Like for matching a 'O' character will require curve (feature) around each corner cube => Convolution of image with all the features
3) ReLU -> Rectified Linear Unit -> Transform Function
4) Pooling Layer -> To reduce the larger feature matrix to get a condensed and approximated output (usually max of a particular window of certain size)
5) RPN -> Region Proposal Network -> Convolution Netowrk used for predicting object bound
6) FPN -> Feature Pyramid Network -> Feature Extractor for single image to provide multi-level feature map in convolutional manner
7) Ellipse R-CNN (Region based convolutional neural network) -> To infer elliptical objects from the image
8) DenseNet -> Dense CNN -> Used during intermediate accuracy improvement during long stage layers in the model by letting the future layers have access of previous layers
	    -> Used for generating the terrain map (To combine and improve the result of each processes that happens during the model creation)
9) Monocular Deep Transform -> Estimating the depth value (distance relative to the camera) of each pixel given a single (monocular) RGB image using gradient descent to train the model.

Process:
1) Image Super Resolution -> Convert low resolution moon image into higher resolution
2) Depth Estimation -> Estimation the depth of the moon surface using the depth mask (For both low res and high res images)
3) Crater Detection -> Detecting craters using the result from depth estimation to get accurate result
4) Use the information from the crater detection (areas regarded as craters) for the accurate creation of hazard map (based of depth => improving depth estimation)
5) Terrain Mapping -> Creating a geographical map based on the output of the above models/processes.
6) Ultimately, a hazard map will be created for the moon excavation missson.

Lunar Navigation:
Generation of Hazard map at 1m grid spacing (1m height resolution) using 5m spatial resolution data for safely navigating a Lander to a safe landing site using Super resolution techniques.

Super Resolution:
1) Objective of this section is to upscale and improve the quality of low resolution images taken by terrain mapping cameras from ISRO.
2) Keras: Model Processing, Training, Evaluation. Main function is Deep Feature Extraction used in perpetual loss.
3) GANS (Generative Adversarial Network): A discriminator network based on super-scaling residual in Residual Dense Network (RSD).

Crater Detection:
1) Objective of this section is to provide with a terrain relative navigation by identifying crater rim from the high resolution TMC images to safe guard the lunar navigation.
2) Ellipse R-CNN model is being used to identify craters from the provided images.
3) Elliptical Object Detection includes two components - Mask R-CNN for elliptical Object retrieval and U-Net Semantic Segmentation) for learning different occlusion patterns
4) Open CV: For using pre-trained Computer Vision Models.

Crater Matching (Crater Pattern Identification):
1) Objective: It is used to identify the pattern in the craters found from the images to provide faster crater recognition.

Visual Terrain Relative Navigation:
1) Identifying Lunar Surface Features like craters depth, boulder, rifts, slope to help avoid hazards during lunar navigation.
2) Terrain Classification: DenseNet is used to segregate the images into different segments based on CNN. It is useful for route planning and obstacle avoidance.
3) Depth Estimation: Pix2Pix and GANs are used for depth estimation of moon's surface from the relative position at which the image was captured.

Show Stopper:
1) The feature-based transform is harder to train, and require some hyperparameter tuning and loss balancing 
2) Lighting Conditions: Moon's surface has areas of permanent shadow or of strong sunlight.
3) Regular Updates: Due to the dynamic environment of the moon.
4) To improve the resolution of the images from terrain relative cameras.

Dependencies:
1) ML models (Deep Learning Models).
2) Computational Resources.
3) Fine tuning of Generative models.
4) Image Processing.