This repo contains shows all results generated in academics and research work during my Masters.
Image inpainting for satelite images of Lunar and Martian surface was performed to fill/predict the mising pixel values. Original Deep image Prior method was modified to obtain good results. Results are shown below
Data augmentation was performed to generate additional/synthetic training samples to train the model. An example of data augmentation is shown.
Two-class image classification with limited no. of samples (less than 100 sample) was performed using augmentation and k-fold cross validation.
Various architectures were tried and results are mentioned below.
Images were cluserted using autoencoders. Both full connected and Convolutional autoencoders were used to perform dimensionality reduction and then clustering techniques were applied. A typical structure of Convolutional autoencoder is shown below.
Entire process of Convolutional Autoencoder + Clustering is shown below.
Training CNN models to perform crater detection and segmentation using a novel method and architecture to beat the SOTA results. An example of visual result is shown.
Consolidate results are as shown below.
Faster RCNN model was trained on custom dataset to perform object detection and counting. The model was trained to detect Cars, Trucks and Humans. This was done as part of the project for ML course in IIT Gandhinagar. The results obtained for some test images are shown below.
Implementation of Computer vision algorithms from scratch in MATLAB. The results obtained are also shown. Following Algorithms are implemented from scratch in MATLAB and their results are also shown.
1. Image and video denoising by sparse 3D transform-domain collaborative filtering Link to Paper
Image de-noising is performed using 3D transform-domain collaborative filtering.
Result for denoising are shown below. Image on the left is noisy input image. Image on the right is de-noised image obtained by code.
Here de-noising is performed for higher noise values in input image. In the second image, the amount of noise present is very high, still the code manages to produce decent output and recover patterns present in original images.
This image shows input image, noisy image, output of first stage and final output of the code.
2. Panorama creation and Image stitching Link to paper
Combining multiple images and creating a panaroma using extracted features and matching keypoints in both images.
This was created withouth using interpolation to fill in the missing pixel values.
Convolution with Gaussian Filters and then using Difference of Gaussian filter to perform edge detection.
4. SIFT (Scale Invariant Feature Transform) for object Recognition.Link to paper
SIFT is a highly cited paper for feature extraction and object recognition. It is implemented from scratch and SIFT features are extracted.
Pixel realignement was performed between a set of stereo images.
Pixels of second image were realigned to resemble/recreate the first image.
Evolutionary Computing algorithms like Genetic Algorithms (GA), Particle Swarm Optimization (PSO), Non-Domination Sorting Algorithms (NSGA-2), Fuzzy Algorithms and ANFIS along with their results. Results are shown in .GIF/.PNG format for all codes.