Projects

Experimented with feature extractors like AlexNet, ResNet50, DenseNet and GoogleNet to extract features from images.

Experimented with different language models to generate words one at a time to describe the content of an image.

Observed that feature extractor ResNet50 along with language model LSTM gave the best results.

Encoded images using HOG (Histogram of oriented gradients) algorithm and a pretrained neural network.

Tackled posing of faces using face landmark estimation algorithm.

Compared encodings from faces found in real time video to the images in the database to mark the attendance.

Image regression is an application of deep learning where we have a dataset that’s annotated in a certain way. For example, for every image in our dataset, we would have the co-ordinates of the eyes of that person. We then train a model to predict these co-ordinates for new images.

Used the ResNet50 architecture to train a model for detecting pneumonia in medical images.

Applied discriminative learning rates and achieved an accuracy of 96.33 %.

Successfully reduced the training time without much compromise in model performance by switching to mixed precision training and achieved an accuracy of 95.82 % .

Implemented neural networks to assign multiple labels to satellite images of the Earth to aid biodiversity conservation.

Leveraged a pretrained model of ResNet50 and achieved an accuracy of 95.97% and f_score of 93.08%.

Secured 80th position on private leaderboard of Kaggle competition.

Utilized the language model built on Wikitext-103 which is good at predicting words in Wikipedia articles.

Trained the language model on the movie dataset to make it good at predicting words in movie reviews with an accuracy of 30%.

Classified movie reviews with an accuracy of 94.39% using the pretrained model.

Categorized images from a dataset consisting of 25000 images, into various types of scenes.

Experimented with different pretrained models like ResNet34, ResNet50 and ResNet101. Observed best accuracy of 94.8% with ResNet101.

Implemented progressive image resizing which resulted in 94.6% accuracy.

Executed by treating image segmentation as a classification problem by predicting what every pixel in the image is.

Leveraged the U-net architecture as they perform better than GANs on applications like generating high resolution images from blurry images, thus color coded every pixel of an image using U-net.

Achieved an accuracy of 92.4%.

Defined a mask of green color to detect the football pitch.

Applied Morphological closing operations on the frames of the video to reduce false detection.

Detected players by finding contours. Applied masking on the players to detect the jersey colors. Detected the football by applying similar masking operations.

Implemented a pipeline using OpenCV for detecting lane lines in images and videos of roads in Python for navigation in an autonomous vehicle.

Used Canny edge detection to detect edges, followed by Hough transform to detect prominent lines in the Region of Interest.

Conducted exploratory data analysis to understand user download behavior across various categories of apps on play store.

Gathered user reviews for the apps downloaded by the users and updated ratings in database.

Conducted time series analysis to understand trends in app download.

Conducted sentiment analysis to give an insight about the kind of apps that can be launched in future.