Fine-grained Leaf Classification (August, 2016 - December, 2016)

I am working on it under Professor Arnav Bhavsar, IIT Mandi. Here I am working on the Leafsnap Dataset towards developing a model that could classify leaf species with high accuracies. Till now, the following has been accomplished:

  • Testing of accuracies by using pre-trained VGG-16 weights trained on ImageNet as a fixed feature extractor and training a fully connected layer on top with different number of neurons.
  • Tweaking the model by varying number of VGG-16 layers and also by varying the number of layers and number of neurons in the final fully connected layers to determine best layer architecture.
  • Creating a class-level error-calculation matrix to analyse and group the classes in which the classifier is getting confused to increase the accuracy of classifier in categorising similar appearing leaves.

Currently working on differentiating leaf species amongst group of classes that appear similar to the classifier to increase overall clasification accuracy.

Sentiment Anaysis on IMDB reviews

Achieved near-perfect accuracies on IMDB movie reviews using Bag of Words and Bag of Clusters model.Code.

  • Clusters were made for words having similar vector representation in word2vec.
  • Random Forest Classifier was used to clasify reviews as positive or negative.

Mobility Assistant for Visually Impaired (August, 2016 - December, 2016)

Mobility Assistant for Visually Impaired(MAVI) is a product being developed by Assistech, IIT Delhi where we are working towards developing a mobility assistant for blind people that would provide them real-time instructions on the scenario in front of them using Image-Processing, Computer Vision and NLP techniques. I am working on the critical aspect of removing shadows from images, which cause serious troubles in other feature detection algorithms. The following has been accomplished till now:

  • Implementation of a 1-D Illuminant Invariant approach:
    1. Illuminant invariant image is obtained using a color constancy approach.
    2. Illuminant-invariant image does not contain shadow-edges and hence difference of Canny-edge detection on original image and illuminant invariant image gives shadow edges.
    3. The derivatives at the shadow edges are set to be zero and the image is integrated to achieve shadow-free image.
  • Using Depth cues for Shadow Removal:
    1. Key assumption: pixels with similar normals, spatial locations and chromaticity should have similar colors.
    2. A modified nonlocal matching is used to compute a shadow confidence map that localizes well hard shadow boundary, thus handling hard and soft shadows within the same framework.
    3. A standard energyminimization using the confidence map is then used to automatically produce the optimized shadowless image.

Currently working towards making the depth algorithm work better for outdoor images.