Various computer vision problems solved using concepts of deep learning
- clone this repository
- run "jupyter notebook" at root folder
- Cricket Shot classification into off-side and on-side shots
- Dataset created using google_images_download python library. Search results were not good so dataset required manual cleaning.
- "cover drive cut cricket shot", and "defense shot cricket" were used as argument to image_download script for label:"0"
- "Sweep shot cricket", "pull hook shot cricket" were used as arguments to iamge_download script for label:"1"
- Pre-trained VGG-16 model used as feature extractor
- Small ConvNet followed by Dense layers trained to decide shot class
- Keras used to build model
- Lung Segmentation from CT images
- Dataset downloaded from https://www.kaggle.com/c/data-science-bowl-2017/data
- Various U-Net architectures experimented with to understand accuracy vs efficiency tradeoff
- Trained networks with dice coefficient as loss
- Some results shown in ipynb
- Pneumonia detection from chest X-rays
- Dataset downloaded from https://www.kaggle.com/parthachakraborty/pneumonia-chest-x-ray
- Built and trained Convolutional Neural Networks of various size.
- CLassification was between normal and pneumonia cases. Binary cross-entropy loss was used.
- Accuracy, precision and recall of the models was examined and confusion matrix plotted
- Scenery Detection
- Dataset created with google_images_download python library.
- Classes are 0: city; 1: ocean; 2: canyon; 3: desert; 4: mountain; 5: river; 6: waterfall
- Convolutional neural network trained from scratch on 1728 training examples. Results were not very good.
- Used Convolutional layers of VGG-16 as a feature extractor. This transfer learning strategy achieves better accuracy.
- Some wrong predictions are shown to understand possible reasons for error