Machine-Vision

This repository consists of algorithms related to Images processing, Features, Segmentation, Pattern Recognition and Deep learning focused on Images. It consists of problems which are manipulated in both the traditional and the AI way.

Installation

All the packages required for the repository, can be obtained using the command

pip install requirements.txt

Datasets:

1. Deep Learning: The dataset for the deep learning code is MNIST image dataset but in the tabular format where every tuple indicates an image.

   Elements per row: ( 28 * 28 ) image with a label = 785 elements

   The dataset is highly visible online

2. Optic Disc Segmentation: The dataset for the optic disc segmentation can be acquired from IDRiD and store the dataset in the Images folder directory.

   The directory representation is as follows:

    -> Images 
          -> IDRiD
               -> optic_disc_segmentation_masks
                    -> Label images (54 samples)
               -> original_retinal_images
                    -> Input images (54 samples)
   

All the datasets required for other related problems are provided with the repository.

Results and Comparisons

1. Feature extraction: Feature extraction is done using the SIFT algorithm which is a texture based feature extractor which extracts keypoints in an image. SIFT is applied on an Eiffel Tower image and we check how SIFT is able to match the keypoints from the original image and the rotated image.

   1. Rotated Image angle 0 degrees:

      

   2. Rotated Image angle of 45 degrees:

      

   3. Rotated Image angle of 90 degrees:

      

   Other comparisons are available in OutputImages/Sift

   Code available at: src/sift.py

2. Segmentation: The segmentation algorithms that are evaluated and compared are Water Shed and Means Shift algorithm.

   1. The segmentaion results on a strawberry toy image is:

      

   Other comparisons are available in OutputImages/Segmentation

   Code available at: src/segmentation.py

3. Pattern Recognition: The confusion matrix for the result obtained from KNN on the scikit-learn digits dataset is:

   

   Code available at: src/pattern_recognition.py

4. Deep Learning: The loss vs epochs graph for the deep_learning code on MNIST dataset is:

   

   Code available at: src/deep_learning.py

5. Oil Painting: Oil painting is implemented from scratch as part of the learning process of manipulating images and it is implemented on a Sydney tram image. Oil painting is done on the image using different filter sizes and the results are:

   

   For detailed description about the project kindly look at the paper report/Oil-Paint-report.pdf where everything about the implementation is explained visually.

   All the codes related are available in the directory src/Oil-Paint.

6. Optic Disc Segmentation: The Optic Disc Segmentation is evaluated on the IDRiD dataset with 54 images:

   1. Image of a good prediction of the optic disc:

      

   2. Image of a bad prediction of the optic disc:

      

   For detailed description about the project kindly look at the paper report/IDRiD/paper.pdf which explains the implementation in detail and also about the reasons for the prediction.

   All the codes related are available in the directory src/IDRiD.

Acknowledgement

I would like to thank Arcot Soumya and her team for supporting me with the materials and the knowledge required to learn Machine Vision.