Summary

Image Processing

• Converting to Lab Color Space.
• Computing Average of L*a*b components of Image.
• Apply Gaussian filtering to the independent L*a*b planes using Kernel.
• Removing Average pixel values from blurred Image using L2 Normalization.
• Compute Precision and Recall using Adaptive Thresholding maps.

Deep Learning - Unsupervised Learning

• Transform Saliency maps into Square images with fixed sizes in case of this project 64*64.
• Transform 2D arrays of images into single arrays using Ravel() function.
• Create function for Normalizing and De-Normalizing Images.
• Decide Encoder Dimension. [256, 128, 64]
• Define an encoding function to create a network which will perform a series of multiplications on flattened array, followed by adding a bias, and wrap all of this in a non-linearity.
• This creates a series of multiplications in the graph which transforms the input of batch size times number of features(64*64) which started as None x 4096, to None x 64.
• Define a Cost Function using squared difference of mean values for input image and computed image.
• Define a learning rate along with an optimizer for improving the model results.
• Define a for loop to train batch of images using encoding function and computing cost.
• Feed the cost to optimizer along with learning rate to improve the efficiency.
• Plot the reconstructed image for every epoch of training.
• Train till loss converges record the epoch and loss.
• Fix the epoch as the range for training the data to create a Model.

Deep Learning - Supervised Learning

• Transform Saliency maps into Square images with fixed sizes in case of this project 64*64.
• Transform 2D arrays of images into single arrays using Ravel() function.
• Define a Sequential Model.
• The model is attributed with 3 activation layers including (RelU, Softmax and RelU) with (4096, 2048, 4096) nodes respectively.
• For optimization I have used "Adam" optimizer along with a loss function of "Binary Cross Entropy".
• Metrics for measuring efficiency is "Accuracy".

Future Scope

• The result for Supervised learning can be improved by fine-tuning the model.
• Deep Learning models require enormous set of data to improve the results.
• If large data set is available feature learning is better.
• TensorFlow - GPU can be used to improve the accuracy and time for model training.