In this project I have taken ten simpsons charcaters and have tried to make end to end image classification where it can detect character which are even blurred.
-
I have taken complete dataset from kaggle (https://www.kaggle.com/alexattia/the-simpsons-characters-dataset) but have used just 1038 images in total for 10 different characters due to computational restriction.
-
Here I have blurred 30% of data in each category.
-
I have made skewed dataset where I have used one category which is just 3% of total dataset.
-
I have done data generator process in Data Genrator file.
-
First Data augmentation was applied for the category which is less than 3% of total dataset.
-
To identify blur images better random blurr ranging from 1 to 50 have been used.
-
For training two transfer learning technique such as VGG16 and ResNet50 were used.
-
Got Better accuracy of in VGG16 as comapre to ResNet50.
Flask and heroku is been used for the deployment.
https://simpsons-character.herokuapp.com/
Detecting blur image: