image-blur-detection

To get all the Dependencies (run the command) after forking / cloning this repository :

pip install requirements.txt

Approach - 1 : Convolutional Neural Network (CNN / ConvNets)

(Step-1) Load & Pickle Train dataset (run the command) :

python train.py

(Step-2) Load & Pickle Test dataset (run the command) :

python test.py

A Convolutional Neural Network is trained over CERTH_ImageBlurDataset (~3.7 GB) yielding accuracy of 58.18% on evaluation dataset. Accuracy can further be improved by increase input dimensions (of first layer) / model's complexity or tweaking number of epochs.

(Step-3) To train the CNN model (run the command) :

python model.py

---

Approach - 2 : Variance of Laplacian

Here we calculate variance of Laplcian; giving value which defines blurry metric. If it's below certain threshold (here it's 435) image can be classified as burry else it is going to be non-blurry. This model gave accuracy of around 87.57%. It's also, performed over the above given dataset only.

(Step-1) To run the script for Laplacian approach (run the command) :

python Laplacian.py

---

NOTE : While running scripts you can ignore / supress warnings / informations which arises due to difference in API's and other dependencies