Classification of Blurred and Non-Blurred Images

CERTH Image Blur Dataset

E. Mavridaki, V. Mezaris, "No-Reference blur assessment in natural images using Fourier transform and spatial pyramids", Proc. IEEE International Conference on Image Processing (ICIP 2014), Paris, France, October 2014.

The dataset consists of undistorted, naturally-blurred and artificially-blurred images for image quality assessment purposes. Download the dataset from here: http://mklab.iti.gr/files/imageblur/CERTH_ImageBlurDataset.zip

Unzip and load the files into a directory named CERTH_ImageBlurDataset

Using OpenCV2

This method yielded an accuracy of 87.29%

cv2.Laplacian(img, cv2.CV_64F).var()

The Laplacian Operator is applied on each of the images. The variation of the result is calculated. If the variation is below the threshold, 400 in this case, the image is classified as blurry. Otherwise, it is classified as non-blurry.

To run this model:

python OpenCV_var.py

To load and pickle train data and its labels:

python load_traindata.py

To load and pickle test data and its labels:

python load_testdata.py

A Convolutional Neural Network is trained which yields an accuracy of 67.70% on the evaluation dataset. The deep learning model has five layers. This accuracy can further be improved by increasing the input dimensions of the first layer in the model and the number of epochs. However, due to constraints related to computational power, I was unable to run the model.

To train the CNN:

python CNN.py

Using OpenCV2

An accuracy of 63.58% is achieved using this method.

gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)

numpy.max(cv2.convertScaleAbs(cv2.Laplacian(gray,3)))

The result of the Laplacian Operator is converted to an absolute scale (0-255). The max of the values is taken for each image. The threshold is set at 215. Values lower than 215, classify the image as out-of-focus or blurry. Greater than 215 is classified as non-blurred.

To run this model:

python OpenCV_max.py