Features Extraction for Detection of Blurred Image Regions
Aneta Bera, Dariusz Sychel · Applied Artificial Intelligence · 2016
The article discusses our study of the problem of blurred image regions detection. Proposed are 21 features: 20 based on the discrete wavelet transform and one being a ratio based on a gray-level co-occurrence matrix. Described features are introduced in order to detect blurred segment regardless of the blur type such as background blur, motion blur, or out of focus. Moreover, proposed features are relatively easy to calculate and parallelize. Multilayer perceptron with a backpropagation algorithm is trained on mentioned features. Three approaches are introduced and tested: (1) with nonoverlapping fixed size windows, (2) with nonoverlapping recursively divided windows, and (3) with addition of a morphological closing operator.