Review — Convolutional Neural Network for Blur Images Detection as an Alternative for Laplacian Method (Blur Classification)

A 2-Conv+4-FC Model to Classify if an Image is Blurry or not

Images described by an expert as blurry (left) and sharp (right).

In this story, Convolutional Neural Network for Blur Images Detection as an Alternative for Laplacian Method,(Szandała SSCI’20), by Wroclaw University of Science and Technology, is briefly reviewed, as I’m studying this problem recently.

The photographers employed for picturing ceremonies such as weddings admit that as many as 40% of the images have insufficient quality to be proposed to a client. One of the key factors that lead to quality degradation is blur.

• In this paper, a CNN is designed to classify if an image is blurred or not.

This is a paper in 2020 SSCI. (Sik-Ho Tsang @ Medium)

Outline

1. Network Architecture
2. Experimental Results

1. Network Architecture

Network Architecture

• As seen above, there are two convolutional layers. (The kernel size and stride are unknown.)
• The activation type is unknown.
• Max pooling is used after each convolution. (The kernel size and stride are unknown.)
• Then, there are four dense layers, i.e. fully-connected layers.
• Dropout is used. (The percentage is unknown.)

2. Experimental Results

• The experiment has been done on two thousand photos of different resolutions. These mostly include the author’s personal photos.
• While 80% of the images were used as the training dataset, the rest were used as validation collection.

Confusion matrix containing the results of the classification for the Laplacian method

Confusion matrix containing the results of classification for the CNN method

• The above shows the confusion matrix, sensitivity and specificity of Laplacian method and the proposed CNN.
• The results by Laplacian show that almost 91% of the images are correctly recognized as blurry. However, one-fourth of the non-blurry images were considered blurry. Overall, the Laplacian method provides an accuracy rate of over 80%.
• The proposed CNN method stays far behind the Laplacian method in the case of sensitivity (66% to over 90%, see the both tables).
• However, surprisingly, CNN has better results in specificity metric.

Reference

[2020 SSCI] [Szandała SSCI’20]
Convolutional Neural Network for Blur Images Detection as an Alternative for Laplacian Method

Blur Classification

2017 [SFA] 2019 [SFA & SFGN] 2020 [Szandała SSCI’20]

My Other Previous Readings