Review — A Blur Classification Approach Using Deep Convolution Neural Network (Blur Classification)

Blur Classification for Hand Gesture Images Using CNN

In this story, A Blur Classification Approach Using Deep Convolution Neural Network, (Tiwari IJISMD’20), by University of Petroleum and Energy Studies, is reviewed. In this paper:

• Four blur categories: motion, defocus, Gaussian, and box blur are classified for hand gesture images by using CNN.
• It is essential to determine the exact blur type for blind image restoration.

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

Outline

1. Blur Models
2. Blur Classification Framework
3. Experimental Results

1. Blur Models

1.1. Motion Blur

• The motion blur appears when the object to be captured moves relative to the imaging device:

• where L is the blur length of uniform motion blur.

1.2. Defocus Blur

• Images that are captured due to a deprived convergence of light from an object on the sensor plane results in defocus blur (or out-of-focus):

• where R is the circular disk of radius.

1.3. Gaussian Blur

• Blur function are usually approximated by the Gaussian function for a large range of devices like optical image camera, microscope, telescope, etc:

• where σ is the Gaussian blur function with a standard deviation.
• Sometimes, this blur is also used to model defocus blur.

1.4. Box Blur

• Box blur is a mean filter where each pixel in the output image has mean of its neighbouring pixels defined in some specific region in the input image.
• Most existing methods model the blur as a box blur, which is caused by constant velocity movement:

• where Suv is the set of points in rectangular blur kernel and m*n is the kernel size.

2. Blur Classification Framework

Blur Classification Framework

2.1. Preprocessing & Fourier Transformation & Cropping

(a) Original image, (b) Fourier spectrum of (a), (c) 1-D Hann Window, (d) 2-D Hann window (e) Hann windowed image, (f) Fourier spectrum of (e)

• The periodic boundary conditions and sharp discontinuities among the borders of an image make a cross patterned in frequency spectrum.
• The Haan Window is used to avoid this problem.
• Then, Fourier transformation is applied.
• And the image is 32×32 cropped.

2.2. Multi-Layer Perceptron Model (MLPM)

A multi-layer perceptron neural network model

• A MLP with just three layers namely input, hidden and output is used.
• The neurons in the first layer i.e. input layer are 1024 and output layer are four, as the input is the 32 × 32 image and the output is a 4 × 1 vector indicating the four blur classes.
• The 500 neurons in the hidden layer are considered.

2.3. Convolution Neural Network (CNN)

A convolution neural network model

• In the passage. authors mentioned the CNN model used for blur classification has three convolutional filter’s layers: 64–3×3, 128–3×3, 512–3×3 convolution filters. (But in the figure, only 2 convolutions.)
• 2×2 max pooling is used.
• The last layer is a dense layer that has a softmax activation function with 4 units.
• Dropout is used to reduce overfitting.

3. Experimental Results

3.1. Datasets

Blurred Gesture Images

• Triesch gesture database consists of 720 images taken by 24 different persons in 3 dissimilar backgrounds comprise of 10 hand gestures for each, is used.
• All the images are blurred separately using each class of blur namely motion, defocus, box and Gaussian blur. Therefore, the blurred image database has the size of 2880 images.

3.2. MLP Results

Performance with multi-layer perceptron-based classification model

• The blurred image dataset is split into training and validation sets.
• 70% and 30% blurred images are utilized for training and validation separately.
• Whole image dataset is used as test set for the trained model.
• (Whole dataset which overlapped with the training set is strange.)
• The overall test loss and the test accuracy of the model are 0.22 and 0.93, respectively.

The accuracy curves and loss curves separately for training and validation data.

3.3. CNN Results

Performance with convolution neural network-based classification model

• The dataset split is the same as the MLP one.
• The overall test loss and the test accuracy of the model are 0.10 and 0.97, respectively.

The accuracy curves and loss curves separately for training and validation data.

3.4. ROC Results

(a) ROC Plots for MLP model, (b) ROC Plots for CNN model

• The above shows the ROC plots.

3.5. Accuracy and SOTA Comparison

• The average accuracies achieved through MLP and CNN models are 93.0 and 97.0, respectively.
• The synthesis database is also tested with Curvelet transform based energy features and a feed forward neural network-based classification model (Tiwari, 2017). The average accuracy using this model is 95.7.

Reference

[2020 IJISMD] [Tiwari IJISMD’20]
A Blur Classification Approach Using Deep Convolution Neural Network

Blur Classification

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

My Other Previous Readings