-The accuracy of segmentation plays an important role in the methods for personal identification by iris images. In this paper, we study the iris segmentation method based on the convolutional neural network of the UNet architecture. As part of the research, manual segmentation of the images of the used dataset was performed, the optimal network training parameters were determined, and the segmentation quality was evaluated. All studies in the paper were conducted using the open MMU Iris Image Database. The results showed that the studied approach provides high precision segmentation of the iris images.
INTRODUCTION
Recently, the most widespread methods of identifying a
person use biometric data (face images [
Unfortunately, the segmentation of the iris performed by these methods is often not accurate enough due to the partial overlap of the iris with the eyelids and eyelashes, as well as the appearance of glare from the light source. Also, the shape of the iris itself is not always well approximated by circles. Thus, a relevant task is to improve the quality of iris segmentation.
With the success of deep neural network models,
researchers are increasingly turning their attention to
convolutional neural networks to further improve the
accuracy of existing iris segmentation methods. To solve the
problem of iris segmentation, we propose to use an approach
based on training the convolutional neural network (CNN) of
the U-Net architecture [
U-Net is considered one of the standard CNN architectures for image segmentation tasks when you need not only to define the entire image class but also to segment regions, i.e. create a mask that will divide the image into several classes. One of the main advantages of the network, Evgeny Myasnikov which is essential in the context of this work, is the ability to train it on a small amount of data.
The paper has the following structure. Section 2 is devoted to the description of the studied approach to iris image segmentation, a description of the network architecture, the features of data preparation, and training. Section 3 describes the dataset used in the paper and provides the results of experimental studies. The paper ends up with the conclusion and reference list.
II.
U-Net is a famous CNN architecture for solving biomedical problems (segmentation of various types of cells, determining the boundaries between dense cell structures, etc.) and other image segmentation tasks. The main advantage of this model is the ability to learn on small datasets and, at the same time, show high accuracy, which is a common problem for computer vision problems.
Fig. 1. The U-Net architecture [
As you can see in the figure, the architecture consists of a contracting path (left) to capture the context and an expanding path (right) to allow precise localization. There are two classes in the paper: iris and non-iris.
The contracting path corresponds to the typical structure of a convolutional neural network. It consists of re-applying two convolutions (kernel size 3x3), followed by a layer of ReLU and max-pooling (kernel size 2x2) with step 2.
Each step of the expanding path contains an expanding feature map layer (up-conv), the reverse of the max-pooling, followed by a convolution (kernel size 2x2) that reduces the number of feature channels. This is followed by concatenation with the appropriate cropped feature map (copy and crop) from the contracting path and two convolutional layers (kernel size 3x3), each of which is followed by a ReLU layer.
B.
At the stage of preparation for training, a pre-trained
model was taken [
Binary cross-entropy was used as a loss function, which
determines how the network quality is estimated on training
data. Adam was chosen as the optimizer determining the
mechanism for updating the weights [
III.
EXPERIMENTS AND RESULTS
For experimental research, implementation in Python was performed using the Keras, Numpy, OpenCV, and Matplotlib libraries.
The MMU Iris Database [
Nine experiments were conducted to determine the best
parameters. The neural network was trained using the found
parameters for 50 epochs. At the stage of searching for the
best parameters during training, the pre-trained neural
network [
To monitor the training process, the value of crossentropy and segmentation accuracy was evaluated using the validation set at the end of each epoch of training.
Segmentation accuracy was estimated as the proportion of correctly classified pixels in the images of the validation set.
The values of the loss function and the accuracy of the trained model on the training and validation data for various values of the batch size and learning rate hyperparameters are shown in Table 1.
The plot of changes in the accuracy values and loss function in the first 10 epochs of neural network training for the two best models can be seen below in figures 2 and 3.
As a result of experiments, the following values of training parameters were selected: learning rate was equal to 10-4, the size of a random sub-sample (batch size) was 40. Training a neural network with the specified parameters for 50 epochs allowed us to achieve accuracy and loss function values equal to 99.53 and 0.0071, respectively, on the validation sample.
The results of the studied approach for the resulting model are shown below in figure 4.
Besides, in this paper, we provide an example of iris
segmentation using the previously developed technique [
A brief description of the previously proposed technique is given below:
1. Determination of the centers and radii of two circles approximated the pupil and the iris using the Hough transform.
2. Crop the image.
3. The resulting image is blurred to eliminate minor noise and smooth the iris.
4. Conversion to a monochrome image. 5. Contour detection. 6. Morphological transformation (closing gaps).
7. Construction of the mask of the open area of the iris (each pixel has a value of 1 if this pixel belongs to the iris, and 0 otherwise).
The use of the aforementioned technique allowed obtaining 90.67% of segment accuracy.
The accuracy of the method considered in this paper is 99.53%, which is 8.86% higher, the use of a CNN significantly compared to the previously proposed technique.
In this paper, we study the iris segmentation method based on the convolutional neural network of the U-Net architecture. To configure and evaluate the method, manual segmentation of images from the open MMU Iris Image Database was performed.
Using the specified dataset, the best values of the training parameters were determined, and segmentation quality was estimated. Studies have shown that the approach under study allows us to achieve 99.53% segmentation on the validation dataset.
The work was partly funded by RFBR according to the research project 18-01-00748 in parts of «II. Method» - «III. Experiments and Results» and by the Russian Federation Ministry of Science and Higher Education within a state contract with the «Crystallography and Photonics» Research Center of the RAS in parts «I. Introduction» and «IV. Conclusion».