Iris recognition based on local circular Gabor filters and multi-scale convolution feature fusion network

With the development of the social economy, the identification of biological characteristics has received more and more attention, and the iris is known as one of the most reliable biometric characteristics due to its stability, uniqueness, stability and non-replication. Although Convolutional Neural Network (CNN) is used widely and can recognize the iris effectively to some extent because of its high robustness in biometrics. However, a large number of training samples are required, and some important features will be lost since only the feature maps of the last CNN layer are used generally. The traditional Gabor wavelet transform can be used to augment the training samples, but its specific directional characteristics need accurate iris registration. In fact, accurate registration is hard to achieve due to the annular structure of the iris. Therefore, the Local Circular Gabor Filters (LCGF) is first used to augment the training set, which can preserve all the directional information and do not need accurate registration. And then. The Multi-scale Convolution Feature Fusion Network (MCFFN) is proposed to extract more features. First, an iris image is processed by iris location, segmentation, and normalization. Second, the preprocessed image is filtered by local circular Gabor filters with four scales of low and medium frequency filters to get four samples. The original image and its filtered images form the augmented database. Third, the MCFFN is constructed based on the augmented database by experiments. The MCFFN extracts the feature maps generated in the last three convolution layers of CNN, and these feature maps are normalized to the same size by adaptive average pooling and form the final iris characteristics by channel splicing. Fourth, the SoftMax method is used to classify the test samples. Finally, the final result is obtained by using the voting method. The method is experimented in CASIA-Iris-Syn and CASIA-Iris-Lamp iris databases, and the experimental results show that this method can effectively perform iris recognition even if the iris is unable to register accurately.