Improving the Accuracy of Iris Recognition Based on Attention Mechanism
Zike Wang, Jianguo Wang, Lei Zhong · 2023
Iris recognition algorithms are becoming more and more widely used in daily life, and the accuracy requirements for iris recognition algorithms are getting higher and higher. Therefore, in this paper, an improved Shufflenetv2 algorithm is proposed as a way to improve the speed and accuracy of current iris recognition algorithms in daily life. In this paper, deep and shallow features are filtered by using the channel attention mechanism, which enables the model to acquire feature representations of both underlying and high-level semantics. In addition, in the loss function, this paper uses the focal loss function instead of the cross entropy loss function to achieve a solution to the imbalance between the number of positive and negative samples and prevent too much learning to negative sample features. By validating on the CASIA-Iris-Thousand dataset, which is commonly used in iris recognition tasks, the proposed model achieves an equal error rate (EER) of 1.66% compared to the Shufflenetv2 model, which has an EER of 5.3%, and optimizes by 3.64%. The model proposed in this paper can effectively improve the accuracy of iris recognition.