The global outbreak of the COVID-19 has greatly changed people's daily life. It has been common measures to wear masks and avoid contact. These measures prevent the spread of the new crown pneumonia virus but hinder the usage of currently popular authentication technologies, such as face recognition, fingerprint recognition, ID card. Therefore, researchers are looking for a high-performance identity authentication technology available in the current epidemic scenario. The iris recognition technology attracts their attention due to its high reliability and contactless acquisition. However, the further promotion of iris recognition technology needs to break through its controlled acquisition condition and explore it in the open-world scenario with fewer acquisition constraints. There are a large number of uncertain acquisition factors in the open-world scenario. This acquisition uncertainty inevitably affects the preprocessing and recognition of iris images, leading to performance degeneration of the iris recognition system. To address these problems in the open-world scenario, we propose several approaches for preprocessing and recognizing iris images. The main contributions in the thesis are summarized as follows.
Considering the impact of uncertain acquisition factors in the preprocessing of iris images, we propose the iris meta super-resolution network based on the distributional discrepancy perception and the bilaterally contextual segmentation network based on predictive discrepancy perception. The first model decomposes the model parameters of the backbone network into fixed parameters determined by the training data and dynamic parameters that change with the input sample. An independent meta network computes the dynamic parameters by perceiving the difference between the input and training images to fit the backbone network to the input image. The second model leverages the predictive discrepancy to estimate the uncertainty of the segmentation result and then utilizes this uncertainty estimate to make the optimizer pay more attention to the high-uncertainty area. In addition, the model learns the bilateral context according to the visual characteristics and spatial layouts of ocular components to segment out the iris region.
We propose a heterogeneous iris recognition method based on device-specific band removal to reduce the distribution gap between samples from different spectra. This method first applies the Gabor function as the prior knowledge to perceive iris textures under different spectra. Then it adopts a trident network to decompose an image into a basic component that is about to be aligned and a residual component containing the device-specific band. The residual component helps the basic component to generate spectral-invariant features. Experiments on two cross-spectral iris datasets and two cross-sensor iris datasets show that our method reduces the distribution gaps between cross-spectral images and between cross-sensor images, significantly improving the recognition performance.
To mitigate the feature ambiguity dilemma caused by the acquisition uncertainty in the open-world scene, we present a probabilistic implicit representation to describe iris images and leverage it to address iris recognition with insufficient labels. This representation applies a probabilistic distribution instead of the previous deterministic point to represent an iris image, in which the mean encodes the most likely identity feature of the iris image, while the variance encodes the data uncertainty from acquisition factors. Each captured image can be regarded as an instantiated iris feature sampled from the parametric probabilistic distribution. Furthermore, we propose contrastive uncertainty learning based on this probabilistic implicit representation for iris recognition under the semi-supervised and unsupervised settings. Experiments on six iris datasets demonstrate the effectiveness of probabilistic implicit representation, while experimental results on semi-supervised and unsupervised settings show that unlabeled data is also beneficial for performance improvement.
To alleviate the local context modeling problem, we propose multi-scale contextual measures to mitigate the degraded recognition. This method analyzes the contexts between iris regions from global and local perspectives. The global context describes the relationships between all iris regions and is robust to local occlusion, while the local context measures the relationships with neighboring regions and is sensitive to local texture details. Experimental results on four datasets show that this method significantly outperforms compared approaches in multiple evaluation metrics. Moreover, extensive experimental results illustrate that global and local contexts are different clues critical for accurate iris recognition.