With the rise in the digitalisation of personal data, the need for robust ways to secure this data to prevent trespassing and to identify criminals has also increased. Palmprint recognition offers a reliable and efficient approach to achieving these targets. However, the efficiency of the subtasks involved in this process, such as cropping of the region of the interest (ROI), feature extraction and fusion, needs to be improved. Existing ROI cropping algorithms suffer from various drawbacks, for example the use of inconsistent information to construct the cropping coordinates (such as the fingers or the spaces between them), extraction of the boundary using algorithms with unsatisfactory performance, losing the palm boundary information during denoising, or ignoring the illumination variations when calculating threshold value to binarise the images. In the process of feature extraction, one of the recently proposed feature extraction algorithm is subset LDA (S-LDA), which clusters the classes into subsets and applies LDA to each subset to achieve better accuracy. However, S-LDA suffers from high sensitivity to the inevitable presence of noise, and this results in mapping the classes into incorrect subsets. State-of-the-art fusion models either have high computational complexity or cost, or are not applicable to identification tasks. Previous work has implemented multiple matching schemes or feature extraction algorithms, without exploiting the availability of the samples in the datasets. A new ROI cropping method is proposed in this work. A median filter is first used to denoise the palmprint image, and Otsu’s binarising method is then used with an adapted and normalised threshold value. Next, Moore’s neighbour boundary tracing algorithm is used to extract the boundary of the palm contour. Following this, the distance between the boundary points on the fingers side and the centre of the image is calculated, based on the x-axis coordinate. Finally, local minima are extracted from the distance values to construct a cropping coordinate system. The features are extracted using a new method called overlapping subset linear discriminant analysis (overlapping S-LDA). In the proposed method, two-dimensional LDA is used to transform the classes into subspaces. The cost matrix is calculated in order to cluster classes that are easily confused into subsets. This method increases the tolerance of the mapping of the classes. A new multi-sample multi-instance fusion scheme at rank level is proposed in this work. In the offline phase, various sample formations are constructed by resampling, and their eigenvectors and feature vectors are extracted from these different formations, while in the online phase, the tested classes are transformed into different subspaces and matched with the classes in corresponding subspace, to generate a range of results that are then fused at rank level. The proposed methods are tested on the PolyU palmprint database for preprocessing, and are then applied to the PolyU and IIT Delhi databases for feature extraction, fusion, and the combination of the proposed methods. When the proposed methods are combined, the system achieves a 100% recognition rate, even when only a single palm is used.
