Biometric authentication via palmprint recognition presents a compelling solution for high-security applications, yet its reliability is frequently challenged by significant geometric variations in pose, scale, and orientation inherent in unconstrained image acquisition. While end-to-end deep learning classifiers have demonstrated considerable promise, that their architectural design conflates the distinct tasks of feature learning and geometric alignment, inherently limits their robustness. To address this limitation, this paper represents a two-stage system founded on the principle of explicit decoupling. The first stage, ROILAnet, employs a bespoke regression network to predict control points for a Thin-Plate Spline (TPS) transformation, which warps the input image into a canonical Region of Interest (ROI). The second stage then utilizes a fine-tuned ResNet-18 architecture to extract a discriminative 512-dimensional feature embedding from this geometrically normalized ROI for verification. The proposed system achieves a verification accuracy of 98.33% and an F1-score of 0.9780, surpassing the best baseline while demonstrating superior class separation and robustness. Furthermore, the simulation results empirically validate the core thesis that explicit geometric normalization is a critical and principled step towards developing state-of-the-art, reliable palmprint recognition systems.