<![CDATA[Fingerprint recognition systems on resource-limited devices often face the challenges of aggressive image dimension compression (resizing) and natural scan tilt variations. This research does not aim to design a commercial identification system, but rather to specifically analyze the limitations of image resolution reduction (64x64, 96x96, 128x128, and 256x256 pixels) and to evaluate the effectiveness of synthetic rotation augmentation in compensating for Support Vector Machine (SVM) classification performance. The test uses a primary dataset (100 images, 20 classes) partitioned stratified (80:20) to prevent data leakage, where the augmentation process produces a total of 1,600 training images. In comparison, 20 test images are retained as pure unseen data. The stage continues with feature extraction using the Rotation Invariant Local Binary Pattern (LBP-RoR, radius 1). The experimental results show that a 64x64-pixel size is the threshold for structural failure, at which the ridge topology is fatally damaged, leading to a test accuracy of 10%. The model exhibited the highest overfitting phenomenon at 128x128 pixel resolution (training accuracy 79.17%, testing 40%). The best generalization equilibrium point was achieved at 256x256 pixels with a testing accuracy of 50%. This maximum achievement, which was stuck at 50%, demonstrates the vulnerability of the LBP and linear SVM margin methods to pixel-artifact distortion (aliasing) caused by digital rotation. This study concludes that spatial data augmentation cannot fully substitute the need for a physical finger alignment module (fingerprint alignment) in the preprocessing stage.]]>
