This study mainly aims to create a wholly automated palm reading system based on computer vision techniques and deep learning models. This process includes the use of MediaPipe landmark detection for the geometric normalization of images, followed by U-Net-based semantic segmentation on a set of one thousand palm images randomly drawn from the hands dataset. Without manually labeling the hands, the ground-truth masks were generated automatically with MediaPipe to accelerate the training phase. The experimental data after 50 epochs revealed that the model was still consistently improving without any signs of overfitting and was capable of achieving a best Intersection-over-Union (IoU) score of 0.2125, final training loss of 0.5446, and a validation loss of 0.5197. Whereas the very last accuracy score obviously shows how segmenting extremely fine palmar creases with auto-generated masks is a very hard task, training plots reveal that the model has a very rudimentary capacity to differentiate semantic line patterns. Hence, this paper sets up a dependable and reproducible experimental baseline for automated chiromancy and also demonstrates the need and potential for very significant improvements through manual dataset annotation.

Keywords: Computer Vision; Deep Learning; Mediapipe; Image Segmentation; Artificial Intelligence; Edge Detection