Advances in computer vision (CV) have led to an increasing market for biometric recognition systems. However, as more users are registered in a system, its expanding dataset will increase the system’s response time and lower its recognition stability. As mentioned above, we propose a new high-performance algorithm suitable for embedded finger-vein recognition systems. First, the semantic segmentation based on DeepLabv3+ filters out the background noise and enhances processing stability. The adaptive symmetric mask-based discrete wavelet transform (A-SMDWT) and adaptive image contrast enhancement were used in the preprocessing of images, and feature extraction was performed through the repeated line tracking (RLT) method. Next, the histogram of oriented gradient (HOG) of the image was computed, after which a support vector machine (SVM) was then used to train a classifier. Finally, a self-established finger-vein image dataset as well as a public dataset was implemented in the Raspberry Pi platform, which is a low-level embedded system. The experimental results indicated that the proposed system offers advantages such as a high accuracy rate, low device cost, and fast response time. Therefore, the three major issues that were encountered in previous embedded finger-vein image verification systems were mitigated in this work.