Venipuncture is a nearly ubiquitous part of modern clinical practice. However, currently venipuncture procedures are mainly applied by manual operation, whose the success rate might decrease below 50% in some situations, including pediatric, and geriatric patients. Thus, robotic technologies to guide autonomous vascular access attracts research attention. For venipuncture robots, near-infrared (NIR) images are widely used for real-time servoing and further to segment subcutaneous vessels for puncture with a series of deep convolutional neural networks. It has been realized that the success rate of puncture largely relies on the performance of segmentation models. However, the small size and low quality of NIR image dataset severely limit the accuracy and efficiency of segmentation models. This paper aims to address this issue by proposing a novel data processing method to improve the performance of segmentation models. With those novel image processing strategies, the segmentation results are improved. The Dice-mean value has increased by an average of 1.12%. Additionally, an algorithm of vessel site selection for puncture is proposed in this paper. Such data processing methods and the puncture site selection algorithm are expected to finally improve the performance of venipuncture robots.