Although the pulse transit time is generally used for blood pressure estimation without a cuff, a method of estimating blood pressure only from photoplethysmography (PPG) based on the relationship between pulse waveform and blood pressure has been studied. This can eliminate the need for an electrocardiogram and allow more continuous and simpler blood pressure measurement. Previous studies have proposed methods of machine learning by extracting features such as wave height and time difference, or generating features with an auto-encoder. In this paper, we propose a method to estimate blood pressure and to automatically generate features from pulse wave using the convolutional neural networks (CNN). By comparing the accuracy of the proposed method with that of the conventional method, the effectiveness of cuffless blood pressure estimation from only PPG by using CNN is examined.