The use of QR Code has been flourishing on mobile and tablet platforms. By scanning the code, we can obtain targeted information synchronously. The regular QR Code, which consists of black and white modules, is neither visually pleasing nor recognizable by human vision. The application of graphic QR Code to product packaging and promotion campaign in the market has skyrocketed nowadays. However, printed graphic QR Code accompanies noise phenomenon that interferes the recognition itself and causes failure when user scanning. Therefore, we produce graphic QR Codes by data hiding with error diffusion techniques that first become training data. Then, we apply convolution neural networks to improve the data point recognition of graphic QR Codes. The experimental results show the superiority of the performance in both accuracy and recognition ability in comparison with normal QR Code readers.