Finger-vein as a high security biometric characteristic has been widely investigated for verification. One of challenges in finger-vein recognition is the image-quality degradation as spurious and missing features in poor quality images can increase the verification error. Despite recent advances in finger-vein quality assessment, these solutions depend on domain knowledge. Training a deep Neural Network (DNN) based on the objective labels selected automatically has compromised this problem, but their performance is still limit because some quality attributes are ignored. In this work, we propose a DNN for representation learning from both grayscale and binary images to predict vein quality. In the proposed approach, the grayscale and binary images are directly input to DNN to learn the joint representations for quality assessment. Experimental results on one large public dataset demonstrates that the proposed method accurately identifies high and low quality images and outperforms other methods in terms of impact on equal error rate (EER) decrease