Recent deep learning techniques are successfully integrated into devices to assist visually impaired people in their daily lives, particularly detecting coins/banknotes. Previous works have focused on well captured devices and examined high-quality images. In this work, we design a framework to recognize Japanese Coin/Banknote (JCB) for low-quality images under various criteria. Discriminate features usually disappear in low-quality images. Consequently, using the depth image in addition to RGB image in processing can be enhanced the accuracy of our system. In this work, we first leverage depth information by using a Monocular Depth Prediction network. Additionally, a pre-trained Deep Convolutional Neural Network process RGB and Depth images, respectively. At last, we combine two networks by an ensemble method to produce more accurate detections. By processing depth images in addition to RGB images, the detection results are thus accurate. As a result, our work achieves 74.1% mean Average Precision (mAP).