Cataract is one of the most prevalent causes of blindness in the industrialized world, accounting for more than 50% of blindness. Early detection and treatment can reduce the suffering of cataract patients and prevent visual impairment from turning into blindness. But the expertise of trained eye specialists is necessary for clinical cataract detection and grading, which may cause difficulties to everybody's early intervention due to the underlying costs. Existing studies on automatic cataract detection and grading based on fundus images utilize a predefined set of image features that may provide an incomplete, redundant, or even noisy representation. This paper aims to investigate the performance and efficiency by using Depp Convolutional Neural Network (DCNN) to detect and grad cataract automatically, it also visualize some of the feature maps at pool5 layer with their high-order empirical semantic meaning, providing a explanation to the feature representation extracted by DCNN. The proposed DCNN classification system is cross validated on different number of population-based clinical retinal fundus images collected from hospital, up to 5620 images. There are two conclusions suggested in this paper: The first one is, the interference of local uneven illumination and the reflection of eyes were overcome by using the retinal fundus images after G-filter, which makes an significant contribution to DCNN classification. The second one is, with the increase of the amount of available samples, the DCNN classification accuracies are increasing, and the fluctuation range of accuracies are more stable. The best accuracy, our method achieved, is 93.52% and 86.69% in cataract detection and grading tasks separately. It is demonstrated in this paper that the DCNN classifier outperforms state-of-the-art in the performance. Further more, The proposed method has the potential to be applied to other eye diseases in future.